WO2024099187A1 - Ai模型策略确定方法、装置、第一设备及第二设备 - Google Patents

Ai模型策略确定方法、装置、第一设备及第二设备 Download PDF

Info

Publication number
WO2024099187A1
WO2024099187A1 PCT/CN2023/128628 CN2023128628W WO2024099187A1 WO 2024099187 A1 WO2024099187 A1 WO 2024099187A1 CN 2023128628 W CN2023128628 W CN 2023128628W WO 2024099187 A1 WO2024099187 A1 WO 2024099187A1
Authority
WO
WIPO (PCT)
Prior art keywords
configuration information
model
adjustable parameter
strategy
algorithm
Prior art date
Application number
PCT/CN2023/128628
Other languages
English (en)
French (fr)
Inventor
周通
袁雁南
孙鹏
Original Assignee
维沃移动通信有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 维沃移动通信有限公司 filed Critical 维沃移动通信有限公司
Publication of WO2024099187A1 publication Critical patent/WO2024099187A1/zh

Links

Definitions

  • the present application belongs to the field of communication technology, and specifically relates to an AI model strategy determination method, apparatus, first device and second device.
  • AI artificial intelligence
  • CSI channel state information
  • AI models are usually generated through offline training or online training.
  • the generated AI models are often only applicable to specific scenarios.
  • AI-based beam management it can be further divided into multiple implementation schemes, such as AI-based transmit and receive beam pair prediction, AI-based transmit beam prediction, and AI-based receive beam prediction.
  • Different models can be trained for a certain scheme. For example, complex network models have high reasoning accuracy but large size; while simple network models are small in size but low in reasoning accuracy.
  • AI models are used in the network, they often need to go through a large number of experiments and obtain sufficient verification data before they can be activated and reused in the existing network, resulting in low efficiency of the communication system.
  • the embodiments of the present application provide an AI model strategy determination method, apparatus, first device and second device, which can solve the problem of low operating efficiency of the communication system.
  • a method for determining an AI model strategy is provided, which is applied to a first device, and the method includes:
  • the first device obtains an adjustable parameter set and/or algorithm configuration information;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value of each adjustable parameter item;
  • the algorithm configuration information is used to indicate the configuration parameters of the target algorithm;
  • N is a positive integer;
  • the first device determines a target strategy for the AI function based on the set of adjustable parameters and/or the algorithm configuration information; the target strategy includes at least one of the following: an AI model deployment strategy; an AI model deactivation strategy; an AI model activation strategy; an AI model training strategy;
  • the first device processes the corresponding AI model according to the target strategy to provide AI services for the terminal.
  • an AI model strategy determination device comprising:
  • An acquisition module used to acquire an adjustable parameter set and/or algorithm configuration information;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value of each adjustable parameter item;
  • the algorithm configuration information is used to indicate the configuration parameters of the target algorithm; N is a positive integer;
  • a determination module configured to determine a target strategy for the AI function based on the set of adjustable parameters and/or the algorithm configuration information; the target strategy includes at least one of the following: an AI model deployment strategy; an AI model deactivation strategy; an AI model activation strategy; an AI model training strategy;
  • the processing module is used to process the corresponding AI model according to the target strategy to provide AI services for the terminal.
  • an AI model strategy determination method is provided, which is applied to a second device, and the method includes:
  • the second device sends any of the following to the first device:
  • a set of adjustable parameters and algorithm configuration information A set of adjustable parameters and algorithm configuration information
  • the adjustable parameter set is used to assist in determining the target strategy for the AI function;
  • the adjustable parameter set includes The invention comprises: N adjustable parameter items and at least one value of each adjustable parameter item, where N is a positive integer;
  • the algorithm configuration information is used to indicate configuration parameters of the target algorithm, and the algorithm configuration information is used to assist in determining a target strategy for the AI function;
  • the adjustable parameter configuration information is used to assist in determining the adjustable parameter set;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items;
  • the target strategy includes at least one of the following: AI model deployment strategy; AI model deactivation strategy; AI model activation strategy; AI model training strategy.
  • an AI model strategy determination device comprising:
  • a sending module configured to send any of the following items to the first device:
  • a set of adjustable parameters and algorithm configuration information A set of adjustable parameters and algorithm configuration information
  • the adjustable parameter set is used to assist in determining a target strategy for an AI function;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value for each adjustable parameter item, where N is a positive integer;
  • the algorithm configuration information is used to indicate configuration parameters of the target algorithm, and the algorithm configuration information is used to assist in determining a target strategy for the AI function;
  • the adjustable parameter configuration information is used to assist in determining the adjustable parameter set;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items;
  • the target strategy includes at least one of the following: AI model deployment strategy; AI model deactivation strategy; AI model activation strategy; AI model training strategy.
  • a first device which includes a processor and a memory, wherein the memory stores a program or instruction that can be executed on the processor, and when the program or instruction is executed by the processor, the steps of the method described in the first aspect are implemented.
  • a first device comprising a processor and a communication interface, wherein the processor is used to obtain an adjustable parameter set and/or algorithm configuration information; the adjustable parameter set comprises: N adjustable parameter items and at least one value for each adjustable parameter item; the algorithm configuration information is used to indicate the configuration parameters of the target algorithm; N is a positive integer; the processor is also used to determine a target strategy for the AI function based on the adjustable parameter set and/or the algorithm configuration information; the target strategy comprises at least one of the following: AI model deployment strategy; AI model deactivation strategy; AI model activation strategy; AI model training strategy; the processor is also used to: process the corresponding AI model according to the target strategy to provide AI services for the terminal.
  • a second device comprising a processor and a memory, wherein the memory stores a program or instruction that can be run on the processor, and when the program or instruction is executed by the processor, the steps of the method described in the third aspect are implemented.
  • a second device including a processor and a communication interface, wherein the communication interface is used to send any of the following items to the first device:
  • a set of adjustable parameters and algorithm configuration information A set of adjustable parameters and algorithm configuration information
  • the adjustable parameter set is used to assist in determining a target strategy for an AI function;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value for each adjustable parameter item, where N is a positive integer;
  • the algorithm configuration information is used to indicate configuration parameters of the target algorithm, and the algorithm configuration information is used to assist in determining a target strategy for the AI function;
  • the adjustable parameter configuration information is used to assist in determining the adjustable parameter set;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items;
  • the target strategy includes at least one of the following: AI model deployment strategy; AI model deactivation strategy; AI model activation strategy; AI model training strategy.
  • an AI model strategy determination system comprising: a first device and a second device, wherein the first device can be used to execute the steps of the AI model strategy determination method as described in the first aspect, and the second device can be used to execute the steps of the AI model strategy determination method as described in the third aspect.
  • a readable storage medium on which a program or instruction is stored.
  • the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented, or the steps of the method described in the third aspect are implemented.
  • a chip comprising a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run a program or instruction to implement the method described in the first aspect, or to implement the method described in the third aspect.
  • a computer program/program product is provided, wherein the computer program/program product is stored in a storage medium, and the computer program/program product is executed by at least one processor to implement the steps of the AI model strategy determination method as described in the first aspect, or to implement the steps of the AI model strategy determination method as described in the third aspect.
  • the first device after the first device obtains the adjustable parameter set and/or algorithm configuration information, it can determine the target strategy for the AI function based on the adjustable parameter set and/or algorithm configuration information, thereby improving the efficiency of determining the target strategy.
  • the first device then processes the corresponding AI model according to the target strategy, provides AI services to the terminal, and can improve the operating efficiency of the communication system.
  • FIG1 is a block diagram of a wireless communication system to which an embodiment of the present application can be applied;
  • FIG2 is a flowchart of a method for determining an AI model strategy according to an embodiment of the present application
  • FIG3 is a second flow chart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG4 is a third flow chart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG5 is a fourth flowchart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG6 is a fifth flowchart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG7 is a sixth flowchart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG8 is a seventh flowchart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG9 is a flowchart of an eighth method for determining an AI model strategy according to an embodiment of the present application.
  • FIG10 is a ninth flowchart of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG11 is a signaling interaction diagram of the AI model strategy determination method provided in an embodiment of the present application.
  • FIG12 is a schematic diagram of a process flow of lifecycle management of an AI model provided in an embodiment of the present application.
  • FIG13 is a schematic diagram of a structure of an AI model strategy determination device according to an embodiment of the present application.
  • FIG14 is a second structural diagram of the AI model strategy determination device provided in an embodiment of the present application.
  • FIG15 is a schematic diagram of the structure of a communication device provided in an embodiment of the present application.
  • FIG16 is one of the structural schematic diagrams of the first device provided in an embodiment of the present application.
  • FIG17 is a second structural schematic diagram of the first device provided in an embodiment of the present application.
  • FIG18 is one of the structural schematic diagrams of the second device provided in an embodiment of the present application.
  • FIG. 19 is a second schematic diagram of the structure of the second device provided in an embodiment of the present application.
  • first, second, etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by “first” and “second” are generally of the same type, and the number of objects is not limited.
  • the first object can be one or more.
  • “and/or” in the specification and claims represents at least one of the connected objects, and the character “/" generally represents that the objects associated with each other are in an "or” relationship.
  • LTE Long Term Evolution
  • LTE-A Long Term Evolution-Advanced
  • CDMA code division multiplexing
  • CDMA Code Division Multiple Access
  • TDMA Time Division Multiple Access
  • FDMA Frequency Division Multiple Access
  • OFDMA Orthogonal Frequency Division Multiple Access
  • SC-FDMA Single-carrier Frequency Division Multiple Access
  • NR New Radio
  • 6G 6th Generation
  • FIG1 is a block diagram of a wireless communication system applicable to an embodiment of the present application.
  • the wireless communication system includes a terminal 11 and a network side device 12 .
  • the terminal 11 can be a mobile phone, a tablet computer (Tablet Personal Computer), a laptop computer (Laptop Computer) or a notebook computer, a personal digital assistant (Personal Digital Assistant, PDA), a handheld computer, a netbook, an ultra-mobile personal computer (ultra-mobile personal computer, UMPC), a mobile Internet device (Mobile Internet Device, MID), augmented reality (augmented reality, AR)/virtual reality (virtual reality, VR) equipment, a robot, a wearable device (Wearable Device), a vehicle-mounted device (VUE), a pedestrian terminal (PUE), a smart home (home appliances with wireless communication functions, such as refrigerators, televisions, washing machines or furniture, etc.), a game console, a personal computer (personal computer, PC), an ATM or a self-service machine and other terminal side devices, and the wear
  • the network side device 12 may include an access network device or a core network device, wherein the access network device 12 may also be referred to as a radio access network device, a radio access network (RAN), a radio access network function or a radio access network unit.
  • the access network device 12 may include a base station, a WLAN access point or a WiFi node, etc.
  • the base station may be referred to as a node B, an evolved node B (eNB), an access point, a base transceiver station (BTS), a radio base station, a radio transceiver, a basic service set (BSS), an extended service set (ESS), a home B node, a home evolved B node, a transmitting and receiving point (TRP) or other appropriate terms in the field, as long as the same technical effect is achieved, the base station is not limited to a specific technical vocabulary, it should be noted that in the embodiment of the present application, only the base station in the NR system is used as an example for introduction, and the specific type of the base station is not limited.
  • the core network equipment may include but is not limited to at least one of the following: core network nodes, core network functions, mobility management entity (Mobility Management Entity, MME), access mobility management function (Access and Mobility Management Function, AMF), session management function (Session Management Function, SMF), user plane function (User Plane Function, UPF), policy control function (Policy Control Function, PCF), policy and charging rules function unit (Policy and Charging Rules Function, PCRF), edge application service discovery function (Edge Application Server Discovery ...
  • MME mobility management entity
  • AMF Access and Mobility Management Function
  • SMF Session Management Function
  • SMF Session Management Function
  • UPF User Plane Function
  • Policy Control Function Policy Control Function
  • PCRF Policy and Charging Rules Function
  • edge application service discovery function Edge Application Server Discovery ...
  • the current mainstream methods include black-box optimization algorithms and multi-fidelity optimization.
  • the classic methods of black-box optimization algorithms include grid search, random search, Bayesian optimization, etc.
  • the classic methods of multi-fidelity optimization algorithms include successive halving and Hyperband algorithm.
  • NAS model network architecture search
  • the current hyperparameter optimization problem mainly solves the problem of how to select the basic model, learning rate, batch size, optimization function and other parameters during model training.
  • the network model search algorithm solves the problem of how to select the network structure of the AI model. The above two methods cannot be directly used for the deployment strategy selection, model activation condition selection, model deactivation condition selection and other issues in the AI use case experiment phase in wireless networks.
  • the AI model strategy determination method provided in the embodiment of the present application can be applied to a first device that needs to determine a target strategy for an AI function.
  • the target strategy may include at least one of the following: an AI model deployment strategy, an AI model deactivation strategy, an AI model activation strategy, and an AI model training strategy.
  • FIG. 2 is one of the flowcharts of the AI model strategy determination method provided in an embodiment of the present application. As shown in FIG. 2 , the method includes step 201 and step 202; wherein:
  • Step 201 The first device obtains an adjustable parameter set and/or algorithm configuration information; the adjustable parameter set includes: N adjustable parameter items and at least one value of each adjustable parameter item; the algorithm configuration information is used to indicate the configuration parameters of the target algorithm; N is a positive integer;
  • Step 202 The first device determines a target strategy for the AI function based on the adjustable parameter set and/or the algorithm configuration information; the target strategy includes at least one of the following: an AI model deployment strategy; an AI model deactivation strategy; an AI model activation strategy; an AI model training strategy;
  • Step 203 The first device processes the corresponding AI model according to the target strategy to provide AI services for the terminal.
  • the first device may obtain an adjustable parameter set and/or algorithm configuration information, wherein the adjustable parameter set is used to assist in determining a target strategy for an AI function, and the adjustable parameter set includes: N adjustable parameter items and at least one value for each adjustable parameter item; the algorithm configuration information is used to indicate configuration parameters of a target algorithm;
  • the first device can determine a target strategy including at least one of an AI model deployment strategy, an AI model deactivation strategy, an AI model activation strategy, and an AI model training strategy based on the adjustable parameter set and/or the algorithm configuration information, so as to process the corresponding AI model according to the target strategy and provide AI services to the terminal.
  • the experimental phase of the AI model can be automated according to the target strategy.
  • the embodiments of the present application can effectively improve the online efficiency of the AI model, thereby improving the performance of providing AI services to the terminal.
  • the AI model is also optional.
  • the target strategy can select one or more AI models from the set 10 AI models to execute the automation algorithm corresponding to the algorithm configuration information.
  • AI functions include, for example, AI-based beam management, AI-based CSI channel compression feedback, AI-based positioning, AI-based base station energy saving, and AI-based load balancing.
  • the first device may include at least one of the following:
  • the entity at the same level as the base station CU may be a newly added entity at the same level as the base station CU.
  • Base station distributed unit (Distributed Unit, DU).
  • the first device after the first device obtains the adjustable parameter set and/or algorithm configuration information, it can determine the target strategy for the AI function based on the adjustable parameter set and/or algorithm configuration information, thereby improving the efficiency of determining the target strategy. Then, the first device processes the corresponding AI model according to the target strategy, provides AI services to the terminal, and can improve the operating efficiency of the communication system.
  • the implementation manner in which the first device acquires the adjustable parameter set and/or the algorithm configuration information may include at least one of the following:
  • the first device receives the adjustable parameter set and/or the algorithm configuration information from the second device;
  • Case 1 The first device receives a set of adjustable parameters from the second device, and the algorithm configuration information is predefined;
  • Case 2 The first device receives algorithm configuration information from the second device, and the adjustable parameter set is predefined;
  • Case 3 The first device receives the adjustable parameter set and algorithm configuration information from the second device.
  • both the adjustable parameter set and the algorithm configuration information may be predefined.
  • the first device receives the adjustable parameter configuration information and/or the algorithm configuration information from the second device, and the first device determines the adjustable parameter set based on the adjustable parameter configuration information;
  • Case 1 The first device receives adjustable parameter configuration information from the second device, the algorithm configuration information is predefined, and the first device determines the adjustable parameter set based on the adjustable parameter configuration information;
  • Case 2 The first device receives algorithm configuration information from the second device, the adjustable parameter configuration information is predefined, and the first device determines the adjustable parameter set based on the adjustable parameter configuration information;
  • Case 3 The first device receives adjustable parameter configuration information and algorithm configuration information from the second device, and the first device determines an adjustable parameter set based on the adjustable parameter configuration information.
  • both the adjustable parameter configuration information and the algorithm configuration information may be predefined.
  • the first device receives the adjustable parameter set from the second device; the first device receives the algorithm configuration information from the third device;
  • the first device receives the adjustable parameter configuration information from the second device, and the first device determines the adjustable parameter set based on the adjustable parameter configuration information; the first device receives the algorithm configuration information from the third device;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items.
  • the adjustable parameter configuration information may include:
  • the second device may include at least one of the following:
  • the entity at the same level as the base station CU may be a newly added entity at the same level as the base station CU.
  • the first device may determine, based on the adjustable parameter configuration information, an implementation manner of the adjustable parameter set that may include at least one of the following:
  • the first device determines the adjustable parameter configuration information as the adjustable parameter set
  • the first device directly obtains the adjustable parameter set from the second device; that is, the second device provides the adjustable parameter set, wherein the adjustable parameter set includes all adjustable parameter items and a discrete value set of each adjustable parameter item.
  • the first device determines at least one value of each adjustable parameter item based on the parameter item configuration information; the first device determines the adjustable parameter set based on the adjustable parameter configuration information and at least one value of each adjustable parameter item;
  • the first device obtains adjustable parameter configuration information from the second device, where the adjustable parameter configuration information includes N
  • the first device can first determine at least one value of each of the N adjustable parameter items based on the parameter item configuration information, and then determine the adjustable parameter set based on the adjustable parameter configuration information and at least one value of each of the N adjustable parameter items; that is, the adjustable parameter configuration information provided by the second device includes all adjustable parameter items, and the first device determines the discrete value set of each adjustable parameter item.
  • the first device determines at least one value of the remaining adjustable parameter items among the N adjustable parameter items based on the parameter item configuration information; the first device determines the adjustable parameter set based on the adjustable parameter configuration information and at least one value of the remaining adjustable parameter items;
  • the first device obtains adjustable parameter configuration information from the second device.
  • the adjustable parameter configuration information includes N adjustable parameter items and at least one value of some of the N adjustable parameter items
  • the first device can first determine at least one value of the remaining adjustable parameter items among the N adjustable parameter items based on the parameter item configuration information, and then determine the adjustable parameter set based on the adjustable parameter configuration information and at least one value of the remaining adjustable parameter items; that is, the adjustable parameter configuration information provided by the second device includes the value sets of all parameter items and some parameter items, and the discrete value sets of the remaining parameter items are determined by the first device.
  • the parameter item configuration information may include an empirical value or a default value.
  • the first device determines at least one value of the first adjustable parameter item and the adjustable parameter item without a value among the N adjustable parameter items based on the parameter item configuration information; the first device determines the adjustable parameter set based on the adjustable parameter configuration information, the first adjustable parameter item and at least one value of the adjustable parameter item without a value; the first adjustable parameter item is the adjustable parameter item among the N adjustable parameter items excluding the some of the adjustable parameter items.
  • the first device obtains adjustable parameter configuration information from the second device.
  • the adjustable parameter configuration information includes some adjustable parameter items and at least one value of at least one adjustable parameter item in the some adjustable parameter items
  • the first device can first determine at least one value of the first adjustable parameter item and the adjustable parameter items without values among the N adjustable parameter items based on the parameter item configuration information.
  • the first adjustable parameter item is an adjustable parameter item among the N adjustable parameter items other than some adjustable parameter items.
  • the first device determines an adjustable parameter set based on the adjustable parameter configuration information, the first adjustable parameter item and at least one value of the adjustable parameter items without values among the N adjustable parameter items; that is, the adjustable parameter configuration information provided by the second device includes some adjustable parameter items and at least one corresponding discrete value set, and the first device determines the remaining adjustable parameter items among the N adjustable parameter items and the discrete value set corresponding to the adjustable parameter items without values.
  • the adjustable parameter configuration information provided by the second device includes some adjustable parameter items and corresponding discrete value sets, and the first device determines the remaining adjustable parameter items and corresponding discrete value sets among the N adjustable parameter items.
  • the adjustable parameter configuration information may include at least one of the following:
  • First configuration information used to assist in determining the AI model deployment strategy
  • the first configuration information may include at least one of the following:
  • the network configuration information may include at least one of the following:
  • the terminal configuration information may include at least one of the following:
  • the network scenario information may include at least one of the following:
  • different rainfall scene information may include information for indicating no, light, medium, or heavy rainfall.
  • the terminal scenario information may include at least one of the following:
  • the candidate solutions corresponding to the AI function include, for example, an AI-based beam management solution, including a base station-side beam prediction solution, a terminal-side receiving beam prediction solution, and a transmitting and receiving beam prediction solution; another example is a candidate solution corresponding to AI-based channel compression feedback.
  • Second configuration information used to assist in determining the AI model activation strategy
  • the second configuration information may include at least one of the following:
  • monitoring indicators include, for example, optimal beam prediction accuracy, prediction error, cosine similarity, throughput, etc.
  • the first parameter may include at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is activated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to activate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that when the count value of the second counter is greater than the second quantity threshold, the first counter is reset or the first timer is paused; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • periodic reset can be a fixed 10,000 samples (sample period) or a fixed 1,000 milliseconds (time period) to reset the first counter; while for timer reset, the timer can be stopped and started midway.
  • the period of the first timer is 10,000 samples or 1,000 ms
  • condition 1 when condition 1 is met, the first timer starts; when condition 2 is met, the first timer stops; when the first timer times out, the AI model is activated.
  • Condition 1 The first counter reaches the first quantity threshold.
  • the first counter increments, and the indicator characterizing the AI model (such as the prediction accuracy of the optimal beam) is higher than the threshold (such as 90%), and the prediction result is good.
  • the reset condition of the first counter is condition 2.
  • Condition 2 The second counter reaches the second quantity threshold.
  • the second counter increments, and the indicator characterizing the AI model (e.g., the prediction accuracy of the optimal beam) is lower than the threshold (e.g., 85%), and the prediction result is not good.
  • the reset condition of the second counter is condition 1.
  • the period of the first timer is 10,000 samples or 1,000 ms
  • condition 1 when condition 1 is met, the first timer starts; when condition 2 is met, the first timer stops and the AI model is activated; when condition 3 is met, the first counter is reset; when the first timer times out, the AI model remains deactivated.
  • Condition 1 Start the model activation evaluation process.
  • Condition 2 The first counter reaches the first quantity threshold.
  • the first counter increments, and the indicator representing the AI model (such as the prediction accuracy of the optimal beam) is higher than the threshold (such as 90%), and the prediction result is good.
  • the reset of the first counter is condition 3.
  • Condition 3 The second counter reaches the second quantity threshold.
  • the second counter increments, and the indicator representing the AI model (such as the prediction accuracy of the optimal beam) is lower than the threshold (such as 85%), and the prediction result is not good.
  • the second counter is reset as condition 2.
  • the third configuration information may include at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is deactivated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to deactivate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that the first counter is reset or the first timer is paused when the count value of the second counter is greater than the second quantity threshold; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the fourth configuration information is used to assist in determining the AI model training strategy.
  • the fourth configuration information may include at least one of the following:
  • the algorithm configuration information may include at least one of the following:
  • the reinforcement learning algorithm configuration information may include at least one of the following:
  • the entity that collects the reward is, for example, a reward collection location.
  • the grid search algorithm configuration information may include at least one of the following:
  • the total search space of AI model deployment strategies for verification can be;
  • the total search space of deactivation strategies for AI models used for verification can be;
  • the total search space of strategies for the AI model activation used for verification can be;
  • it can be the total search space of AI model training strategies used for verification.
  • the random search algorithm configuration information may include at least one of the following:
  • it can be the maximum number of AI model training strategies used for verification.
  • the continuous halving algorithm configuration information includes at least one of the following:
  • the number of strategies deployed for AI models used for verification and the verification duration of each strategy can be determined
  • the number of deactivation strategies for the AI model used for verification and the verification duration of each strategy can be determined
  • the number of strategies activated for the AI model used for verification and the verification duration of each strategy can be set;
  • it can be the number of AI model training strategies used for verification and the verification time of each strategy.
  • the Hyperband algorithm configuration information may include at least one of the following:
  • the number of executions of the continuous halving algorithm for the AI model deployment strategy used for verification, as well as the number of strategies and verification duration in the initial continuous halving algorithm can be calculated;
  • the number of executions of the continuous halving algorithm for the AI model activation strategy used for verification, as well as the number of strategies and verification duration in the initial continuous halving algorithm can be used;
  • it can be the number of executions of the continuous halving algorithm of the AI model training strategy used for verification, as well as the number of strategies and verification time in the initial continuous halving algorithm.
  • the AI model strategy determination method provided in the embodiment of the present application can be applied to a second device that sends at least one of an adjustable parameter set, adjustable parameter configuration information, and algorithm configuration information to a first device.
  • FIG. 3 is a second flow chart of the AI model strategy determination method provided in an embodiment of the present application. As shown in FIG. 3 , the method includes step 301; wherein:
  • Step 301 The second device sends any one of the following to the first device: an adjustable parameter set and algorithm configuration information; adjustable parameter configuration information and algorithm configuration information; an adjustable parameter set; adjustable parameter configuration information; algorithm configuration information;
  • the adjustable parameter set is used to assist in determining a target strategy for an AI function;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value for each adjustable parameter item, where N is a positive integer;
  • the algorithm configuration information is used to indicate configuration parameters of the target algorithm, and the algorithm configuration information is used to assist in determining a target strategy for the AI function;
  • the adjustable parameter configuration information is used to assist in determining the adjustable parameter set;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items;
  • the target strategy includes at least one of the following:
  • the second device may send any of the following items to the first device: an adjustable parameter set and algorithm configuration information; adjustable parameter configuration information and algorithm configuration information; an adjustable parameter set; adjustable parameter configuration information; algorithm configuration information; so that the first device obtains the adjustable parameter set and/or algorithm configuration information based on any of the above items, and then determines the target strategy for the AI function based on the adjustable parameter set and/or algorithm configuration information.
  • the first device may include at least one of the following:
  • Base station DU Base station DU.
  • the second device may include at least one of the following:
  • the second device sends any of the following items to the first device: an adjustable parameter set and algorithm configuration information; adjustable parameter configuration information and algorithm configuration information; adjustable parameter set; adjustable parameter configuration information; algorithm configuration information; after the first device obtains the adjustable parameter set and/or algorithm configuration information, the target strategy for the AI function is determined based on the adjustable parameter set and/or algorithm configuration information, thereby improving the efficiency of determining the target strategy, and then the first device processes the corresponding AI model according to the target strategy, provides AI services to the terminal, and can improve the operating efficiency of the communication system.
  • the adjustable parameter configuration information may include at least one of the following:
  • First configuration information used to assist in determining the AI model deployment strategy
  • Second configuration information used to assist in determining the AI model activation strategy
  • the fourth configuration information is used to assist in determining the AI model training strategy.
  • the first configuration information may include at least one of the following:
  • the network configuration information may include at least one of the following:
  • the terminal configuration information may include at least one of the following:
  • the network scenario information may include at least one of the following:
  • the terminal scenario information may include at least one of the following:
  • the second configuration information includes at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is activated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to activate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that the first counter is reset or the first timer is paused when the count value of the second counter is greater than the second quantity threshold; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the third configuration information may include at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is deactivated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to deactivate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that the first counter is reset or the first timer is paused when the count value of the second counter is greater than the second quantity threshold; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the first parameter may include at least one of the following:
  • the fourth configuration information may include at least one of the following:
  • the algorithm configuration information may include at least one of the following:
  • the reinforcement learning algorithm configuration information may include at least one of the following:
  • the grid search algorithm configuration information may include at least one of the following:
  • the random search algorithm configuration information may include at least one of the following:
  • the continuous halving algorithm configuration information may include at least one of the following:
  • the Hyperband algorithm configuration information may include at least one of the following:
  • the following example illustrates the AI model strategy determination method provided in the embodiments of the present application.
  • Embodiment 1 A set of adjustable parameters for selecting an AI model deployment strategy.
  • Table 1 shows the set of adjustable parameters for AI model deployment strategy selection.
  • Table 1 A set of adjustable parameters for AI model deployment strategy selection
  • the first device After the first device executes the automation algorithm, it generates an AI model deployment strategy to be verified. In each deployment strategy, a value is selected for each adjustable parameter item.
  • the number of terminal beams is 8, and 8 is selected from ⁇ 2, 4, 8, 16 ⁇ ;
  • the number of terminal antenna panels is 2, and 2 is selected from ⁇ 1, 2, 3 ⁇ ;
  • Line-of-sight and non-line-of-sight scenes are line-of-sight scenes, and the line-of-sight scene is selected from ⁇ line-of-sight scene, mixed scene ⁇ ;
  • the terminal moving speed is 30km/h, select 30km/h from ⁇ 30km/h, 60km/h, 90km/h ⁇ ;
  • the AI model used for reasoning is the model with model ID 2, and model ID 2 is selected from ⁇ model ID 1, model ID 2 ⁇ .
  • the first device deploys the AI model deployment strategy to be verified to the corresponding base station and terminal, and then verifies the performance of this AI model deployment strategy.
  • Embodiment 2 A set of adjustable parameters for selecting an AI model deactivation strategy.
  • Table 2 shows a set of adjustable parameters for AI model deactivation strategy selection.
  • Table 2 A set of adjustable parameters for AI model deactivation strategy selection
  • the first device After the first device executes the automation algorithm, it generates a deactivation strategy for the AI model to be verified. In each strategy, a value is selected for each adjustable parameter item.
  • the indication information for indicating whether the monitoring indicator is activated is 1, selected from ⁇ 0, 1 ⁇ , indicating that this monitoring indicator is activated.
  • the indicator threshold is 97%, selected from ⁇ 90%, 95%, 97% ⁇ ;
  • the sample period is 1000, selected from ⁇ 1000, 2000, 5000 ⁇ ;
  • the sample proportion is 5%, selected from ⁇ 5%, 10%, 20%, 100% ⁇ ;
  • the sample distribution is a pre-distribution, selected from ⁇ random distribution, pre-distribution, post-distribution ⁇ .
  • the pre-distribution means that from 1000 samples, the first 50 samples are selected for monitoring and the average accuracy of the optimal beam prediction is calculated.
  • the first quantity threshold (event activation quantity threshold) is 2, selected from ⁇ 1, 2, 4 ⁇ . In this example, if the average accuracy calculated from the first 50 samples out of 1000 samples is less than 97%, the first counter (event activation counter) is accumulated to 1; when the first counter is accumulated to 2, the AI model is deactivated.
  • the reset mode (counter reset rule) of the first counter is event reset, which is selected from ⁇ periodic reset, event reset ⁇ .
  • event reset means that the first counter can be reset after a specific event is met.
  • the reset threshold is 99%, selected from ⁇ 99%, 100% ⁇ . In this example, it means that if the average accuracy of the first 50 samples out of 1000 samples is greater than 99%, the cumulative reset counter (second counter) can be increased by 1;
  • the reset counter threshold is 2, selected from ⁇ 1, 2, 4 ⁇ , indicating that when the second counter (reset counter) accumulates to 2, the first counter can be reset.
  • the indication information for indicating whether the monitoring indicator is activated is 0, selected from ⁇ 0, 1 ⁇ , indicating that this monitoring indicator is not considered.
  • the indication information for indicating whether the monitoring indicator is activated is 0, selected from ⁇ 0, 1 ⁇ , indicating that this monitoring indicator is not considered.
  • the indication information for indicating whether the monitoring indicator is activated is 1, selected from ⁇ 0, 1 ⁇ , indicating that this monitoring indicator is activated.
  • the indicator threshold is 20%, selected from ⁇ 10%, 20%, 30% ⁇ ;
  • the sample period is 1000, selected from ⁇ 1000, 2000, 5000 ⁇ ;
  • the sample proportion is 5%, selected from ⁇ 5%, 10%, 20%, 100% ⁇ ;
  • the sample distribution is a pre-distribution, selected from ⁇ random distribution, pre-distribution, post-distribution ⁇ .
  • the pre-distribution means that from 1000 samples, the first 50 samples are selected for monitoring, and the gain of the average throughput compared with the baseline scheme is calculated.
  • the first quantity threshold (event activation quantity threshold) is 1, selected from ⁇ 1, 2, 4 ⁇ . In this example, if the gain of the average throughput calculated from the first 50 samples out of 1000 samples compared with the baseline solution is less than 20%, the first counter is accumulated to 1; when the first counter is accumulated to 1, the AI model is deactivated.
  • the reset mode (counter reset rule) of the first counter is periodic reset, which is selected from ⁇ periodic reset ⁇ . At this time, only one optional item is configured, and periodic reset is fixedly selected.
  • the reset period (counter reset period) of the periodic reset is 40000, selected from ⁇ 10000, 20000, 40000 ⁇ , which means in this example that if the first counter does not reach the first quantity threshold within 40000, the first counter is reset to 0.
  • the first device deploys the AI model deactivation strategy to be verified to the corresponding functional entity, and then verifies the performance of this strategy.
  • Embodiment 3 The first device is a base station DU, and the second device is a base station CU.
  • the first device is a base station DU
  • the second device is a base station CU
  • the automation algorithm corresponding to the algorithm configuration information is a reinforcement learning algorithm
  • Figure 4 is a flow chart of the third AI model strategy determination method provided in an embodiment of the present application, as shown in Figure 4.
  • Step 1 The base station DU receives the algorithm configuration information and adjustable parameter configuration information sent by the base station CU.
  • the algorithm configuration information includes the entity user equipment (UE) for collecting rewards for reinforcement learning, the reinforcement learning hyperparameter discount factor, and the number of training times;
  • UE entity user equipment
  • Adjustable parameter configuration information includes network configuration information, terminal configuration information, scenario configuration information (including network scenario information and terminal scenario information), candidate AI model information, and corresponding value sets.
  • Step 2 The base station DU generates a set of adjustable parameters and executes the automated algorithm
  • Step 3 The base station DU determines the activation conditions of the base station and UE based on the AI model strategy output by the automation algorithm;
  • Step 4 Assuming that the AI model is performing UE inference, the base station DU sends the activation condition to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assume that UE monitors key performance indicators (KPIs), and UE feeds back KPI results to DU.
  • KPIs include optimal beam prediction accuracy, beam management loss, and throughput.
  • Step 7 DU converts the KPI result into reward and updates the network parameters of its own controller.
  • Embodiment 4 The first device is a base station DU, and the second device is a newly added node of the base station CU, serving as a node parallel to the base station CU. level entity.
  • the first device is the base station DU
  • the second device is a newly added node of the base station CU
  • the automation algorithm corresponding to the algorithm configuration information is a random search algorithm.
  • FIG5 is a fourth flow chart of the AI model strategy determination method provided in an embodiment of the present application, as shown in FIG5 .
  • Step 1 The base station DU receives the algorithm configuration information and adjustable parameter configuration information sent by the newly added node of the base station CU.
  • the algorithm configuration information includes the maximum number of random searches; the adjustable parameter configuration information includes base station configuration, terminal configuration and corresponding value sets.
  • Step 2 The base station DU generates a set of adjustable parameters and executes an automated algorithm.
  • the base station DU determines the scenario configuration information and the candidate AI model information, as well as the corresponding value set, as an adjustable parameter set.
  • Step 3 The strategy output by the base station DU automation algorithm determines the activation conditions of the base station and UE;
  • Step 4 Assuming that the AI model is performing UE inference, the base station DU sends the activation condition to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assume that the UE performs KPI monitoring and feeds back the KPI results to the DU.
  • the KPIs include the optimal beam prediction accuracy, beam management loss, and throughput.
  • Step 7 DU records the current strategy and the corresponding KPI, converts the KPI result into reward, and updates the network parameters of its own controller.
  • Embodiment 5 The first device is a base station CU, and the second device is a newly added node of the base station CU, serving as an entity at the same level as the base station CU.
  • the first device is the base station CU
  • the second device is a newly added node of the base station CU
  • the automation algorithm corresponding to the algorithm configuration information is a grid search algorithm.
  • FIG6 is a fifth flow chart of the AI model strategy determination method provided in an embodiment of the present application, as shown in FIG6 .
  • Step 1 The base station CU receives the algorithm configuration information and adjustable parameter configuration information sent by the newly added node of the base station CU.
  • the algorithm configuration information can be the same as the adjustable parameter configuration information, because for the grid search algorithm, the total search space of the target strategy can be the same as the value set of the adjustable parameter item.
  • the adjustable parameter configuration information includes base station configuration information, terminal configuration information, scenario configuration information, candidate AI model information, and the corresponding value set.
  • Step 2 The base station CU generates a set of adjustable parameters and executes the automated algorithm
  • Step 3 The base station CU determines the activation conditions of the base station and UE based on the AI model strategy output by the automation algorithm;
  • Step 4 Assuming that the AI model is performing UE inference, the base station CU sends the activation condition to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assuming that the UE performs KPI monitoring, the UE feeds back the KPI results to the CU.
  • KPIs include the best beam prediction accuracy, beam management loss, and throughput.
  • Step 7 CU records this strategy and the corresponding KPI.
  • CU selects the strategy with the best KPI based on the recorded strategies and corresponding KPIs.
  • Embodiment 6 The first device is a base station DU, and the second device is a core network (CN) function.
  • CN core network
  • the first device is a base station DU
  • the second device is a core network function
  • the automation algorithm corresponding to the algorithm configuration information is a continuous halving algorithm.
  • FIG. 7 is a sixth flowchart of the AI model strategy determination method provided in an embodiment of the present application, as shown in FIG. 7 .
  • Step 1 The base station DU receives the algorithm configuration information and adjustable parameter configuration information sent by the core network function through the CU.
  • the CU may be a base station CU, or an entity at the same level as the base station CU (eg, a newly added node of the base station CU).
  • the algorithm configuration information includes the number of target strategies and the verification time of each strategy.
  • the adjustable parameter configuration information includes base station configuration information, terminal configuration information, scenario configuration information, candidate AI model information, and a set of values for each adjustable parameter item.
  • Step 2 The base station DU generates a set of adjustable parameters and executes the automated algorithm.
  • Step 3 The base station DU determines the activation conditions of the base station and UE based on the AI model strategy output by the automation algorithm;
  • Step 4 Assuming that the AI model is inferring in the UE, the base station DU sends the activation conditions to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assume that the UE performs KPI monitoring and the UE feeds back the KPI results to the DU.
  • the KPIs include the optimal beam prediction accuracy, beam management loss, and throughput.
  • Step 7 DU records this strategy and the corresponding KPI.
  • DU selects the strategy with the best KPI based on the recorded strategies and corresponding KPIs.
  • Embodiment 7 The first device is a base station CU, and the second device is a core network function.
  • the first device is a base station CU
  • the second device is a core network function
  • the automation algorithm corresponding to the algorithm configuration information is a grid search algorithm.
  • FIG8 is a seventh flow chart of the AI model strategy determination method provided in an embodiment of the present application, as shown in FIG8 .
  • Step 1 The base station CU receives the algorithm configuration information and adjustable parameter configuration information sent by the core network function.
  • the algorithm configuration information can be the same as the adjustable parameter configuration information, because for the grid search algorithm, the total search space of the target strategy can be the same as the value set of the adjustable parameter item.
  • the adjustable parameter configuration information includes base station configuration information, terminal configuration information, scenario configuration information, candidate AI model information, and the corresponding value set.
  • Step 2 The base station CU generates a set of adjustable parameters and executes the automated algorithm
  • Step 3 The base station CU determines the activation conditions of the base station and UE based on the AI model strategy output by the automation algorithm;
  • Step 4 Assuming that the AI model is performing UE inference, the base station CU sends the activation condition to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assume that the UE performs KPI monitoring and the UE feeds back the KPI results to the CU.
  • the KPIs include the optimal beam prediction accuracy, beam management loss, and throughput.
  • Step 7 CU records this strategy and the corresponding KPI.
  • CU selects the strategy with the best KPI based on the recorded strategies and corresponding KPIs.
  • Embodiment 8 The first device is a base station, and the second device is a core network function.
  • the first device is a base station
  • the second device is a core network function
  • the automation algorithm corresponding to the algorithm configuration information is a random search algorithm.
  • FIG9 is an eighth flowchart of the AI model strategy determination method provided in an embodiment of the present application, as shown in FIG9 .
  • Step 1 The base station receives the algorithm configuration information and adjustable parameter configuration information sent by the core network function.
  • the algorithm configuration information includes the maximum search number of the target strategy.
  • the adjustable parameter configuration information includes base station configuration information, terminal configuration information, scenario configuration information, candidate AI model information, and corresponding value sets.
  • Step 2 The base station generates a set of adjustable parameters and executes the automated algorithm
  • Step 3 The base station determines the activation conditions for the base station and UE based on the AI model strategy output by the automation algorithm;
  • Step 4 Assuming that the AI model is performing UE inference, the base station sends the activation condition to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assume that the UE performs KPI monitoring and feeds back the KPI results to the base station.
  • the KPIs include the optimal beam prediction accuracy, beam management loss, and throughput.
  • Step 7 The base station records this strategy and the corresponding KPI.
  • the base station selects the strategy with the best KPI based on the recorded strategies and corresponding KPIs.
  • Embodiment 9 The first device is a newly added node at the same level as the base station CU, and the second device is a core network function.
  • the first device is a newly added node at the same level as the base station CU
  • the second device is a core network function
  • the automation algorithm corresponding to the algorithm configuration information is the Hyperband algorithm.
  • FIG10 is a ninth flowchart of the AI model strategy determination method provided in an embodiment of the present application, as shown in FIG10 .
  • Step 1 The newly added node at the same level as the base station CU receives the algorithm configuration information and adjustable parameter configuration information sent by the core network function.
  • the algorithm configuration information includes the number of executions of the continuous halving algorithm of the target strategy, as well as the number of strategies and verification time in the initial continuous halving algorithm.
  • the adjustable parameter configuration information includes base station configuration information, terminal configuration information, field Scene configuration information, candidate AI model information, and corresponding value sets.
  • Step 2 The newly added node at the same level as the base station CU generates a set of adjustable parameters and executes the automated algorithm;
  • Step 3 The newly added node at the CU level of the base station determines the activation conditions of the base station and UE based on the AI model strategy output by the automated algorithm;
  • Step 4 Assuming that the AI model is performing UE inference, the newly added node at the base station CU level sends the activation condition to the UE;
  • Step 5 UE activates the AI model based on the terminal scenario information
  • the UE compares its own moving speed to see whether it matches the terminal moving speed configured in the terminal scene information, and determines whether to activate the AI model based on the matching result.
  • Step 6 Assuming that the UE performs KPI monitoring, the UE feeds back the KPI results to the newly added node at the same level as the CU.
  • the KPIs include the optimal beam prediction accuracy, beam management loss, and throughput.
  • Step 7 The newly added node at the same level as the base station CU records this strategy and the corresponding KPI.
  • the newly added nodes at the same level as the base station CU select the strategy with the best KPI based on the recorded strategies and corresponding KPIs.
  • Embodiment 10 The first device is a core network function, and the second device is a core network function.
  • the first device is a network function entity (Network Function, NF) 2 of a core network function
  • the second device is NF1 of another core network function
  • the automation algorithm corresponding to the algorithm configuration information is a grid search algorithm; the two devices can both be existing core network functions, or both are newly added core network functions, or one is a newly added core network function and the other is an existing core network function.
  • NF Network Function
  • FIG 11 is a signaling interaction diagram of the AI model strategy determination method provided in an embodiment of the present application. As shown in Figure 11, only step 1 is shown in the figure.
  • Step 1 NF2 receives the algorithm configuration information and adjustable parameter configuration information sent by NF1.
  • the algorithm configuration information can be the same as the adjustable parameter configuration information, because for the grid search algorithm, the total search space of the target strategy can be the same as the value set of the adjustable parameter item.
  • the adjustable parameter configuration information includes base station configuration information, terminal configuration information, scenario configuration information, candidate AI model information, and the corresponding value set.
  • Step 2 NF2 generates a set of adjustable parameters and executes the automated algorithm
  • Step 3 NF2 determines the activation conditions of the base station and UE based on the AI model strategy output by the automation algorithm;
  • Step 4 Assuming that the AI model is performing UE inference, NF2 sends the activation conditions to the UE;
  • Step 5 UE activates the AI model according to the scenario
  • Step 6 Assume that the UE performs KPI monitoring and the UE feeds back the KPI results to NF2. KPIs include the best beam prediction accuracy, beam management loss, and throughput.
  • Step 7 NF2 records this strategy and the corresponding KPI.
  • NF2 selects the strategy with the best KPI based on the recorded strategies and corresponding KPIs.
  • Example 11 Schematic diagram of AI model lifecycle management process.
  • FIG12 is a flowchart of the lifecycle management of the AI model provided by the embodiment of the present application.
  • the AI model lifecycle management studied in the existing 5G system includes: model training (steps 21 to 27 in the figure), model deployment (steps 3 to 6 in the figure), model transfer (step 5 in the figure), model activation (steps 7 and 8 in the figure), model reasoning (steps 9 and 10 in the figure), model monitoring (steps 11 to 16 in the figure), model deactivation (steps 17 to 18 in the figure), model rollback (step 20 in the figure), model switching (step 19 in the figure), etc.
  • model training steps 21 to 27 in the figure
  • model deployment steps 3 to 6 in the figure
  • model transfer step 5 in the figure
  • model activation steps 7 and 8 in the figure
  • model reasoning steps 9 and 10 in the figure
  • model monitoring steps 11 to 16 in the figure
  • model deactivation steps 17 to 18 in the figure
  • model rollback step 20 in the figure
  • model switching step 19 in the figure
  • the first device receives the algorithm configuration information and/or the adjustable parameter configuration information sent by the second device, the first device generates an adjustable parameter set, and/or the first device executes the automation algorithm corresponding to the algorithm configuration information.
  • the algorithm configuration information includes:
  • Hyperparameter optimization parameters the first device uses algorithms such as grid search, random search, continuous halving, Hyperband, etc.
  • the adjustable parameter configuration information includes:
  • adjustable parameter items for generating policy selection and at least one value for each adjustable parameter item, such as base station configuration information, terminal configuration information, scenario configuration information, candidate AI model information, etc.;
  • adjustable parameter items for deactivation and at least one value of each adjustable parameter item such as monitoring indicators, indicators Timers and counters related to target calculation, indicator thresholds, etc.
  • the wireless network can automatically perform operations such as AI model strategy selection, AI model activation condition selection, AI model deactivation condition selection, AI model monitoring configuration, and AI model training control. This avoids manual intervention and improves the system operation efficiency and the online efficiency of the AI model.
  • the AI model strategy determination method provided in the embodiment of the present application can be executed by an AI model strategy determination device.
  • the AI model strategy determination device executing the AI model strategy determination method is taken as an example to illustrate the AI model strategy determination device provided in the embodiment of the present application.
  • FIG. 13 is a schematic diagram of a structure of an AI model strategy determination device provided in an embodiment of the present application. As shown in FIG. 13 , the AI model strategy determination device 1300 includes:
  • the acquisition module 1301 is used to acquire an adjustable parameter set and/or algorithm configuration information;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value of each adjustable parameter item;
  • the algorithm configuration information is used to indicate the configuration parameters of the target algorithm; N is a positive integer;
  • a determination module 1302 is used to determine a target strategy for the AI function based on the adjustable parameter set and/or the algorithm configuration information; the target strategy includes at least one of the following: an AI model deployment strategy; an AI model deactivation strategy; an AI model activation strategy; an AI model training strategy;
  • the processing module 1303 is used to process the corresponding AI model according to the target strategy to provide AI services for the terminal.
  • the determination module can determine the target strategy for the AI function based on the adjustable parameter set and/or algorithm configuration information, thereby improving the efficiency of determining the target strategy.
  • the processing module then processes the corresponding AI model according to the target strategy, provides AI services for the terminal, and can improve the operating efficiency of the communication system.
  • the acquisition module 1301 is specifically used for:
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items.
  • the acquisition module 1301 is further specifically configured to:
  • the adjustable parameter configuration information includes all adjustable parameter items of the N adjustable parameter items and at least one value of each adjustable parameter item, determining the adjustable parameter configuration information as the adjustable parameter set;
  • the adjustable parameter configuration information includes all adjustable parameter items in the N adjustable parameter items, determining at least one value of each adjustable parameter item based on the parameter item configuration information; determining the adjustable parameter set based on the adjustable parameter configuration information and at least one value of each adjustable parameter item;
  • the adjustable parameter configuration information includes all adjustable parameter items of the N adjustable parameter items and at least one value of some adjustable parameter items of the N adjustable parameter items, determining at least one value of the remaining adjustable parameter items of the N adjustable parameter items based on the parameter item configuration information; determining the adjustable parameter set based on the adjustable parameter configuration information and at least one value of the remaining adjustable parameter items;
  • the adjustable parameter configuration information includes some of the N adjustable parameter items and at least one value of at least one of the some of the adjustable parameter items, determine at least one value of the first adjustable parameter item and the adjustable parameter items without values among the N adjustable parameter items based on the parameter item configuration information; determine the adjustable parameter set based on the adjustable parameter configuration information, the first adjustable parameter item and at least one value of the adjustable parameter items without values; the first adjustable parameter item is the adjustable parameter item among the N adjustable parameter items excluding the some of the adjustable parameter items.
  • the adjustable parameter configuration information may include at least one of the following:
  • First configuration information used to assist in determining the AI model deployment strategy
  • Second configuration information used to assist in determining the AI model activation strategy
  • the fourth configuration information is used to assist in determining the AI model training strategy.
  • the first configuration information may include at least one of the following:
  • the network configuration information may include at least one of the following:
  • the terminal configuration information may include at least one of the following:
  • the network scenario information may include at least one of the following:
  • the terminal scenario information includes at least one of the following:
  • the second configuration information includes at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is activated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to activate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that when the count value of the second counter is greater than the second quantity threshold, the first counter is reset or the first timer is paused; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the third configuration information may include at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is deactivated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to deactivate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that when the count value of the second counter is greater than the second quantity threshold, the first counter is reset or the first timer is paused; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the first parameter may include at least one of the following:
  • the fourth configuration information may include at least one of the following:
  • the algorithm configuration information may include at least one of the following:
  • the reinforcement learning algorithm configuration information may include at least one of the following:
  • the grid search algorithm configuration information may include at least one of the following:
  • the random search algorithm configuration information may include at least one of the following:
  • the continuous halving algorithm configuration information may include at least one of the following:
  • the Hyperband algorithm configuration information may include at least one of the following:
  • the first device may include at least one of the following:
  • Base station DU Base station DU.
  • the second device may include at least one of the following:
  • FIG. 14 is a second structural diagram of an AI model strategy determination device provided in an embodiment of the present application. As shown in FIG. 14 , the AI model strategy determination device 1400 includes:
  • the sending module 1401 is configured to send any of the following items to the first device:
  • the adjustable parameter set is used to assist in determining a target strategy for an AI function;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value for each adjustable parameter item, where N is a positive integer;
  • the algorithm configuration information is used to indicate the configuration parameters of the target algorithm, and the algorithm configuration information is used to assist in determining the target algorithm.
  • Target strategy for AI capabilities Target strategy for AI capabilities
  • the adjustable parameter configuration information is used to assist in determining the adjustable parameter set;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items;
  • the target strategy includes at least one of the following:
  • a sending module sends any of the following items to the first device: an adjustable parameter set and algorithm configuration information; adjustable parameter configuration information and algorithm configuration information; adjustable parameter set; adjustable parameter configuration information; algorithm configuration information; after the first device obtains the adjustable parameter set and/or algorithm configuration information, a target strategy for the AI function is determined based on the adjustable parameter set and/or algorithm configuration information to improve the efficiency of determining the target strategy, and then the first device processes the corresponding AI model according to the target strategy, provides AI services to the terminal, and can improve the operating efficiency of the communication system.
  • the adjustable parameter configuration information may include at least one of the following:
  • First configuration information used to assist in determining the AI model deployment strategy
  • Second configuration information used to assist in determining the AI model activation strategy
  • the fourth configuration information is used to assist in determining the AI model training strategy.
  • the first configuration information may include at least one of the following:
  • the network configuration information may include at least one of the following:
  • the terminal configuration information may include at least one of the following:
  • the network scenario information may include at least one of the following:
  • the terminal scenario information may include at least one of the following:
  • the second configuration information includes at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is activated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to activate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that the first counter is reset or the first timer is paused when the count value of the second counter is greater than the second quantity threshold; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the third configuration information may include at least one of the following:
  • a first quantity threshold used to indicate that when the count value of the first counter is greater than the first quantity threshold, the first timer is started, or the first timer is paused and the AI model is deactivated; the first counter is used to record the cumulative number of times the monitoring indicator meets the indicator threshold;
  • a first time threshold used to deactivate the AI model when the first timer exceeds the first time threshold
  • a reset method of the first counter including at least one of periodic reset, event reset and timer reset;
  • a second quantity threshold used to indicate that when the count value of the second counter is greater than the second quantity threshold, the first counter is reset or the first timer is paused; the second counter is used to record the cumulative number of times the monitoring indicator meets the reset threshold;
  • a reset period of periodic reset including a sample period and/or a time period; the sample period is used to reset the first counter when the number of consecutive monitoring samples is greater than or equal to the sample period and the first counter does not reach a first quantity threshold; the time period is used to reset the first counter when the continuous timing of the second timer exceeds the time period and the first counter does not reach the first quantity threshold;
  • a reset threshold for resetting the timer used to indicate that the first counter is reset when the accumulated timing of the second timer exceeds the second time threshold and the first counter has not reached the first quantity threshold; the second timer is used to record the counting time of the first counter.
  • the first parameter may include at least one of the following:
  • the fourth configuration information may include at least one of the following:
  • the algorithm configuration information may include at least one of the following:
  • the reinforcement learning algorithm configuration information may include at least one of the following:
  • the grid search algorithm configuration information may include at least one of the following:
  • the random search algorithm configuration information may include at least one of the following:
  • the continuous halving algorithm configuration information may include at least one of the following:
  • the Hyperband algorithm configuration information may include at least one of the following:
  • the first device may include at least one of the following:
  • Base station DU Base station DU.
  • the second device may include at least one of the following:
  • the AI model strategy determination device in the embodiment of the present application can be an electronic device, such as an electronic device with an operating system, or a component in an electronic device, such as an integrated circuit or a chip.
  • the electronic device can be a network-side device, or it can be a device other than a network-side device.
  • the network-side device can include but is not limited to the types of network-side devices 12 listed above, and other devices can be servers, network attached storage (NAS), etc., which are not specifically limited in the embodiment of the present application.
  • the AI model strategy determination device provided in the embodiment of the present application can implement the various processes implemented by the method embodiments of Figures 2 to 12 and achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • FIG15 is a schematic diagram of the structure of a communication device provided in an embodiment of the present application.
  • an embodiment of the present application further provides a communication device 1500, including a processor 1501 and a memory 1502.
  • the memory 1502 stores a program or instruction that can be run on the processor 1501.
  • the communication device 1500 is a first device
  • the program or instruction is executed by the processor 1501 to implement the various steps of the embodiment of the AI model strategy determination method on the first device side, and can achieve the same technical effect.
  • the communication device 1500 is a second device
  • the program or instruction is executed by the processor 1501 to implement the various steps of the embodiment of the AI model strategy determination method on the second device side, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • An embodiment of the present application also provides a first device, including a processor and a communication interface, the processor being used to: obtain an adjustable parameter set and/or algorithm configuration information; the adjustable parameter set including: N adjustable parameter items and at least one value for each adjustable parameter item; the algorithm configuration information is used to indicate the configuration parameters of the target algorithm; N is a positive integer; the processor is also used to: determine a target strategy for the AI function based on the adjustable parameter set and/or the algorithm configuration information; the target strategy includes at least one of the following: AI model deployment strategy; AI model deactivation strategy; AI model activation strategy; AI model training strategy; the processor is also used to: process the corresponding AI model according to the target strategy to provide AI services for the terminal.
  • the first device embodiment corresponds to the above-mentioned first device side method embodiment.
  • Each implementation process and implementation method of the above-mentioned method embodiment can be applied to the first device embodiment and can achieve the same technical effect.
  • the embodiment of the present application also provides a first device.
  • Figure 16 is one of the structural schematic diagrams of the first device provided in the embodiment of the present application.
  • the first device 1600 includes: an antenna 1601, a radio frequency device 1602, a baseband device 1603, a processor 1604 and a memory 1605.
  • the antenna 1601 is connected to the radio frequency device 1602.
  • the radio frequency device 1602 receives information through the antenna 1601 and sends the received information to the baseband device 1603 for processing.
  • the baseband device 1603 processes the information to be sent and sends it to the radio frequency device 1602.
  • the radio frequency device 1602 processes the received information and sends it out through the antenna 1601.
  • the method executed by the first device in the above embodiment may be implemented in the baseband device 1603, which includes a baseband processor.
  • the baseband device 1603 may include, for example, at least one baseband board, on which multiple chips are arranged, as shown in Figure 16, one of which is, for example, a baseband processor, which is connected to the memory 1605 through a bus interface to call the program in the memory 1605 and execute the network device operations shown in the above method embodiment.
  • the first device may also include a network interface 1606, which is, for example, a common public radio interface (CPRI).
  • a network interface 1606 which is, for example, a common public radio interface (CPRI).
  • CPRI common public radio interface
  • the first device 1600 of the embodiment of the present application also includes: instructions or programs stored in the memory 1605 and executable on the processor 1604.
  • the processor 1604 calls the instructions or programs in the memory 1605 to execute the methods executed by the modules shown in Figure 13 and achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • FIG17 is a second structural diagram of the first device provided in the embodiment of the present application.
  • the first device 1700 includes: a processor 1701, a network interface 1702, and a memory 1703.
  • the network interface 1702 is, for example, a common public radio interface (CPRI).
  • CPRI common public radio interface
  • the first device 1700 of the embodiment of the present application also includes: instructions or programs stored in the memory 1703 and executable on the processor 1701.
  • the processor 1701 calls the instructions or programs in the memory 1703 to execute the methods executed by the modules shown in Figure 13 and achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • the embodiment of the present application further provides a second device, including a processor and a communication interface, where the communication interface is used to send any of the following items to the first device:
  • a set of adjustable parameters and algorithm configuration information A set of adjustable parameters and algorithm configuration information
  • the adjustable parameter set is used to assist in determining a target strategy for an AI function;
  • the adjustable parameter set includes: N adjustable parameter items and at least one value for each adjustable parameter item, where N is a positive integer;
  • the algorithm configuration information is used to indicate configuration parameters of the target algorithm, and the algorithm configuration information is used to assist in determining a target strategy for the AI function;
  • the adjustable parameter configuration information is used to assist in determining the adjustable parameter set;
  • the adjustable parameter configuration information includes: some or all of the N adjustable parameter items, and at least one value of 0 or at least one of the N adjustable parameter items;
  • the target strategy includes at least one of the following: AI model deployment strategy; AI model deactivation strategy; AI model activation strategy; AI model training strategy.
  • the second device embodiment corresponds to the above-mentioned second device side method embodiment.
  • Each implementation process and implementation method of the above-mentioned method embodiment can be applied to the second device embodiment and can achieve the same technical effect.
  • the embodiment of the present application also provides a second device.
  • Figure 18 is one of the structural schematic diagrams of the second device provided in the embodiment of the present application.
  • the second device 1800 includes: an antenna 1801, a radio frequency device 1802, a baseband device 1803, a processor 1804 and a memory 1805.
  • the antenna 1801 is connected to the radio frequency device 1802.
  • the radio frequency device 1802 receives information through the antenna 1801 and sends the received information to the baseband device 1803 for processing.
  • the baseband device 1803 processes the information to be sent and sends it to the radio frequency device 1802.
  • the radio frequency device 1802 processes the received information and sends it out through the antenna 1801.
  • the method executed by the second device in the above embodiment may be implemented in the baseband device 1803, which includes a baseband processor.
  • the baseband device 1803 may include, for example, at least one baseband board, on which multiple chips are arranged, as shown in Figure 18, one of which is, for example, a baseband processor, which is connected to the memory 1805 through a bus interface to call the program in the memory 1805 and execute the network device operations shown in the above method embodiment.
  • the second device may also include a network interface 1806, which is, for example, a common public radio interface (CPRI).
  • a network interface 1806, which is, for example, a common public radio interface (CPRI).
  • CPRI common public radio interface
  • the second device 1800 of the embodiment of the present application also includes: instructions or programs stored in the memory 1805 and executable on the processor 1804.
  • the processor 1804 calls the instructions or programs in the memory 1805 to execute the methods executed by the modules shown in Figure 14 and achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • FIG. 19 is a second structural diagram of the second device provided in the embodiment of the present application.
  • the second device 1900 includes: a processor 1901, a network interface 1902, and a memory 1903.
  • the network interface 1902 is, for example, a common public radio interface (CPRI).
  • CPRI common public radio interface
  • the second device 1900 of the embodiment of the present application also includes: instructions or programs stored in the memory 1903 and executable on the processor 1901.
  • the processor 1901 calls the instructions or programs in the memory 1903 to execute the methods executed by the modules shown in Figure 14 and achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored.
  • a program or instruction is stored.
  • the program or instruction is executed by a processor, each process of the above-mentioned AI model strategy determination method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.
  • the processor is the processor in the network side device described in the above embodiment.
  • the readable storage medium includes a computer readable storage medium, such as a computer read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk.
  • the present application embodiment further provides a chip, the chip comprising a processor and a communication interface, the communication interface and The processor is coupled, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned AI model strategy determination method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.
  • the embodiments of the present application further provide a computer program/program product, which is stored in a storage medium, and is executed by at least one processor to implement the various processes of the above-mentioned AI model strategy determination method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.
  • An embodiment of the present application also provides an AI model strategy determination system, including: a first device and a second device, wherein the first device can be used to execute the steps of the AI model strategy determination method as described above, and the second device can be used to execute the steps of the AI model strategy determination method as described above.
  • the technical solution of the present application can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM/RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.
  • a storage medium such as ROM/RAM, a magnetic disk, or an optical disk
  • a terminal which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Mobile Radio Communication Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Medical Informatics (AREA)

Abstract

本申请公开了一种AI模型策略确定方法、装置、第一设备及第二设备,属于通信技术领域,本申请实施例的AI模型策略确定方法包括:第一设备获取可调参数集合和/或算法配置信息;该可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;该算法配置信息用于指示目标算法的配置参数;N为正整数(201);该第一设备基于该可调参数集合和/或该算法配置信息,确定AI功能的目标策略;该目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略(202);AI模型训练策略;该第一设备根据该目标策略对相应的AI模型进行处理,为终端提供AI服务(203)。

Description

AI模型策略确定方法、装置、第一设备及第二设备
本申请要求于2022年11月10日提交国家知识产权局、申请号为202211406405.6、申请名称为“AI模型策略确定方法、装置、第一设备及第二设备”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请属于通信技术领域,具体涉及一种AI模型策略确定方法、装置、第一设备及第二设备。
背景技术
在移动通信系统中,开始有越来越多的用例结合人工智能(Artificial Intelligence,AI)。例如在物理层有基于AI的信道状态信息(channel state information,CSI)反馈压缩、基于AI的波束管理及基于AI的定位,基于AI的节能、基于AI的负载均衡等。在未来会有更多的结合AI的用例在移动通信系统中出现。
AI模型通常是通过离线训练或者在线训练的方式产生的,产生的AI模型往往只适用于特定场景,针对某个AI功能,例如基于AI的波束管理,进一步可以分为多种实现方案,例如基于AI的收发波束对预测、基于AI的发送波束预测和基于AI的接收波束预测等。针对某种方案,又可以训练出不同的模型,例如,复杂网络模型推理精度高,但尺寸较大;而简单网络模型尺寸小,但推理精度不高。
目前,AI模型在网络中使用之前,往往需要经过大量的实验,获得足够的验证数据才会在现网中激活再使用,导致通信系统运行效率低。
发明内容
本申请实施例提供一种AI模型策略确定方法、装置、第一设备及第二设备,能够解决通信系统运行效率低的问题。
第一方面,提供了一种AI模型策略确定方法,应用于第一设备,该方法包括:
第一设备获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;
所述第一设备基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;
所述第一设备根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
第二方面,提供了一种AI模型策略确定装置,包括:
获取模块,用于获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;
确定模块,用于基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;
处理模块,用于根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
第三方面,提供了一种AI模型策略确定方法,应用于第二设备,该方法包括:
第二设备向第一设备发送以下任一项:
可调参数集合和算法配置信息;
可调参数配置信息和算法配置信息;
可调参数集合;
可调参数配置信息;
算法配置信息;
其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包 括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针对AI功能的目标策略;
所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略。
第四方面,提供了一种AI模型策略确定装置,包括:
发送模块,用于向第一设备发送以下任一项:
可调参数集合和算法配置信息;
可调参数配置信息和算法配置信息;
可调参数集合;
可调参数配置信息;
算法配置信息;
其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针对AI功能的目标策略;
所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略。
第五方面,提供了一种第一设备,该第一设备包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如第一方面所述的方法的步骤。
第六方面,提供了一种第一设备,包括处理器及通信接口,其中,所述处理器用于获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;所述处理器还用于基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;所述处理器还用于:根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
第七方面,提供了一种第二设备,该第二设备包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如第三方面所述的方法的步骤。
第八方面,提供了一种第二设备,包括处理器及通信接口,其中,所述通信接口用于向第一设备发送以下任一项:
可调参数集合和算法配置信息;
可调参数配置信息和算法配置信息;
可调参数集合;
可调参数配置信息;
算法配置信息;
其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针对AI功能的目标策略;
所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略。
第九方面,提供了一种AI模型策略确定系统,包括:第一设备及第二设备,所述第一设备可用于执行如第一方面所述的AI模型策略确定方法的步骤,所述第二设备可用于执行如第三方面所述的AI模型策略确定方法的步骤。
第十方面,提供了一种可读存储介质,所述可读存储介质上存储程序或指令,所述程序或指令被处理器执行时实现如第一方面所述的方法的步骤,或者实现如第三方面所述的方法的步骤。
第十一方面,提供了一种芯片,所述芯片包括处理器和通信接口,所述通信接口和所述处理器耦合,所述处理器用于运行程序或指令,实现如第一方面所述的方法,或实现如第三方面所述的方法。
第十二方面,提供了一种计算机程序/程序产品,所述计算机程序/程序产品被存储在存储介质中,所述计算机程序/程序产品被至少一个处理器执行以实现如第一方面所述的AI模型策略确定方法的步骤,或实现如第三方面所述的AI模型策略确定方法的步骤。
在本申请实施例中,第一设备获取到可调参数集合和/或算法配置信息之后,可以基于可调参数集合和/或算法配置信息确定针对AI功能的目标策略,提高目标策略的确定效率,进而第一设备根据目标策略对相应的AI模型进行处理,为终端提供AI服务,能够提高通信系统运行效率。
附图说明
图1是本申请实施例可应用的一种无线通信系统的框图;
图2是本申请实施例提供的AI模型策略确定方法的流程示意图之一;
图3是本申请实施例提供的AI模型策略确定方法的流程示意图之二;
图4是本申请实施例提供的AI模型策略确定方法的流程示意图之三;
图5是本申请实施例提供的AI模型策略确定方法的流程示意图之四;
图6是本申请实施例提供的AI模型策略确定方法的流程示意图之五;
图7是本申请实施例提供的AI模型策略确定方法的流程示意图之六;
图8是本申请实施例提供的AI模型策略确定方法的流程示意图之七;
图9是本申请实施例提供的AI模型策略确定方法的流程示意图之八;
图10是本申请实施例提供的AI模型策略确定方法的流程示意图之九;
图11是本申请实施例提供的AI模型策略确定方法的信令交互图;
图12是本申请实施例提供的AI模型的生命周期管理的流程示意图;
图13是本申请实施例提供的AI模型策略确定装置的结构示意图之一;
图14是本申请实施例提供的AI模型策略确定装置的结构示意图之二;
图15是本申请实施例提供的通信设备的结构示意图;
图16是本申请实施例提供的第一设备的结构示意图之一;
图17是本申请实施例提供的第一设备的结构示意图之二;
图18是本申请实施例提供的第二设备的结构示意图之一;
图19是本申请实施例提供的第二设备的结构示意图之二。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员所获得的所有其他实施例,都属于本申请保护的范围。
本申请的说明书和权利要求书中的术语“第一”、“第二”等是用于区别类似的对象,而不用于描述特定的顺序或先后次序。应该理解这样使用的术语在适当情况下可以互换,以便本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施,且“第一”、“第二”所区别的对象通常为一类,并不限定对象的个数,例如第一对象可以是一个,也可以是多个。此外,说明书以及权利要求中“和/或”表示所连接对象的至少其中之一,字符“/”一般表示前后关联对象是一种“或”的关系。
值得指出的是,本申请实施例所描述的技术不限于长期演进型(Long Term Evolution,LTE)/LTE的演进(LTE-Advanced,LTE-A)系统,还可用于其他无线通信系统,诸如码分 多址(Code Division Multiple Access,CDMA)、时分多址(Time Division Multiple Access,TDMA)、频分多址(Frequency Division Multiple Access,FDMA)、正交频分多址(Orthogonal Frequency Division Multiple Access,OFDMA)、单载波频分多址(Single-carrier Frequency Division Multiple Access,SC-FDMA)和其他系统。本申请实施例中的术语“系统”和“网络”常被可互换地使用,所描述的技术既可用于以上提及的系统和无线电技术,也可用于其他系统和无线电技术。以下描述出于示例目的描述了新空口(New Radio,NR)系统,并且在以下大部分描述中使用NR术语,但是这些技术也可应用于NR系统应用以外的应用,如第6代(6th Generation,6G)通信系统。
图1是本申请实施例可应用的一种无线通信系统的框图。无线通信系统包括终端11和网络侧设备12。其中,终端11可以是手机、平板电脑(Tablet Personal Computer)、膝上型电脑(Laptop Computer)或称为笔记本电脑、个人数字助理(Personal Digital Assistant,PDA)、掌上电脑、上网本、超级移动个人计算机(ultra-mobile personal computer,UMPC)、移动上网装置(Mobile Internet Device,MID)、增强现实(augmented reality,AR)/虚拟现实(virtual reality,VR)设备、机器人、可穿戴式设备(Wearable Device)、车载设备(VUE)、行人终端(PUE)、智能家居(具有无线通信功能的家居设备,如冰箱、电视、洗衣机或者家具等)、游戏机、个人计算机(personal computer,PC)、柜员机或者自助机等终端侧设备,可穿戴式设备包括:智能手表、智能手环、智能耳机、智能眼镜、智能首饰(智能手镯、智能手链、智能戒指、智能项链、智能脚镯、智能脚链等)、智能腕带、智能服装等。需要说明的是,在本申请实施例并不限定终端11的具体类型。网络侧设备12可以包括接入网设备或核心网设备,其中,接入网设备12也可以称为无线接入网设备、无线接入网(Radio Access Network,RAN)、无线接入网功能或无线接入网单元。接入网设备12可以包括基站、WLAN接入点或WiFi节点等,基站可被称为节点B、演进节点B(eNB)、接入点、基收发机站(Base Transceiver Station,BTS)、无线电基站、无线电收发机、基本服务集(Basic Service Set,BSS)、扩展服务集(Extended Service Set,ESS)、家用B节点、家用演进型B节点、发送接收点(Transmitting Receiving Point,TRP)或所述领域中其他某个合适的术语,只要达到相同的技术效果,所述基站不限于特定技术词汇,需要说明的是,在本申请实施例中仅以NR系统中的基站为例进行介绍,并不限定基站的具体类型。核心网设备可以包含但不限于如下至少一项:核心网节点、核心网功能、移动管理实体(Mobility Management Entity,MME)、接入移动管理功能(Access and Mobility Management Function,AMF)、会话管理功能(Session Management Function,SMF)、用户平面功能(User Plane Function,UPF)、策略控制功能(Policy Control Function,PCF)、策略与计费规则功能单元(Policy and Charging Rules Function,PCRF)、边缘应用服务发现功能(Edge Application Server Discovery Function,EASDF)、统一数据管理(Unified Data Management,UDM),统一数据仓储(Unified Data Repository,UDR)、归属用户服务器(Home Subscriber Server,HSS)、集中式网络配置(Centralized network configuration,CNC)、网络存储功能(Network Repository Function,NRF),网络开放功能(Network Exposure Function,NEF)、本地NEF(Local NEF,或L-NEF)、绑定支持功能(Binding Support Function,BSF)、应用功能(Application Function,AF)等。需要说明的是,在本申请实施例中仅以NR系统中的核心网设备为例进行介绍,并不限定核心网设备的具体类型。
下面结合附图,通过一些实施例及其应用场景对本申请实施例提供的AI模型策略确定方法、装置、第一设备及第二设备进行详细地说明。
为了便于更加清晰地理解本申请各实施例,首先对一些相关的技术知识进行如下介绍。
一、超参数优化算法
针对AI模型训练的超参数优化方法,目前比较主流的方法包括黑盒优化算法和多保真度优化(Multi-fidelity optimization)。其中,黑盒优化算法的经典方法又包括了网格搜索,随机搜索,贝叶斯优化等。多保真度优化算法的经典方法包括连续减半(Successive Halving)和Hyperband算法。
二、网络模型搜索算法
在AI模型训练优化中,可以通过搜索不同的网络结构以提升AI模型的预测精度。目前较为主流的模型网络结构搜索(network architecture search,NAS)算法包括随机搜索、强化学习搜索和训练超级网络等方法。
目前的超参数优化问题,主要解决的是模型训练时如何选择基础模型、学习率、批大小、优化函数等参数。而网络模型搜索算法解决的是AI模型网络结构如何选择的问题。上述两种方法并不能直接用于无线网络中AI用例实验阶段的部署策略选择、模型激活条件选择、模型去激活条件选择等问题。
目前针对无线网络中AI用例实验阶段的部署策略选择、模型激活条件选择、模型去激活条件选择等自动化处理并没有成熟的方法。
本申请实施例提供的AI模型策略确定方法,可应用于需要确定针对AI功能的目标策略的第一设备,目标策略可以包括以下至少一项:AI模型部署策略、AI模型去激活策略、AI模型激活策略及AI模型训练策略。
图2是本申请实施例提供的AI模型策略确定方法的流程示意图之一,如图2所示,该方法包括步骤201和步骤202;其中:
步骤201、第一设备获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;
步骤202、所述第一设备基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;
步骤203、所述第一设备根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
具体地,相关技术中,AI模型在网络中使用之前,往往需要经过大量的实验,获得足够的验证数据才会在现网中激活再使用,但是,现有的超参数优化、网络模型搜索等算法,并不能直接用于无线网络中AI用例实验阶段的部署策略选择、模型激活条件选择、模型去激活条件选择等问题,目前AI模型的实验阶段往往需要大量的人工干预配置,导致AI模型的上线效率低。
本申请实施例中,第一设备可以获取可调参数集合和/或算法配置信息,其中,可调参数集合用于辅助确定针对AI功能的目标策略,可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;算法配置信息用于指示目标算法的配置参数;
第一设备在获取可调参数集合和/或算法配置信息后,可以基于可调参数集合和/或算法配置信息,确定包括AI模型部署策略、AI模型去激活策略、AI模型激活策略及AI模型训练策略中至少一项的目标策略,以根据目标策略对相应的AI模型进行处理,为终端提供AI服务,例如可以根据目标策略使得AI模型的实验阶段自动化进行,相较于在实验阶段对AI模型进行大量人工干预配置,本申请实施例可以有效提高AI模型的上线效率,进而提高了对终端提供AI服务的性能。
需要说明的是,AI模型也是可选的,目标策略例如可以在设置的10个AI模型中,选择一个或多个AI模型执行算法配置信息对应的自动化算法。
可选地,AI功能,例如包括基于AI的波束管理、基于AI的CSI信道压缩反馈、基于AI的定位、基于AI的基站节能、基于AI的负载均衡。
可选地,所述第一设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站中心单元(Centralized Unit,CU);
(5)与基站CU平级的实体;
具体地,与基站CU平级的实体可以是与基站CU平级的新增实体。
(6)基站分布单元(Distributed Unit,DU)。
在本申请实施例提供的AI模型策略确定方法中,第一设备获取到可调参数集合和/或算法配置信息之后,可以基于可调参数集合和/或算法配置信息确定针对AI功能的目标策略,提高目标策略的确定效率,进而第一设备根据目标策略对相应的AI模型进行处理,为终端提供AI服务,能够提高通信系统运行效率。
可选地,所述第一设备获取可调参数集合和/或算法配置信息的实现方式可以包括以下至少一项:
1、所述第一设备接收来自第二设备的所述可调参数集合和/或所述算法配置信息;
具体地,可以分为以下3种情况:
情况1、第一设备接收来自第二设备的可调参数集合,算法配置信息是预定义的;
情况2、第一设备接收来自第二设备的算法配置信息,可调参数集合是预定义的;
情况3、第一设备接收来自第二设备的可调参数集合和算法配置信息。
可选地,可调参数集合和算法配置信息也可以都是预定义的。
2、所述第一设备接收来自第二设备的可调参数配置信息和/或所述算法配置信息,所述第一设备基于所述可调参数配置信息,确定所述可调参数集合;
具体地,可以分为以下3种情况:
情况1、第一设备接收来自第二设备的可调参数配置信息,算法配置信息是预定义的,第一设备基于可调参数配置信息,确定可调参数集合;
情况2、第一设备接收来自第二设备的算法配置信息,可调参数配置信息是预定义的,第一设备基于可调参数配置信息,确定可调参数集合;
情况3、第一设备接收来自第二设备的可调参数配置信息和算法配置信息,第一设备基于可调参数配置信息,确定可调参数集合。
可选地,可调参数配置信息和算法配置信息也可以都是预定义的。
3、所述第一设备接收来自第二设备的所述可调参数集合;所述第一设备接收来自第三设备的所述算法配置信息;
4、所述第一设备接收来自第二设备的可调参数配置信息,所述第一设备基于所述可调参数配置信息,确定所述可调参数集合;所述第一设备接收来自第三设备的所述算法配置信息;
其中,所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值。
可选地,可调参数配置信息可以包括:
a.用于AI模型部署策略选择的全部(N个可调参数项)或部分可调参数项;可选地,还可以包括可调参数项取值;
b.用于AI模型去激活策略选择的全部(N个可调参数项)或部分可调参数项;可选地,还可以包括可调参数项取值;
c.用于AI模型激活策略选择的全部(N个可调参数项)或部分可调参数项;可选地,还可以包括可调参数项取值;
d.用于AI模型训练的全部(N个可调参数项)或部分可调参数项;可选地,还可以包括可调参数项取值。
可选地,所述第二设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体。
具体地,与基站CU平级的实体可以是与基站CU平级的新增实体。
可选地,所述第一设备基于所述可调参数配置信息,确定所述可调参数集合的实现方式可以包括以下至少一项:
1)在所述可调参数配置信息包括N个可调参数项中的全部可调参数项,及各可调参数项的至少一个取值的情况下,所述第一设备将所述可调参数配置信息确定为所述可调参数集合;
具体地,第一设备直接从第二设备获取可调参数集合;也即,由第二设备提供可调参数集合,其中,可调参数集合包括了所有可调参数项,以及每个可调参数项的离散取值集合。
2)在所述可调参数配置信息包括N个可调参数项中的全部可调参数项的情况下,所述第一设备基于参数项配置信息,确定各可调参数项的至少一个取值;所述第一设备基于所述可调参数配置信息及各可调参数项的至少一个取值,确定所述可调参数集合;
具体地,第一设备从第二设备获取可调参数配置信息,在可调参数配置信息包括N个 可调参数项的情况下,第一设备可以基于参数项配置信息,先确定N个可调参数项各自的至少一个取值,再基于可调参数配置信息及N个可调参数项各自的至少一个取值,确定可调参数集合;也即,第二设备提供的可调参数配置信息包括了所有可调参数项,由第一设备确定每个可调参数项的离散取值集合。
3)在所述可调参数配置信息包括N个可调参数项中的全部可调参数项,及所述N个可调参数项中的部分可调参数项的至少一个取值的情况下,所述第一设备基于参数项配置信息,确定所述N个可调参数项中的剩余可调参数项的至少一个取值;所述第一设备基于所述可调参数配置信息及所述剩余可调参数项的至少一个取值,确定所述可调参数集合;
具体地,第一设备从第二设备获取可调参数配置信息,在可调参数配置信息包括N个可调参数项,及N个可调参数项中的部分可调参数项的至少一个取值的情况下,第一设备可以基于参数项配置信息,先确定N个可调参数项中的剩余可调参数项的至少一个取值,再基于可调参数配置信息及剩余可调参数项的至少一个取值,确定可调参数集合;也即,第二设备提供的可调参数配置信息包括了所有参数项和部分参数项的取值集合,由第一设备确定剩余参数项的离散取值集合。
可选地,参数项配置信息可以包括经验值或默认值。
4)在所述可调参数配置信息包括所述N个可调参数项中的部分可调参数项,及所述部分可调参数项中的至少一个可调参数项的至少一个取值的情况下,所述第一设备基于参数项配置信息,确定第一可调参数项及所述N个可调参数项中无取值的可调参数项的至少一个取值;所述第一设备基于所述可调参数配置信息、所述第一可调参数项及所述无取值的可调参数项的至少一个取值,确定所述可调参数集合;所述第一可调参数项为所述N个可调参数项中除所述部分可调参数项之外的可调参数项。
具体地,第一设备从第二设备获取可调参数配置信息,在可调参数配置信息包括部分可调参数项,及部分可调参数项中的至少一个可调参数项的至少一个取值的情况下,第一设备可以基于参数项配置信息,先确定第一可调参数项及N个可调参数项中无取值的可调参数项的至少一个取值,第一可调参数项为N个可调参数项中除部分可调参数项之外的可调参数项,第一设备再基于可调参数配置信息、第一可调参数项及N个可调参数项中无取值的可调参数项的至少一个取值,确定可调参数集合;也即,第二设备提供的可调参数配置信息包括了部分可调参数项和至少一个对应的离散取值集合,由第一设备确定N个可调参数项中的剩余可调参数项,和无取值的可调参数项对应的离散取值集合。
在一个实施例中,第二设备提供的可调参数配置信息包括了部分可调参数项和对应的离散取值集合,由第一设备确定N个可调参数项中的剩余可调参数项和对应的离散取值集合。
可选地,所述可调参数配置信息可以包括以下至少一项:
1、第一配置信息,用于辅助确定所述AI模型部署策略;
可选地,所述第一配置信息可以包括以下至少一项:
1)网络配置信息;
可选地,所述网络配置信息可以包括以下至少一项:
a.基站天线配置信息;
b.基站波束配置信息;
c.基站波束个数信息;
d.基站高度信息;
e.基站站间距信息。
2)终端配置信息;
可选地,所述终端配置信息可以包括以下至少一项:
a.终端天线配置信息;
b.终端波束配置信息;
c.终端波束个数信息;
d.终端天线面板个数信息。
3)网络场景信息;
可选地,所述网络场景信息可以包括以下至少一项:
a.视距场景信息;
b.非视距场景信息;
c.室外场景信息;
d.室内场景信息;
e.用户密集场景信息;
f.用户稀疏场景信息;
g.不同降雨量场景信息。
具体地,不同降雨量场景信息,可以包括用于指示降雨量无、小、中、大的信息。
4)终端场景信息;
可选地,所述终端场景信息可以包括以下至少一项:
a.终端移动速度;
b.终端旋转速度。
5)候选AI模型信息;
6)所述AI功能对应的候选方案;
具体地,AI功能对应的候选方案,例如为基于AI的波束管理方案,具体方案有基站侧的波束预测方案、终端侧接收波束预测方案,还可以是收发波束预测方案;又例如为基于AI的信道压缩反馈对应的候选方案。
7)候选推理侧信息。
2、第二配置信息,用于辅助确定所述AI模型激活策略;
可选地,所述第二配置信息可以包括以下至少一项:
1)至少一个监视指标;
具体地,监视指标例如为最优波束预测准确率,预测误差,余弦相似度,吞吐量等。
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
可选地,所述第一参数可以包括以下至少一项:
a.样本周期;
b.样本占比;
c.样本分布。
4)指标门限;
5)启动条件,用于指示在开启模型激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及激活AI模型;第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
需要说明的是,周期性重置和计时器重置的主要区别在于,周期性重置可以是固定10000个样本(样本周期),或者固定1000毫秒ms(时间周期)对第一计数器进行重置;而计时器重置,计时器是可以中途停止和启动的。
举例来说,假设事件是激活AI模型,当前AI模型还未激活。
假设第一计时器的周期是10000个样本或者1000ms,当满足条件1时,第一计时器启动;当满足条件2时,第一计时器停止;当第一计时器超时,AI模型激活。
条件1:第一计数器到达第一数量门限。其中,第一计数器递增,表征AI模型的指标(例如最优波束的预测准确率)高于门限(例如90%),预测结果好。其中,第一计数器的重置条件为条件2。
条件2:第二计数器到达第二数量门限。其中,第二计数器递增,表征AI模型的指标(例如最优波束的预测准确率)低于门限(例如85%),预测结果不好。其中,第二计数器的重置条件为条件1。
又例如,假设事件是激活AI模型,当前AI模型还未激活。
假设第一计时器的周期是10000个样本或者1000ms,当满足条件1时,第一计时器启动;当满足条件2时,第一计时器停止,AI模型激活;当满足条件3时,第一计数器重置;当第一计时器超时,AI模型维持去激活。
条件1:开启模型激活评估流程。
条件2:第一计数器到达第一数量门限。其中,第一计数器递增,表征AI模型的指标(例如最优波束的预测准确率)高于门限(例如90%),预测结果好。第一计数器的重置为条件3。
条件3:第二计数器到达第二数量门限。其中,第二计数器递增,表征AI模型的指标(例如最优波束的预测准确率)低于门限(例如85%),预测结果不好。第二计数器的重置为条件2。
3、第三配置信息,用于辅助确定所述AI模型去激活策略;
可选地,所述第三配置信息可以包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型去激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及去激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,去激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
4、第四配置信息,用于辅助确定所述AI模型训练策略。
可选地,所述第四配置信息可以包括以下至少一项:
1)模型输入参数;
2)预处理策略;
3)后处理策略;
4)模型输出参数;
5)训练超参数;
6)模型网络结构。
可选地,所述算法配置信息可以包括以下至少一项:
1、强化学习算法配置信息;
可选地,所述强化学习算法配置信息可以包括以下至少一项:
1)收集奖励的实体;
具体地,收集奖励的实体例如为奖励(reward)收集位置。
2)奖励计算公式;
3)奖励对应的参数;
4)折扣因子;
5)贪婪因子最大值;
6)贪婪因子最小值;
7)贪婪因子变化步长。
2、网格搜索算法配置信息;
可选地,所述网格搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的总搜索空间;
具体地,可以为用于验证的AI模型部署策略的总搜索空间;
2)所述AI模型去激活策略的总搜索空间;
具体地,可以为用于验证的AI模型去激活策略的总搜索空间;
3)所述AI模型激活策略的总搜索空间;
具体地,可以为用于验证的AI模型激活策略的总搜索空间;
4)所述AI模型训练策略的总搜索空间。
具体地,可以为用于验证的AI模型训练策略的总搜索空间。
3、随机搜索算法配置信息;
可选地,所述随机搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的最大个数;
具体地,可以为用于验证的AI模型部署策略的最大个数;
2)所述AI模型去激活策略的最大个数;
具体地,可以为用于验证的AI模型去激活策略的最大个数;
3)所述AI模型激活策略的最大个数;
具体地,可以为用于验证的AI模型激活策略的最大个数;
4)所述AI模型训练策略的最大个数。
具体地,可以为用于验证的AI模型训练策略的最大个数。
4、连续减半算法配置信息;
可选地,所述连续减半算法配置信息包括以下至少一项:
1)所述AI模型部署策略的数目和每个策略的验证时长;
具体地,可以为用于验证的AI模型部署策略的数目和每个策略的验证时长;
2)所述AI模型去激活策略的数目和每个策略的验证时长;
具体地,可以为用于验证的AI模型去激活策略的数目和每个策略的验证时长;
3)所述AI模型激活策略的数目和每个策略的验证时长;
具体地,可以为用于验证的AI模型激活策略的数目和每个策略的验证时长;
4)所述AI模型训练策略的数目和每个策略的验证时长。
具体地,可以为用于验证的AI模型训练策略的数目和每个策略的验证时长。
5、超参数优化Hyperband算法配置信息。
可选地,所述Hyperband算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
具体地,可以为用于验证的AI模型部署策略的连续减半算法执行次数,以及初始连续减半算法中策略个数和验证时长;
2)所述AI模型去激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
具体地,可以为用于验证的AI模型去激活策略的连续减半算法执行次数,以及初始连续减半算法中策略个数和验证时长;
3)所述AI模型激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
具体地,可以为用于验证的AI模型激活策略的连续减半算法执行次数,以及初始连续减半算法中策略个数和验证时长;
4)所述AI模型训练策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。
具体地,可以为用于验证的AI模型训练策略的连续减半算法执行次数,以及初始连续减半算法中策略个数和验证时长。
本申请实施例提供的AI模型策略确定方法,可应用于向第一设备发送可调参数集合、可调参数配置信息及算法配置信息中至少一项的第二设备。
图3是本申请实施例提供的AI模型策略确定方法的流程示意图之二,如图3所示,该方法包括步骤301;其中:
步骤301、第二设备向第一设备发送以下任一项:可调参数集合和算法配置信息;可调参数配置信息和算法配置信息;可调参数集合;可调参数配置信息;算法配置信息;
其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针对AI功能的目标策略;
所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
所述目标策略包括以下至少一项:
1)AI模型部署策略;
2)AI模型去激活策略;
3)AI模型激活策略;
4)AI模型训练策略。
具体地,第二设备可以向第一设备发送以下任一项:可调参数集合和算法配置信息;可调参数配置信息和算法配置信息;可调参数集合;可调参数配置信息;算法配置信息;以由第一设备基于上述任一项获取可调参数集合和/或算法配置信息,进而基于可调参数集合和/或算法配置信息,确定针对AI功能的目标策略。
可选地,所述第一设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体;
(6)基站DU。
可选地,所述第二设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体。
在本申请实施例提供的AI模型策略确定方法中,第二设备向第一设备发送以下任一项:可调参数集合和算法配置信息;可调参数配置信息和算法配置信息;可调参数集合;可调参数配置信息;算法配置信息;以由第一设备获取到可调参数集合和/或算法配置信息之后,基于可调参数集合和/或算法配置信息确定针对AI功能的目标策略,提高目标策略的确定效率,进而第一设备根据目标策略对相应的AI模型进行处理,为终端提供AI服务,能够提高通信系统运行效率。
可选地,所述可调参数配置信息可以包括以下至少一项:
1、第一配置信息,用于辅助确定所述AI模型部署策略;
2、第二配置信息,用于辅助确定所述AI模型激活策略;
3、第三配置信息,用于辅助确定所述AI模型去激活策略;
4、第四配置信息,用于辅助确定所述AI模型训练策略。
可选地,所述第一配置信息可以包括以下至少一项:
1)网络配置信息;
2)终端配置信息;
3)网络场景信息;
4)终端场景信息;
5)候选AI模型信息;
6)所述AI功能对应的候选方案;
7)候选推理侧信息。
其中,所述网络配置信息可以包括以下至少一项:
a.基站天线配置信息;
b.基站波束配置信息;
c.基站波束个数信息;
d.基站高度信息;
e.基站站间距信息。
所述终端配置信息可以包括以下至少一项:
a)终端天线配置信息;
b)终端波束配置信息;
c)终端波束个数信息;
d)终端天线面板个数信息。
所述网络场景信息可以包括以下至少一项:
a.视距场景信息;
b.非视距场景信息;
c.室外场景信息;
d.室内场景信息;
e.用户密集场景信息;
f.用户稀疏场景信息;
g.不同降雨量场景信息。
所述终端场景信息可以包括以下至少一项:
a)终端移动速度;
b)终端旋转速度。
可选地,所述第二配置信息包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
可选地,所述第三配置信息可以包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型去激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及去激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,去激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
可选地,所述第一参数可以包括以下至少一项:
a.样本周期;
b.样本占比;
c.样本分布。
可选地,所述第四配置信息可以包括以下至少一项:
1)模型输入参数;
2)预处理策略;
3)后处理策略;
4)模型输出参数;
5)训练超参数;
6)模型网络结构。
可选地,所述算法配置信息可以包括以下至少一项:
1、强化学习算法配置信息;
2、网格搜索算法配置信息;
3、随机搜索算法配置信息;
4、连续减半算法配置信息;
5、Hyperband算法配置信息。
其中,所述强化学习算法配置信息可以包括以下至少一项:
1)收集奖励的实体;
2)奖励计算公式;
3)奖励对应的参数;
4)折扣因子;
5)贪婪因子最大值;
6)贪婪因子最小值;
7)贪婪因子变化步长。
所述网格搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的总搜索空间;
2)所述AI模型去激活策略的总搜索空间;
3)所述AI模型激活策略的总搜索空间;
4)所述AI模型训练策略的总搜索空间。
所述随机搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的最大个数;
2)所述AI模型去激活策略的最大个数;
3)所述AI模型激活策略的最大个数;
4)所述AI模型训练策略的最大个数。
所述连续减半算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的数目和每个策略的验证时长;
2)所述AI模型去激活策略的数目和每个策略的验证时长;
3)所述AI模型激活策略的数目和每个策略的验证时长;
4)所述AI模型训练策略的数目和每个策略的验证时长。
所述Hyperband算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
2)所述AI模型去激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
3)所述AI模型激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
4)所述AI模型训练策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。
下面举例说明本申请实施例提供的AI模型策略确定方法。
实施例一、AI模型部署策略选择的可调参数集合。
表1示出了AI模型部署策略选择的可调参数集合。
表1 AI模型部署策略选择的可调参数集合
第一设备执行完自动化算法后,产生一个待验证的AI模型部署策略。在每个部署策略中,每个可调参数项会选出一个值。
例如,基站波束个数为16,从{8,16,32,64}中选出16;
终端波束个数为8,从{2,4,8,16}中选出8;
终端天线面板个数为2,从{1,2,3}中选出2;
视距、非视距场景为视距场景,从{视距场景,混合场景}中选出视距场景;
终端移动速度为30km/h,从{30km/h,60km/h,90km/h}中选出30km/h;
用于推理的AI模型为模型标识2的模型,从{模型标识1,模型标识2}中选出模型标识2。
第一设备将待验证的AI模型部署策略部署到对应的基站和终端,进而验证本次AI模型部署策略的性能。
实施例二、AI模型去激活策略选择的可调参数集合。
表2示出了AI模型去激活策略选择的可调参数集合。
表2 AI模型去激活策略选择的可调参数集合
第一设备执行完自动化算法后,产生一个待验证的AI模型去激活策略。在每个策略中,每个可调参数项会选出一个值。
(一)以监视指标-最优波束预测准确率为例进行说明:
用于指示监视指标是否激活的指示信息为1,从{0,1}中选择,表示激活这项监视指标。
指标门限为97%,从{90%,95%,97%}中选择;
样本周期为1000,从{1000,2000,5000}中选择;
样本占比为5%,从{5%,10%,20%,100%}中选择;
样本分布为前置分布,从{随机分布,前置分布,后置分布}中选择。在本例中,前置分布表示从1000个样本中,选择前50个样本用于监视,并计算最优波束预测平均准确率。
第一数量门限(事件激活数量门限)为2,从{1,2,4}中选择。在本例中,表示从1000个样本中的前50个样本计算的平均准确率如果小于97%,则第一计数器(事件激活计数器)累加为1;当第一计数器累加到2时,则进行AI模型去激活。
第一计数器的重置方式(计数器重置规则)为事件重置,从{周期性重置,事件重置}中选择。在本例中,事件重置表示第一计数器可以在满足特定事件后进行重置。
重置门限为99%,从{99%,100%}中选择。在本例中,表示如果1000个样本中前50个的平均准确率大于99%,则可以累计重置计数器(第二计数器)加1;
重置计数器门限为2,从{1,2,4}中选择,表示当第二计数器(重置计数器)累加到2后,可以对第一计数器进行重置。
(二)以监视指标-吞吐量为例进行说明:
用于指示监视指标是否激活的指示信息为0,从{0,1}中选择,表示不考虑这项监视指标。
(三)以监视指标-最优波束预测准确率为例进行说明:
用于指示监视指标是否激活的指示信息为0,从{0,1}中选择,表示不考虑这项监视指标。
(四)以监视指标-吞吐量为例进行说明:
用于指示监视指标是否激活的指示信息为1,从{0,1}中选择,表示激活这项监视指标。
指标门限为20%,从{10%,20%,30%}中选择;
样本周期为1000,从{1000,2000,5000}中选择;
样本占比为5%,从{5%,10%,20%,100%}中选择;
样本分布为前置分布,从{随机分布,前置分布,后置分布}中选择。在本例中,前置分布表示从1000个样本中,选择前50个样本用于监视,并计算平均吞吐量相比与基线方案的增益。
第一数量门限(事件激活数量门限)为1,从{1,2,4}中选择。在本例中,表示从1000个样本中的前50个样本计算的平均吞吐量相比与基线方案的增益如果小于20%,则第一计数器累加为1;当第一计数器累加到1时,则进行AI模型去激活。
第一计数器的重置方式(计数器重置规则)为周期性重置,从{周期性重置}中选择。此时只配置了一种可选项目,则固定选择周期性重置。
周期性重置的重置周期(计数器重置周期)为40000,从{10000,20000,40000}中选择,在此例中表示,如果40000内,第一计数器没有达到第一数量门限时,则重置第一计数器为0。
第一设备将待验证的AI模型去激活策略部署到对应的功能实体,进而验证本次策略的性能。
实施例三、第一设备是基站DU,第二设备是基站CU。
第一设备是基站DU,第二设备是基站CU,算法配置信息对应的自动化算法为强化学习算法,图4是本申请实施例提供的AI模型策略确定方法的流程示意图之三,如图4所示。
步骤1:基站DU接收基站CU发送的算法配置信息和可调参数配置信息。
其中,算法配置信息包括强化学习收集奖励的实体用户设备(User Equipment,UE)、强化学习超参数折扣因子、训练次数;
可调参数配置信息包括网络配置信息、终端配置信息、场景配置信息(包括网络场景信息和终端场景信息)、候选AI模型信息,以及对应的取值集合。
步骤2:基站DU生成可调参数集合,执行自动化算法;
步骤3:基站DU基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,基站DU把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做关键绩效指标(Key Performance Indicator,KPI)监视,UE反馈KPI结果给DU,KPI例如为最优波束预测准确率、波束管理损失、吞吐量。
步骤7:DU把KPI结果换算为reward,对自身控制器网络参数进行更新。
实施例四、第一设备是基站DU,第二设备是基站CU的新增节点,作为与基站CU平 级的实体。
第一设备是基站DU,第二设备是基站CU的新增节点,算法配置信息对应的自动化算法为随机搜索算法。
图5是本申请实施例提供的AI模型策略确定方法的流程示意图之四,如图5所示。
步骤1:基站DU接收基站CU的新增节点发送的算法配置信息和可调参数配置信息。
其中,算法配置信息包括随机搜索的最大次数;可调参数配置信息包括基站配置,终端配置以及对应取值集合。
步骤2:基站DU生成可调参数集合,执行自动化算法。
具体地,基站DU确定场景配置信息和候选AI模型信息,以及对应取值集合,作为可调参数集合。
步骤3:基站DU自动化算法输出的策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,基站DU把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做KPI监视,UE反馈KPI结果给DU,KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:DU记录本次策略和对应的KPI,DU把KPI结果换算为reward,对自身控制器网络参数进行更新。
实施例五、第一设备是基站CU,第二设备是基站CU的新增节点,作为与基站CU平级的实体。
第一设备是基站CU,第二设备是基站CU的新增节点,算法配置信息对应的自动化算法为网格搜索算法。
图6是本申请实施例提供的AI模型策略确定方法的流程示意图之五,如图6所示。
步骤1:基站CU接收基站CU的新增节点发送的算法配置信息和可调参数配置信息。
需要说明的是,对于网格搜索算法来说,算法配置信息可以与可调参数配置信息相同,这是因为对于网格搜索算法,目标策略的总搜索空间可以与可调参数项的取值集合相同。
其中,可调参数配置信息包括基站配置信息、终端配置信息、场景配置信息、候选AI模型信息,以及对应的取值集合。
步骤2:基站CU生成可调参数集合,执行自动化算法;
步骤3:基站CU基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,基站CU把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做KPI监视,UE反馈KPI结果给CU。KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:CU记录本次策略和对应的KPI。
在网格搜索结束后,CU根据记录的策略和对应的KPI,选择出KPI最优的策略。
实施例六、第一设备是基站DU,第二设备是核心网(Core Network,CN)功能。
第一设备是基站DU,第二设备是核心网功能,算法配置信息对应的自动化算法为连续减半算法。
图7是本申请实施例提供的AI模型策略确定方法的流程示意图之六,如图7所示。
步骤1:基站DU通过CU接收核心网功能发送的算法配置信息和可调参数配置信息。
可选地,CU可以为基站CU,或与基站CU平级的实体(例如为基站CU的新增节点)。
其中,算法配置信息包括目标策略的数目和每个策略的验证时长。
可调参数配置信息包括了基站配置信息、终端配置信息、场景配置信息、候选AI模型信息,以及每个可调参数项的取值集合。
步骤2:基站DU生成可调参数集合,执行自动化算法。
步骤3:基站DU基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在UE推理,基站DU把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做KPI监视,UE反馈KPI结果给DU,KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:DU记录本次策略和对应的KPI。
在连续减半算法结束后,DU根据记录的策略和对应的KPI,选择出KPI最优的策略。
实施例七、第一设备是基站CU,第二设备是核心网功能。
第一设备是基站CU,第二设备是核心网功能,算法配置信息对应的自动化算法为网格搜索算法。
图8是本申请实施例提供的AI模型策略确定方法的流程示意图之七,如图8所示。
步骤1:基站CU接收核心网功能发送的算法配置信息和可调参数配置信息。
需要说明的是,对于网格搜索算法来说,算法配置信息可以与可调参数配置信息相同,这是因为对于网格搜索算法,目标策略的总搜索空间可以与可调参数项的取值集合相同。
其中,可调参数配置信息包括基站配置信息、终端配置信息、场景配置信息、候选AI模型信息,以及对应的取值集合。
步骤2:基站CU生成可调参数集合,执行自动化算法;
步骤3:基站CU基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,基站CU把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做KPI监视,UE反馈KPI结果给CU,KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:CU记录本次策略和对应的KPI。
在网格搜索算法结束后,CU根据记录的策略和对应的KPI,选择出KPI最优的策略。
实施例八、第一设备是基站,第二设备是核心网功能。
第一设备是基站,第二设备是核心网功能,算法配置信息对应的自动化算法为随机搜索算法。
图9是本申请实施例提供的AI模型策略确定方法的流程示意图之八,如图9所示。
步骤1:基站接收核心网功能发送的算法配置信息和可调参数配置信息。
其中,算法配置信息包括目标策略的最大搜索数目。可调参数配置信息包括基站配置信息、终端配置信息、场景配置信息、候选AI模型信息,以及对应的取值集合。
步骤2:基站生成可调参数集合,执行自动化算法;
步骤3:基站基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,基站把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做KPI监视,UE反馈KPI结果给基站,KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:基站记录本次策略和对应的KPI。
在随机搜索算法结束后,基站根据记录的策略和对应的KPI,选择出KPI最优的策略。
实施例九、第一设备是基站CU平级的新增节点,第二设备是核心网功能。
第一设备是基站CU平级的新增节点,第二设备是核心网功能,算法配置信息对应的自动化算法为Hyperband算法。
图10是本申请实施例提供的AI模型策略确定方法的流程示意图之九,如图10所示。
步骤1:基站CU平级的新增节点接收核心网功能发送的算法配置信息和可调参数配置信息。
其中,算法配置信息包括目标策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。可调参数配置信息包括基站配置信息、终端配置信息、场 景配置信息、候选AI模型信息,以及对应的取值集合。
步骤2:基站CU平级的新增节点生成可调参数集合,执行自动化算法;
步骤3:基站CU平级的新增节点基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,基站CU平级的新增节点把激活条件发送给UE;
步骤5:UE根据终端场景信息激活AI模型;
例如,UE对比自身的移动速度,与终端场景信息中配置的终端移动速度是否匹配,并基于匹配结果确定是否激活AI模型。
步骤6:假设UE做KPI监视,UE反馈KPI结果给CU平级的新增节点,KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:基站CU平级的新增节点记录本次策略和对应的KPI。
在Hyperband算法结束后,基站CU平级的新增节点根据记录的策略和对应的KPI,选择出KPI最优的策略。
实施例十、第一设备是核心网功能,第二设备是核心网功能。
第一设备是核心网功能的网络功能实体(Network Function,NF)2,第二设备是另一个核心网功能的NF1,算法配置信息对应的自动化算法为网格搜索算法;两个设备可以都是现有的核心网功能,或者都是新增的核心网功能,或者一个是新增的核心网功能,一个是现有的核心网功能。
图11是本申请实施例提供的AI模型策略确定方法的信令交互图,如图11所示,图中仅示出了步骤1。
步骤1:NF2接收NF1发送的算法配置信息和可调参数配置信息。
需要说明的是,对于网格搜索算法来说,算法配置信息可以与可调参数配置信息相同,这是因为对于网格搜索算法,目标策略的总搜索空间可以与可调参数项的取值集合相同。
其中,可调参数配置信息包括基站配置信息、终端配置信息、场景配置信息、候选AI模型信息,以及对应的取值集合。
步骤2:NF2生成可调参数集合,执行自动化算法;
步骤3:NF2基于自动化算法输出的AI模型策略,确定基站和UE的激活条件;
步骤4:假设AI模型在进行UE推理,NF2把激活条件发送给UE;
步骤5:UE根据场景激活AI模型;
步骤6:假设UE做KPI监视,UE反馈KPI结果给NF2。KPI例如最优波束预测准确率、波束管理损失、吞吐量。
步骤7:NF2记录本次策略和对应的KPI。
在网格搜索算法结束后,NF2根据记录的策略和对应的KPI,选择出KPI最优的策略。
实施例十一、AI模型生命周期管理流程示意图。
图12是本申请实施例提供的AI模型的生命周期管理的流程示意图,如图12所示,现有5G系统研究的AI模型生命周期管理包括:模型训练(图中步骤21至步骤27)、模型部署(图中步骤3至步骤6)、模型传递(图中步骤5)、模型激活(图中步骤7和步骤8)、模型推理(图中步骤9和步骤10)、模型监视(图中步骤11至步骤16)、模型去激活(图中步骤17至步骤18)、模型回退(图中步骤20)、模型切换(图中步骤19)等。而在AI功能部署初期,即在实验阶段(图中步骤1和步骤2),AI模型策略选择、AI模型激活条件选择、AI模型去激活条件选择、AI模型监视配置以及AI模型训练控制仍需要人为干预。
本发明实施例中,第一设备接收第二设备发送的算法配置信息和/或可调参数配置信息,第一设备生成可调参数集合,和/或,第一设备执行算法配置信息对应的自动化算法。
其中,算法配置信息包括:
a)强化学习参数、reward收集位置(第一设备采用强化学习算法);
b)超参数优化参数(第一设备采用网格搜索、随机搜索、连续减半、Hyperband等算法);
可调参数配置信息包括:
a)用于产生策略选择的可调参数项和每个可调参数项的至少一个取值,例如,基站配置信息、终端配置信息、场景配置信息、候选AI模型信息等;
b)用于去激活的可调参数项和每个可调参数项的至少一个取值,例如,监视指标、指 标计算相关的计时器和计数器、指标门限等。
本发明实施例中,通过将第二设备的算法配置信息和可调参数配置信息发送给第一设备,使得无线网络可以自动化地进行AI模型策略选择、AI模型激活条件选择、AI模型去激活条件选择、AI模型监视配置,以及AI模型训练控制等操作。从而避免了人工干预,提升了系统运行效率及AI模型的上线效率。
本申请实施例提供的AI模型策略确定方法,执行主体可以为AI模型策略确定装置。本申请实施例中以AI模型策略确定装置执行AI模型策略确定方法为例,说明本申请实施例提供的AI模型策略确定装置。
图13是本申请实施例提供的AI模型策略确定装置的结构示意图之一,如图13所示,AI模型策略确定装置1300包括:
获取模块1301,用于获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;
确定模块1302,用于基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;
处理模块1303,用于根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
在本申请实施例提供的AI模型策略确定装置中,获取模块在获取到可调参数集合和/或算法配置信息之后,确定模块可以基于可调参数集合和/或算法配置信息确定针对AI功能的目标策略,提高目标策略的确定效率,进而处理模块根据目标策略对相应的AI模型进行处理,为终端提供AI服务,能够提高通信系统运行效率。
可选地,获取模块1301具体用于:
1、接收来自第二设备的所述可调参数集合和/或所述算法配置信息;
2、接收来自第二设备的可调参数配置信息和/或所述算法配置信息,基于所述可调参数配置信息,确定所述可调参数集合;
3、接收来自第二设备的所述可调参数集合;接收来自第三设备的所述算法配置信息;
4、接收来自第二设备的可调参数配置信息,基于所述可调参数配置信息,确定所述可调参数集合;接收来自第三设备的所述算法配置信息;
其中,所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值。
可选地,获取模块1301还具体用于:
1)在所述可调参数配置信息包括所述N个可调参数项中的全部可调参数项,及各可调参数项的至少一个取值的情况下,将所述可调参数配置信息确定为所述可调参数集合;
2)在所述可调参数配置信息包括所述N个可调参数项中的全部可调参数项的情况下,基于参数项配置信息,确定各可调参数项的至少一个取值;基于所述可调参数配置信息及各可调参数项的至少一个取值,确定所述可调参数集合;
3)在所述可调参数配置信息包括所述N个可调参数项中的全部可调参数项,及所述N个可调参数项中的部分可调参数项的至少一个取值的情况下,基于参数项配置信息,确定所述N个可调参数项中的剩余可调参数项的至少一个取值;基于所述可调参数配置信息及所述剩余可调参数项的至少一个取值,确定所述可调参数集合;
4)在所述可调参数配置信息包括所述N个可调参数项中的部分可调参数项,及所述部分可调参数项中的至少一个可调参数项的至少一个取值的情况下,基于参数项配置信息,确定第一可调参数项及所述N个可调参数项中无取值的可调参数项的至少一个取值;基于所述可调参数配置信息、所述第一可调参数项及所述无取值的可调参数项的至少一个取值,确定所述可调参数集合;所述第一可调参数项为所述N个可调参数项中除所述部分可调参数项之外的可调参数项。
可选地,所述可调参数配置信息可以包括以下至少一项:
1、第一配置信息,用于辅助确定所述AI模型部署策略;
2、第二配置信息,用于辅助确定所述AI模型激活策略;
3、第三配置信息,用于辅助确定所述AI模型去激活策略;
4、第四配置信息,用于辅助确定所述AI模型训练策略。
可选地,所述第一配置信息可以包括以下至少一项:
1)网络配置信息;
2)终端配置信息;
3)网络场景信息;
4)终端场景信息;
5)候选AI模型信息;
6)所述AI功能对应的候选方案;
7)候选推理侧信息。
其中,所述网络配置信息可以包括以下至少一项:
a.基站天线配置信息;
b.基站波束配置信息;
c.基站波束个数信息;
d.基站高度信息;
e.基站站间距信息。
所述终端配置信息可以包括以下至少一项:
a)终端天线配置信息;
b)终端波束配置信息;
c)终端波束个数信息;
d)终端天线面板个数信息。
所述网络场景信息可以包括以下至少一项:
a.视距场景信息;
b.非视距场景信息;
c.室外场景信息;
d.室内场景信息;
e.用户密集场景信息;
f.用户稀疏场景信息;
g.不同降雨量场景信息。
所述终端场景信息包括以下至少一项:
a)终端移动速度;
b)终端旋转速度。
可选地,所述第二配置信息包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
可选地,所述第三配置信息可以包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型去激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及去激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,去激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
可选地,所述第一参数可以包括以下至少一项:
a.样本周期;
b.样本占比;
c.样本分布。
可选地,所述第四配置信息可以包括以下至少一项:
1)模型输入参数;
2)预处理策略;
3)后处理策略;
4)模型输出参数;
5)训练超参数;
6)模型网络结构。
可选地,所述算法配置信息可以包括以下至少一项:
1、强化学习算法配置信息;
2、网格搜索算法配置信息;
3、随机搜索算法配置信息;
4、连续减半算法配置信息;
5、Hyperband算法配置信息。
其中,所述强化学习算法配置信息可以包括以下至少一项:
1)收集奖励的实体;
2)奖励计算公式;
3)奖励对应的参数;
4)折扣因子;
5)贪婪因子最大值;
6)贪婪因子最小值;
7)贪婪因子变化步长。
所述网格搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的总搜索空间;
2)所述AI模型去激活策略的总搜索空间;
3)所述AI模型激活策略的总搜索空间;
4)所述AI模型训练策略的总搜索空间。
所述随机搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的最大个数;
2)所述AI模型去激活策略的最大个数;
3)所述AI模型激活策略的最大个数;
4)所述AI模型训练策略的最大个数。
所述连续减半算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的数目和每个策略的验证时长;
2)所述AI模型去激活策略的数目和每个策略的验证时长;
3)所述AI模型激活策略的数目和每个策略的验证时长;
4)所述AI模型训练策略的数目和每个策略的验证时长。
所述Hyperband算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
2)所述AI模型去激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
3)所述AI模型激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
4)所述AI模型训练策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。
可选地,所述第一设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体;
(6)基站DU。
可选地,所述第二设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体。
图14是本申请实施例提供的AI模型策略确定装置的结构示意图之二,如图14所示,AI模型策略确定装置1400包括:
发送模块1401,用于向第一设备发送以下任一项:
1、可调参数集合和算法配置信息;
2、可调参数配置信息和算法配置信息;
3、可调参数集合;
4、可调参数配置信息;
5、算法配置信息;
其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针 对AI功能的目标策略;
所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
所述目标策略包括以下至少一项:
1)AI模型部署策略;
2)AI模型去激活策略;
3)AI模型激活策略;
4)AI模型训练策略。
在本申请实施例提供的AI模型策略确定装置中,由发送模块向第一设备发送以下任一项:可调参数集合和算法配置信息;可调参数配置信息和算法配置信息;可调参数集合;可调参数配置信息;算法配置信息;以由第一设备获取到可调参数集合和/或算法配置信息之后,基于可调参数集合和/或算法配置信息确定针对AI功能的目标策略,提高目标策略的确定效率,进而第一设备根据目标策略对相应的AI模型进行处理,为终端提供AI服务,能够提高通信系统运行效率。
可选地,所述可调参数配置信息可以包括以下至少一项:
1、第一配置信息,用于辅助确定所述AI模型部署策略;
2、第二配置信息,用于辅助确定所述AI模型激活策略;
3、第三配置信息,用于辅助确定所述AI模型去激活策略;
4、第四配置信息,用于辅助确定所述AI模型训练策略。
可选地,所述第一配置信息可以包括以下至少一项:
1)网络配置信息;
2)终端配置信息;
3)网络场景信息;
4)终端场景信息;
5)候选AI模型信息;
6)所述AI功能对应的候选方案;
7)候选推理侧信息。
其中,所述网络配置信息可以包括以下至少一项:
a.基站天线配置信息;
b.基站波束配置信息;
c.基站波束个数信息;
d.基站高度信息;
e.基站站间距信息。
所述终端配置信息可以包括以下至少一项:
a)终端天线配置信息;
b)终端波束配置信息;
c)终端波束个数信息;
d)终端天线面板个数信息。
所述网络场景信息可以包括以下至少一项:
a.视距场景信息;
b.非视距场景信息;
c.室外场景信息;
d.室内场景信息;
e.用户密集场景信息;
f.用户稀疏场景信息;
g.不同降雨量场景信息。
所述终端场景信息可以包括以下至少一项:
a)终端移动速度;
b)终端旋转速度。
可选地,所述第二配置信息包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
可选地,所述第三配置信息可以包括以下至少一项:
1)至少一个监视指标;
2)指示信息,用于指示监视指标是否激活;
3)第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
4)指标门限;
5)启动条件,用于指示在开启模型去激活评估流程的情况下,启动第一计时器;
6)第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及去激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
7)第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,去激活AI模型;
8)次数累计方式,包括可重置的累计方式或不可重置的累计方式;
9)第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
10)重置门限;
11)第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
12)周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
13)第二时间门限;
14)计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
可选地,所述第一参数可以包括以下至少一项:
a.样本周期;
b.样本占比;
c.样本分布。
可选地,所述第四配置信息可以包括以下至少一项:
1)模型输入参数;
2)预处理策略;
3)后处理策略;
4)模型输出参数;
5)训练超参数;
6)模型网络结构。
可选地,所述算法配置信息可以包括以下至少一项:
1、强化学习算法配置信息;
2、网格搜索算法配置信息;
3、随机搜索算法配置信息;
4、连续减半算法配置信息;
5、Hyperband算法配置信息。
其中,所述强化学习算法配置信息可以包括以下至少一项:
1)收集奖励的实体;
2)奖励计算公式;
3)奖励对应的参数;
4)折扣因子;
5)贪婪因子最大值;
6)贪婪因子最小值;
7)贪婪因子变化步长。
所述网格搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的总搜索空间;
2)所述AI模型去激活策略的总搜索空间;
3)所述AI模型激活策略的总搜索空间;
4)所述AI模型训练策略的总搜索空间。
所述随机搜索算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的最大个数;
2)所述AI模型去激活策略的最大个数;
3)所述AI模型激活策略的最大个数;
4)所述AI模型训练策略的最大个数。
所述连续减半算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的数目和每个策略的验证时长;
2)所述AI模型去激活策略的数目和每个策略的验证时长;
3)所述AI模型激活策略的数目和每个策略的验证时长;
4)所述AI模型训练策略的数目和每个策略的验证时长。
所述Hyperband算法配置信息可以包括以下至少一项:
1)所述AI模型部署策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
2)所述AI模型去激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
3)所述AI模型激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
4)所述AI模型训练策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。
可选地,所述第一设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体;
(6)基站DU。
可选地,所述第二设备可以包括以下至少一项:
(1)自动化流程控制实体;
(2)核心网功能;
(3)基站;
(4)基站CU;
(5)与基站CU平级的实体。
本申请实施例中的AI模型策略确定装置可以是电子设备,例如具有操作系统的电子设备,也可以是电子设备中的部件,例如集成电路或芯片。该电子设备可以是网络侧设备,也可以为除网络侧设备之外的其他设备。示例性地,网络侧设备可以包括但不限于上述所列举的网络侧设备12的类型,其他设备可以为服务器、网络附属存储器(Network Attached Storage,NAS)等,本申请实施例不作具体限定。
本申请实施例提供的AI模型策略确定装置能够实现图2至图12的方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
可选地,图15是本申请实施例提供的通信设备的结构示意图,如图15所示,本申请实施例还提供一种通信设备1500,包括处理器1501和存储器1502,存储器1502上存储有可在所述处理器1501上运行的程序或指令,例如,该通信设备1500为第一设备时,该程序或指令被处理器1501执行时实现上述第一设备侧的AI模型策略确定方法实施例的各个步骤,且能达到相同的技术效果。该通信设备1500为第二设备时,该程序或指令被处理器1501执行时实现上述第二设备侧的AI模型策略确定方法实施例的各个步骤,且能达到相同的技术效果,为避免重复,这里不再赘述。
本申请实施例还提供一种第一设备,包括处理器和通信接口,处理器用于:获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;处理器还用于:基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;处理器还用于:根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
该第一设备实施例与上述第一设备侧方法实施例对应,上述方法实施例的各个实施过程和实现方式均可适用于该第一设备实施例中,且能达到相同的技术效果。
具体地,本申请实施例还提供了一种第一设备。图16是本申请实施例提供的第一设备的结构示意图之一,如图16所示,该第一设备1600包括:天线1601、射频装置1602、基带装置1603、处理器1604和存储器1605。天线1601与射频装置1602连接。在上行方向上,射频装置1602通过天线1601接收信息,将接收的信息发送给基带装置1603进行处理。在下行方向上,基带装置1603对要发送的信息进行处理,并发送给射频装置1602,射频装置1602对收到的信息进行处理后经过天线1601发送出去。
以上实施例中第一设备执行的方法可以在基带装置1603中实现,该基带装置1603包括基带处理器。
基带装置1603例如可以包括至少一个基带板,该基带板上设置有多个芯片,如图16所示,其中一个芯片例如为基带处理器,通过总线接口与存储器1605连接,以调用存储器1605中的程序,执行以上方法实施例中所示的网络设备操作。
该第一设备还可以包括网络接口1606,该接口例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本申请实施例的第一设备1600还包括:存储在存储器1605上并可在处理器1604上运行的指令或程序,处理器1604调用存储器1605中的指令或程序执行图13所示各模块执行的方法,并达到相同的技术效果,为避免重复,故不在此赘述。
具体地,本申请实施例还提供了一种第一设备。图17是本申请实施例提供的第一设备的结构示意图之二,如图17所示,该第一设备1700包括:处理器1701、网络接口1702和存储器1703。其中,网络接口1702例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本申请实施例的第一设备1700还包括:存储在存储器1703上并可在处理器1701上运行的指令或程序,处理器1701调用存储器1703中的指令或程序执行图13所示各模块执行的方法,并达到相同的技术效果,为避免重复,故不在此赘述。
本申请实施例还提供一种第二设备,包括处理器和通信接口,通信接口用于:向第一设备发送以下任一项:
可调参数集合和算法配置信息;
可调参数配置信息和算法配置信息;
可调参数集合;
可调参数配置信息;
算法配置信息;
其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针对AI功能的目标策略;
所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略。
该第二设备实施例与上述第二设备侧方法实施例对应,上述方法实施例的各个实施过程和实现方式均可适用于该第二设备实施例中,且能达到相同的技术效果。
具体地,本申请实施例还提供了一种第二设备。图18是本申请实施例提供的第二设备的结构示意图之一,如图18所示,该第二设备1800包括:天线1801、射频装置1802、基带装置1803、处理器1804和存储器1805。天线1801与射频装置1802连接。在上行方向上,射频装置1802通过天线1801接收信息,将接收的信息发送给基带装置1803进行处理。在下行方向上,基带装置1803对要发送的信息进行处理,并发送给射频装置1802,射频装置1802对收到的信息进行处理后经过天线1801发送出去。
以上实施例中第二设备执行的方法可以在基带装置1803中实现,该基带装置1803包括基带处理器。
基带装置1803例如可以包括至少一个基带板,该基带板上设置有多个芯片,如图18所示,其中一个芯片例如为基带处理器,通过总线接口与存储器1805连接,以调用存储器1805中的程序,执行以上方法实施例中所示的网络设备操作。
该第二设备还可以包括网络接口1806,该接口例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本申请实施例的第二设备1800还包括:存储在存储器1805上并可在处理器1804上运行的指令或程序,处理器1804调用存储器1805中的指令或程序执行图14所示各模块执行的方法,并达到相同的技术效果,为避免重复,故不在此赘述。
具体地,本申请实施例还提供了一种第二设备。图19是本申请实施例提供的第二设备的结构示意图之二,如图19所示,该第二设备1900包括:处理器1901、网络接口1902和存储器1903。其中,网络接口1902例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本申请实施例的第二设备1900还包括:存储在存储器1903上并可在处理器1901上运行的指令或程序,处理器1901调用存储器1903中的指令或程序执行图14所示各模块执行的方法,并达到相同的技术效果,为避免重复,故不在此赘述。
本申请实施例还提供一种可读存储介质,所述可读存储介质上存储有程序或指令,该程序或指令被处理器执行时实现上述AI模型策略确定方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
其中,所述处理器为上述实施例中所述的网络侧设备中的处理器。所述可读存储介质,包括计算机可读存储介质,如计算机只读存储器ROM、随机存取存储器RAM、磁碟或者光盘等。
本申请实施例另提供了一种芯片,所述芯片包括处理器和通信接口,所述通信接口和 所述处理器耦合,所述处理器用于运行程序或指令,实现上述AI模型策略确定方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
应理解,本申请实施例提到的芯片还可以称为系统级芯片,系统芯片,芯片系统或片上系统芯片等。
本申请实施例另提供了一种计算机程序/程序产品,所述计算机程序/程序产品被存储在存储介质中,所述计算机程序/程序产品被至少一个处理器执行以实现上述AI模型策略确定方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
本申请实施例还提供了一种AI模型策略确定系统,包括:第一设备及第二设备,所述第一设备可用于执行如上所述的AI模型策略确定方法的步骤,所述第二设备可用于执行如上所述的AI模型策略确定方法的步骤。
需要说明的是,在本文中,术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。此外,需要指出的是,本申请实施方式中的方法和装置的范围不限按示出或讨论的顺序来执行功能,还可包括根据所涉及的功能按基本同时的方式或按相反的顺序来执行功能,例如,可以按不同于所描述的次序来执行所描述的方法,并且还可以添加、省去、或组合各种步骤。另外,参照某些示例所描述的特征可在其他示例中被组合。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以计算机软件产品的形式体现出来,该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端(可以是手机,计算机,服务器,空调器,或者网络设备等)执行本申请各个实施例所述的方法。
上面结合附图对本申请的实施例进行了描述,但是本申请并不局限于上述的具体实施方式,上述的具体实施方式仅仅是示意性的,而不是限制性的,本领域的普通技术人员在本申请的启示下,在不脱离本申请宗旨和权利要求所保护的范围情况下,还可做出很多形式,均属于本申请的保护之内。

Claims (27)

  1. 一种人工智能AI模型策略确定方法,其中,包括:
    第一设备获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;
    所述第一设备基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;
    所述第一设备根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
  2. 根据权利要求1所述的方法,其中,所述第一设备获取可调参数集合和/或算法配置信息,包括以下至少一项:
    所述第一设备接收来自第二设备的所述可调参数集合和/或所述算法配置信息;
    所述第一设备接收来自第二设备的可调参数配置信息和/或所述算法配置信息,所述第一设备基于所述可调参数配置信息,确定所述可调参数集合;
    所述第一设备接收来自第二设备的所述可调参数集合;所述第一设备接收来自第三设备的所述算法配置信息;
    所述第一设备接收来自第二设备的可调参数配置信息,所述第一设备基于所述可调参数配置信息,确定所述可调参数集合;所述第一设备接收来自第三设备的所述算法配置信息;
    其中,所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值。
  3. 根据权利要求2所述的方法,其中,所述第一设备基于所述可调参数配置信息,确定所述可调参数集合,包括以下至少一项:
    在所述可调参数配置信息包括所述N个可调参数项中的全部可调参数项,及各可调参数项的至少一个取值的情况下,所述第一设备将所述可调参数配置信息确定为所述可调参数集合;
    在所述可调参数配置信息包括所述N个可调参数项中的全部可调参数项的情况下,所述第一设备基于参数项配置信息,确定各可调参数项的至少一个取值;所述第一设备基于所述可调参数配置信息及各可调参数项的至少一个取值,确定所述可调参数集合;
    在所述可调参数配置信息包括所述N个可调参数项中的全部可调参数项,及所述N个可调参数项中的部分可调参数项的至少一个取值的情况下,所述第一设备基于参数项配置信息,确定所述N个可调参数项中的剩余可调参数项的至少一个取值;所述第一设备基于所述可调参数配置信息及所述剩余可调参数项的至少一个取值,确定所述可调参数集合;
    在所述可调参数配置信息包括所述N个可调参数项中的部分可调参数项,及所述部分可调参数项中的至少一个可调参数项的至少一个取值的情况下,所述第一设备基于参数项配置信息,确定第一可调参数项及所述N个可调参数项中无取值的可调参数项的至少一个取值;所述第一设备基于所述可调参数配置信息、所述第一可调参数项及所述无取值的可调参数项的至少一个取值,确定所述可调参数集合;所述第一可调参数项为所述N个可调参数项中除所述部分可调参数项之外的可调参数项。
  4. 根据权利要求2或3所述的方法,其中,所述可调参数配置信息包括以下至少一项:
    第一配置信息,用于辅助确定所述AI模型部署策略;
    第二配置信息,用于辅助确定所述AI模型激活策略;
    第三配置信息,用于辅助确定所述AI模型去激活策略;
    第四配置信息,用于辅助确定所述AI模型训练策略。
  5. 根据权利要求4所述的方法,其中,所述第一配置信息包括以下至少一项:网络配置信息;终端配置信息;网络场景信息;终端场景信息;候选AI模型信息;所述AI功能对应的候选方案;候选推理侧信息;其中,
    所述网络配置信息包括以下至少一项:
    基站天线配置信息;
    基站波束配置信息;
    基站波束个数信息;
    基站高度信息;
    基站站间距信息;
    所述终端配置信息包括以下至少一项:
    终端天线配置信息;
    终端波束配置信息;
    终端波束个数信息;
    终端天线面板个数信息;
    所述网络场景信息包括以下至少一项:
    视距场景信息;
    非视距场景信息;
    室外场景信息;
    室内场景信息;
    用户密集场景信息;
    用户稀疏场景信息;
    不同降雨量场景信息;
    所述终端场景信息包括以下至少一项:
    终端移动速度;
    终端旋转速度。
  6. 根据权利要求4所述的方法,其中,所述第二配置信息包括以下至少一项:
    至少一个监视指标;
    指示信息,用于指示监视指标是否激活;
    第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
    指标门限;
    启动条件,用于指示在开启模型激活评估流程的情况下,启动第一计时器;
    第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及激活AI模型;第一计数器用于记录监视指标满足指标门限的累计次数;
    第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,激活AI模型;
    次数累计方式,包括可重置的累计方式或不可重置的累计方式;
    第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
    重置门限;
    第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;第二计数器用于记录监视指标满足重置门限的累计次数;
    周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
    第二时间门限;
    计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
  7. 根据权利要求4所述的方法,其中,所述第三配置信息包括以下至少一项:
    至少一个监视指标;
    指示信息,用于指示监视指标是否激活;
    第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
    指标门限;
    启动条件,用于指示在开启模型去激活评估流程的情况下,启动第一计时器;
    第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及去激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
    第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,去激活AI模型;
    次数累计方式,包括可重置的累计方式或不可重置的累计方式;
    第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
    重置门限;
    第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
    周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
    第二时间门限;
    计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
  8. 根据权利要求6或7所述的方法,其中,所述第一参数包括以下至少一项:
    样本周期;
    样本占比;
    样本分布。
  9. 根据权利要求4所述的方法,其中,所述第四配置信息包括以下至少一项:
    模型输入参数;
    预处理策略;
    后处理策略;
    模型输出参数;
    训练超参数;
    模型网络结构。
  10. 根据权利要求1至9任一项所述的方法,其中,所述算法配置信息包括以下至少一项:强化学习算法配置信息;网格搜索算法配置信息;随机搜索算法配置信息;连续减半算法配置信息;Hyperband算法配置信息;其中,
    所述强化学习算法配置信息包括以下至少一项:
    收集奖励的实体;
    奖励计算公式;
    奖励对应的参数;
    折扣因子;
    贪婪因子最大值;
    贪婪因子最小值;
    贪婪因子变化步长;
    所述网格搜索算法配置信息包括以下至少一项:
    所述AI模型部署策略的总搜索空间;
    所述AI模型去激活策略的总搜索空间;
    所述AI模型激活策略的总搜索空间;
    所述AI模型训练策略的总搜索空间;
    所述随机搜索算法配置信息包括以下至少一项:
    所述AI模型部署策略的最大个数;
    所述AI模型去激活策略的最大个数;
    所述AI模型激活策略的最大个数;
    所述AI模型训练策略的最大个数;
    所述连续减半算法配置信息包括以下至少一项:
    所述AI模型部署策略的数目和每个策略的验证时长;
    所述AI模型去激活策略的数目和每个策略的验证时长;
    所述AI模型激活策略的数目和每个策略的验证时长;
    所述AI模型训练策略的数目和每个策略的验证时长;
    所述Hyperband算法配置信息包括以下至少一项:
    所述AI模型部署策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
    所述AI模型去激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
    所述AI模型激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
    所述AI模型训练策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。
  11. 根据权利要求1至10任一项所述的方法,其中,所述第一设备包括以下至少一项:
    自动化流程控制实体;
    核心网功能;
    基站;
    基站中心单元CU;
    与基站CU平级的实体;
    基站分布单元DU。
  12. 根据权利要求3所述的方法,其中,所述第二设备包括以下至少一项:
    自动化流程控制实体;
    核心网功能;
    基站;
    基站CU;
    与基站CU平级的实体。
  13. 一种人工智能AI模型策略确定方法,其中,包括:
    第二设备向第一设备发送以下任一项:
    可调参数集合和算法配置信息;
    可调参数配置信息和算法配置信息;
    可调参数集合;
    可调参数配置信息;
    算法配置信息;
    其中,所述可调参数集合用于辅助确定AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
    所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定AI功能的目标策略;
    所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
    所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略。
  14. 根据权利要求13所述的方法,其中,所述可调参数配置信息包括以下至少一项:
    第一配置信息,用于辅助确定所述AI模型部署策略;
    第二配置信息,用于辅助确定所述AI模型激活策略;
    第三配置信息,用于辅助确定所述AI模型去激活策略;
    第四配置信息,用于辅助确定所述AI模型训练策略。
  15. 根据权利要求14所述的方法,其中,所述第一配置信息包括以下至少一项:网络配置信息;终端配置信息;网络场景信息;终端场景信息;候选AI模型信息;所述AI功能对应的候选方案;候选推理侧信息;其中,
    所述网络配置信息包括以下至少一项:
    基站天线配置信息;
    基站波束配置信息;
    基站波束个数信息;
    基站高度信息;
    基站站间距信息;
    所述终端配置信息包括以下至少一项:
    终端天线配置信息;
    终端波束配置信息;
    终端波束个数信息;
    终端天线面板个数信息;
    所述网络场景信息包括以下至少一项:
    视距场景信息;
    非视距场景信息;
    室外场景信息;
    室内场景信息;
    用户密集场景信息;
    用户稀疏场景信息;
    不同降雨量场景信息;
    所述终端场景信息包括以下至少一项:
    终端移动速度;
    终端旋转速度。
  16. 根据权利要求14所述的方法,其中,所述第二配置信息包括以下至少一项:
    至少一个监视指标;
    指示信息,用于指示监视指标是否激活;
    第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
    指标门限;
    启动条件,用于指示在开启模型激活评估流程的情况下,启动第一计时器;
    第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
    第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,激活AI模型;
    次数累计方式,包括可重置的累计方式或不可重置的累计方式;
    第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
    重置门限;
    第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
    周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
    第二时间门限;
    计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
  17. 根据权利要求14所述的方法,其中,所述第三配置信息包括以下至少一项:
    至少一个监视指标;
    指示信息,用于指示监视指标是否激活;
    第一参数,用于确定监视样本,所述监视样本用于计算监视指标;
    指标门限;
    启动条件,用于指示在开启模型去激活评估流程的情况下,启动第一计时器;
    第一数量门限,用于指示在第一计数器的计数值大于所述第一数量门限的情况下,启动第一计时器,或暂停第一计时器以及去激活AI模型;所述第一计数器用于记录监视指标满足指标门限的累计次数;
    第一时间门限,用于在第一计时器超过所述第一时间门限的情况下,去激活AI模型;
    次数累计方式,包括可重置的累计方式或不可重置的累计方式;
    第一计数器的重置方式,包括周期性重置、事件重置及计时器重置中至少一项;
    重置门限;
    第二数量门限,用于指示在第二计数器的计数值大于所述第二数量门限的情况下,重置第一计数器或暂停第一计时器;所述第二计数器用于记录监视指标满足重置门限的累计次数;
    周期性重置的重置周期,包括样本周期和/或时间周期;所述样本周期用于在连续的监视样本数目大于或等于所述样本周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述时间周期用于在第二计时器的连续计时超过所述时间周期,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;
    第二时间门限;
    计时器重置的重置门限,用于指示在第二计时器的累计计时超过第二时间门限,且第一计数器未达到第一数量门限的情况下,对第一计数器进行重置;所述第二计时器用于记录第一计数器的计数时间。
  18. 根据权利要求16或17所述的方法,其中,所述第一参数包括以下至少一项:
    样本周期;
    样本占比;
    样本分布。
  19. 根据权利要求14所述的方法,其中,所述第四配置信息包括以下至少一项:
    模型输入参数;
    预处理策略;
    后处理策略;
    模型输出参数;
    训练超参数;
    模型网络结构。
  20. 根据权利要求13至19任一项所述的方法,其中,所述算法配置信息包括以下至少一项:强化学习算法配置信息;网格搜索算法配置信息;随机搜索算法配置信息;连续减半算法配置信息;Hyperband算法配置信息;其中,
    所述强化学习算法配置信息包括以下至少一项:
    收集奖励的实体;
    奖励计算公式;
    奖励对应的参数;
    折扣因子;
    贪婪因子最大值;
    贪婪因子最小值;
    贪婪因子变化步长;
    所述网格搜索算法配置信息包括以下至少一项:
    所述AI模型部署策略的总搜索空间;
    所述AI模型去激活策略的总搜索空间;
    所述AI模型激活策略的总搜索空间;
    所述AI模型训练策略的总搜索空间;
    所述随机搜索算法配置信息包括以下至少一项:
    所述AI模型部署策略的最大个数;
    所述AI模型去激活策略的最大个数;
    所述AI模型激活策略的最大个数;
    所述AI模型训练策略的最大个数;
    所述连续减半算法配置信息包括以下至少一项:
    所述AI模型部署策略的数目和每个策略的验证时长;
    所述AI模型去激活策略的数目和每个策略的验证时长;
    所述AI模型激活策略的数目和每个策略的验证时长;
    所述AI模型训练策略的数目和每个策略的验证时长;
    所述Hyperband算法配置信息包括以下至少一项:
    所述AI模型部署策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
    所述AI模型去激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
    所述AI模型激活策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长;
    所述AI模型训练策略的连续减半算法的执行次数,以及初始连续减半算法中的策略的数目及验证时长。
  21. 根据权利要求13至20任一项所述的方法,其中,所述第一设备包括以下至少一项:
    自动化流程控制实体;
    核心网功能;
    基站;
    基站中心单元CU;
    与基站CU平级的实体;
    基站分布单元DU。
  22. 根据权利要求13至21任一项所述的方法,其中,所述第二设备包括以下至少一项:
    自动化流程控制实体;
    核心网功能;
    基站;
    基站CU;
    与基站CU平级的实体。
  23. 一种人工智能AI模型策略确定装置,其中,包括:
    获取模块,用于获取可调参数集合和/或算法配置信息;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值;所述算法配置信息用于指示目标算法的配置参数;N为正整数;
    确定模块,用于基于所述可调参数集合和/或所述算法配置信息,确定AI功能的目标策略;所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略;
    处理模块,用于根据所述目标策略对相应的AI模型进行处理,为终端提供AI服务。
  24. 一种人工智能AI模型策略确定装置,其中,包括:
    发送模块,用于向第一设备发送以下任一项:
    可调参数集合和算法配置信息;
    可调参数配置信息和算法配置信息;
    可调参数集合;
    可调参数配置信息;
    算法配置信息;
    其中,所述可调参数集合用于辅助确定针对AI功能的目标策略;所述可调参数集合包括:N个可调参数项及每个可调参数项的至少一个取值,N为正整数;
    所述算法配置信息用于指示目标算法的配置参数,所述算法配置信息用于辅助确定针对AI功能的目标策略;
    所述可调参数配置信息用于辅助确定所述可调参数集合;所述可调参数配置信息包括:所述N个可调参数项中部分或全部可调参数项,及所述N个可调参数项中0个或至少一个可调参数项的至少一个取值;
    所述目标策略包括以下至少一项:AI模型部署策略;AI模型去激活策略;AI模型激活策略;AI模型训练策略。
  25. 一种第一设备,其中,包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如权利要求1至12任一项所述的AI模型策略确定方法的步骤。
  26. 一种第二设备,其中,包括处理器和存储器,所述存储器存储可在所述处理器上运 行的程序或指令,所述程序或指令被所述处理器执行时实现如权利要求13至22任一项所述的AI模型策略确定方法的步骤。
  27. 一种可读存储介质,其中,所述可读存储介质上存储程序或指令,所述程序或指令被处理器执行时实现如权利要求1至12任一项所述的AI模型策略确定方法,或者实现如权利要求13至22任一项所述的AI模型策略确定方法的步骤。
PCT/CN2023/128628 2022-11-10 2023-10-31 Ai模型策略确定方法、装置、第一设备及第二设备 WO2024099187A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211406405.6 2022-11-10
CN202211406405.6A CN118019022A (zh) 2022-11-10 2022-11-10 Ai模型策略确定方法、装置、第一设备及第二设备

Publications (1)

Publication Number Publication Date
WO2024099187A1 true WO2024099187A1 (zh) 2024-05-16

Family

ID=90954921

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/128628 WO2024099187A1 (zh) 2022-11-10 2023-10-31 Ai模型策略确定方法、装置、第一设备及第二设备

Country Status (2)

Country Link
CN (1) CN118019022A (zh)
WO (1) WO2024099187A1 (zh)

Also Published As

Publication number Publication date
CN118019022A (zh) 2024-05-10

Similar Documents

Publication Publication Date Title
Yu et al. Computation offloading for mobile edge computing: A deep learning approach
CN109982352B (zh) Tdd-lte网络中的异频切换方法及装置
WO2024099187A1 (zh) Ai模型策略确定方法、装置、第一设备及第二设备
CN107734712A (zh) 一种随机接入方法和设备
CN102857939B (zh) 小区参数的配置方法及装置
CN116846771A (zh) 业务操作方法、装置、终端及可读存储介质
WO2023213288A1 (zh) 模型获取方法及通信设备
US20220007450A1 (en) Dynamic Adjustment Method and Apparatus for PDU Session
WO2023216961A1 (zh) 隐私保护信息处理方法、装置及通信设备
WO2023207980A1 (zh) 模型信息获取方法、发送方法、装置、节点和储存介质
CN117062098A (zh) 模型获取方法及通信设备
WO2023185818A1 (zh) 设备确定方法、装置及通信设备
WO2023169425A1 (zh) 通信网络中的数据处理方法及网络侧设备
WO2024078400A1 (zh) 模型请求方法、装置、通信设备及可读存储介质
WO2023179709A1 (zh) 信息处理方法、装置、通信设备及可读存储介质
WO2024017244A1 (zh) 信息发送方法、信息接收方法、装置及相关设备
WO2023169402A1 (zh) 模型的准确度确定方法、装置及网络侧设备
WO2023169378A1 (zh) 模型的准确度确定方法、装置及网络侧设备
WO2023179558A1 (zh) 数据采集方法及通信设备
WO2023186091A1 (zh) 样本确定方法、装置及设备
US20230114869A1 (en) Method and article of manufacture for providing a mobility management with optimized paging
WO2024061254A1 (zh) 数据获取方法、装置、系统和设备
US20230067970A1 (en) Electronic device for providing information for reinforcement learning and method for operating thereof
CN106470493A (zh) 中继节点的处理方法、基站和中继节点
CN117440449A (zh) Ai模型信息传输方法、装置及设备