TWI806496B - Construction systems for recurrent bayesian networks, construction methods for recurrent bayesian networks, computer readable recording medium with stored program, non-transitory computer program products, and wireless network control systems - Google Patents

Abstract

The present invention provides a construction system for recurrent Bayesian networks, a construction method for recurrent Bayesian networks, a computer readable recording medium with stored program, a non-transitory computer program product, and a wireless network control system. A processor performs: creating an initial population and setting the initial population to be a current population; generating a recurrent Bayesian network for each compositional pattern-producing network in the current population to obtain a set of recurrent Bayesian networks corresponding to the current population; evolving the current population using an evolutionary algorithm and a fitness function to obtain a next population; determining, based on the fitness function and the set of recurrent Bayesian networks corresponding to the current population whether a termination condition is satisfied. If the termination condition is not satisfied, repeating the above steps, and if the termination condition is satisfied, selecting a solution network of the current population as a task network based on the fitness function.

Description

Recursive Bayesian network construction system, recursive Bayesian network construction method, computer-readable recording medium, non-transitory computer program product, and wireless network control system

The invention relates to the genetic algorithm technology applied to the recursive Bayesian network, especially the recursive Bayesian network applicable to the wireless network control system.

Self-organizing network (SON) refers to the minimization of human intervention in mobile network automation and cellular/wireless network management. This concept was introduced by 3GPP in Release 8. The main goals of self-organizing networks can be roughly divided into three points. First, to introduce intelligence and autonomous adaptability into cellular networks; second, to reduce capital and operating expenses; and third, to improve network performance in terms of network capacity, coverage, provided services and experience. However, many operations of existing self-organizing network solutions on the market are still done manually (for example, network faults are usually directly repaired); and many operations are still done manually (for example, network faults are usually directly repaired by engineers ) as well as some open challenges that remain unresolved, such as the coordination of the self-organizing network functions, or properly addressing the trade-offs between centralized and decentralized self-organizing network implementations.

Recurrent Bayesian networks have the advantage of being able to use prior knowledge to model the causal relationship between problem features as dependencies between random variables, and to deal with the inherent uncertainty in the problem task. However, training a large recurrent Bayesian network usually requires a large number of training samples to learn useful things from the data. Moreover, the structure of the recurrent Bayesian network to be trained is predetermined, which may lead to low utilization efficiency of computing resources.

In view of this, some embodiments of the present invention provide a recursive Bayesian network construction system, a recursive Bayesian network construction method, a computer-readable recording medium, a non-transitory computer program product, and a wireless network control system, to Improve existing technical problems.

An embodiment of the present invention provides a system for constructing a recurrent Bayesian network, which includes at least one processor. The aforementioned at least one processor is configured to perform the following steps to generate a task network: establish an initial group, wherein the initial group includes a plurality of combination patterns to generate a network, and set the initial group as the current group; for each combination pattern in the current group Generate a network and establish a corresponding recursive Bayesian network to obtain a recursive Bayesian network set corresponding to the current population; use evolutionary algorithms and fitness functions to evolve the current population to obtain the next population, and set the next population to The current group; judge whether the termination condition is satisfied according to the fitness function and the recursive Bayesian network set corresponding to the current group; function, select the solution network in the current group as the task network.

An embodiment of the present invention provides a method for constructing a recursive Bayesian network, which is executed by a processor. A recurrent Bayesian network construction method is used to generate task networks. The recursive Bayesian network construction method includes: establishing an initial group, wherein the initial group contains a plurality of combination mode generation networks, and setting the initial group as the current group; for each combination mode generation network in the current group, establishing a corresponding Recursive Bayesian network to obtain a recursive Bayesian network set corresponding to the current population; use the evolutionary algorithm and fitness function to evolve the current population to obtain the next population, and set the next population as the current population; according to the fitness function And the recursive Bayesian network set corresponding to the current group determines whether the termination condition is satisfied; and in response to the termination condition not being satisfied, repeat the above steps, in response to the termination condition being satisfied, according to the fitness function, select one of the current group The solution network serves as the task network.

Some embodiments of the present invention provide a computer-readable medium storing a program and a non-transitory computer program product. After the processor loads the program and executes it, the aforementioned recursive Bayesian network construction method can be completed.

An embodiment of the invention provides a wireless network control system. The wireless network control system applies the task network generated by the aforementioned recursive Bayesian network construction system. The wireless network control system is coupled to a radio access network (Radio Access Network, RAN) and includes an optimizer and a control unit. The optimizer is configured to receive the network status and target policy of the wireless access network, and output the optimal configuration to the wireless access network for configuring and optimizing the wireless access network. The control unit includes the aforementioned task network. Wherein, the optimizer inputs a plurality of configuration feature vectors of each solution to the control unit based on the plurality of solutions in the solution space. The task network evaluates at least one Key Performance Indicator (KPI) of each configuration feature vector to obtain the KPI value of each configuration feature vector. The optimizer chooses an optimal solution in the solution space based on the KPI values for each configuration eigenvector. The optimizer outputs the best configuration to the radio access network based on the best solution.

Based on the above, some embodiments of the present invention provide a recursive Bayesian network construction system, a construction method, a computer-readable recording medium, and a non-transitory computer program product. The group formed by the network can be generated by an evolutionary combination model. Efficiently construct a suitable recurrent Bayesian network. Some embodiments of the present invention provide a wireless network control system, which can automatically complete a self-organizing network solution by applying the task network generated by the aforementioned recursive Bayesian network construction system.

The aforementioned and other technical contents, features and effects of the present invention will be clearly presented in the following detailed description of the embodiments with reference to the drawings. The ratio or size of each element in the drawings is expressed in an exaggerated, omitted or approximate manner, for the understanding and reading of those familiar with the art, and is not used to limit the conditions for the implementation of the present invention, so there is no technical limitation Substantive meaning, any modification of structure, change of proportional relationship or adjustment of size shall still fall within the scope covered by the technical content disclosed in the present invention without affecting the functions and objectives of the present invention. Inside. The same reference numbers will be used throughout the drawings to refer to the same or similar elements. The term "coupled" or "connected" mentioned in the following embodiments may refer to any direct or indirect, wired or wireless connection means.

FIG. 1 is a block diagram of a recurrent Bayesian network construction system according to an embodiment of the present invention. Please refer to FIG. 1 , the recurrent Bayesian network construction system 100 includes a processor 101 and a memory 102 . The memory 102 is configured to store intermediate results that need to be temporarily stored during the calculation process. It should be noted that although FIG. 1 only shows a single processor 101 , the recurrent Bayesian network construction system 100 can also be configured with multiple processors to speed up the calculation.

The method for constructing a recursive Bayesian network according to the embodiment of the present invention and how the various modules of the recursive Bayesian network construction system 100 cooperate with each other will be described in detail below with reference to the drawings.

FIG. 2 is a schematic diagram of a compositional pattern-producing network (CPPN) according to an embodiment of the present invention. FIG. 3 is a schematic diagram of neural network coding according to an embodiment of the present invention. FIG. 4 is a schematic diagram of a neural network mutation process according to an embodiment of the present invention. FIG. 5 is a schematic diagram of a neural network crossover process according to an embodiment of the present invention. FIG. 9 is a flowchart of a method for constructing a recurrent Bayesian network according to an embodiment of the present invention. Please refer to FIG. 2 to FIG. 5 and FIG. 9 together.

In the embodiment shown in FIG. 9, the combined pattern generation network is a neural network. Taking the combined pattern generating network 205 shown in FIG. 2 as an example, the combined pattern generating network 205 includes an input layer 2051 , a hidden layer 2052 and an output layer 2053 . Wherein, each node of the hidden layer 2052 may adopt different activation functions, such as trigonometric sine, Gaussian and so on. It is worth noting that the combined pattern generating network 205 is only used to illustrate a general combined pattern generating network, and the present invention is not limited to the connection relationship of the combined pattern generating network 205 .

Using the combination mode to generate the network, the connection mode of the neural network can be stored as a four-dimensional hypercube, where each point of the four-dimensional hypercube can be expressed as (x ₁ ,y ₁ ,x ₂ ,y ₂ ), where x ₁ , y ₁ , x ₂ , y ₂ are the coordinate components of the points of the four-dimensional hypercube. Each point encodes a link between two nodes. The combined pattern generation network 205 can be regarded as a four-dimensional function CPPN(1): w=CPPN(x ₁ , y ₁ , x ₂ , y ₂ ) (1). Where w is the output value of the combined mode generation network 205 . As shown in Figure 2, the connection weight between node (x ₁ ,y ₁ ) and node (x ₂ ,y ₂ ) can be obtained by inputting (x ₁ ,y ₁ ,x ₂ ,y ₂ ) into the four-dimensional In the function CPPN(1), the obtained output w is the connection weight between the node (x ₁ , y ₁ ) and the node (x ₂ , y ₂ ). Taking the example shown in FIG. 2 as an illustration, since the output of the four-dimensional function CPPN(1) corresponding to (-1,1,1,1) is w ₁ , the node 201 (encoded as (-1,1)) to the node 202 (encoded as (1,1)) is w ₁ ; since the output of the four-dimensional function CPPN(1) corresponding to (0,-1,1,-1) is w ₂ , node 203 (encoded as ( 0,-1)) to the node 204 (encoded as (1,-1)) is w ₂ . It is worth noting that, in this embodiment, a minimum threshold w _min is given first, and when the output w is smaller than the minimum threshold w _min , it means that the two nodes are not connected. In the manner described above, the combined pattern generation network 205 and the corresponding four-dimensional function CPPN(1) can establish a corresponding neural network topology.

Please refer to FIG. 3 , a general neural network can represent a neural network by numbering its nodes and links so that the neural network can perform evolutionary operations. Taking the example shown in FIG. 3 as an example, by numbering the nodes of the combination pattern generation network 205 from 1 to 7, a numbered combination pattern generation network 303 can be obtained. The numbered combined pattern generation network 303 can be expressed as a node gene 301 and a link gene 302 . The "input" recorded in the node gene 301 indicates that this node is an input node; the "output" recorded in the node gene 301 indicates that this node is an output node; the "hidden" recorded in the node gene 301 indicates that this node is a hidden node. The "in" recorded in the connection gene 302 represents the starting node of the connection; the "out" recorded in the connection gene 302 represents the end node of the connection; the "weight" recorded in the connection gene 302 represents the weight value of the connection; the "innovation" recorded in the connection gene 302 "(innovation) indicates the innovation value of the link. The connected innovation value is used to match the genome when performing the exchange process in the neural network evolution operation, and the use of the connected innovation value will be further introduced in the following embodiments.

After the neural network is represented as node genes 301 and link genes 302 as shown in FIG. 3 , the neural network can further perform evolutionary calculations. As shown in FIG. 4 , the neural network 402 has link genes 401 . By randomly increasing the linking gene elements of the linking gene 401, for example, in the linking gene 403, the "in" is 2, the "out" is 5, the "weight" is 0.1, and the "innovation" is 6 linking gene elements. The link gene 403 corresponds to the neural network 404 . The process of the neural network 402 evolving to the neural network 404 is called the neural network mutation process.

It is worth noting that the aforementioned random addition of elements connecting the gene 401 can randomly generate appropriate "in", "out", "weight" and "innovation" to achieve.

Please refer to FIG. 5 , the neural network 503 has a link gene 501 , and the link gene 501 has link gene elements 5011 , 5012 , 5013 and 5014 . The neural network 504 has a link gene 502 , and the link gene 502 has link gene elements 5021 , 5022 , 5023 , 5024 and 5025 . To perform the exchange process in the neural network evolution operation on the neural network 503 and the neural network 504 to generate a new neural network, firstly, the linking gene elements of the linking genes of the two are matched based on the "innovation" number. Using the example shown in FIG. 5 to illustrate, the link gene element 5011 matches the link gene element 5021 , the link gene element 5012 matches the link gene element 5022 , and the link gene element 5013 matches the link gene element 5023 . The link gene element 5014 does not match the link gene element of the link gene 502 , and the link gene element 5024 and the link gene element 5025 do not match the link gene element of the link gene 501 . For the aforementioned matching gene elements, randomly select one of them and put it into the connection gene of the new neural network.

Taking the foregoing example as an example, the linking gene element 5011 matches the linking gene element 5021 , so one of the linking gene element 5011 and the linking gene element 5021 is randomly selected and put into the linking gene 505 of the neural network 506 . The link gene element 5012 matches the link gene element 5022, so one of the link gene element 5012 and the link gene element 5022 is randomly selected and put into the link gene 505 of the neural network 506, and so on. For unmatched link gene elements, such as link gene elements 5014 , 5024 and 5025 , they are unconditionally put into the link gene 505 of the neural network 506 . Finally, a link gene 505 and a corresponding neural network 506 can be obtained, wherein the link gene 505 includes link gene elements 5051 , 5052 , 5053 , 5054 , 5055 and 5056 . The neural network 503 and the neural network 504 are called the parents of the neural network 506 , and the neural network 506 is called the offspring of the neural network 503 and the neural network 504 .

The algorithm for evolving the topology and connection weights of the neural network shown in FIGS. 3 to 5 is called the NeuroEvolution of Augmenting Topologies method.

It is worth noting that one of the gene elements 5012 and 5022 can be randomly connected to one of the gene elements 5012 and 5022 to achieve the aforementioned random selection of the gene element 5012 and the process of linking one of the genetic elements 5022.

In step S901 of FIG. 9 , the processor 101 establishes an initial group, and sets the initial group as the current group. Wherein, the initial group includes a plurality of combined pattern generating networks randomly generated by the processor 101 (for example, the combined pattern generating network 205 shown in FIG. 2 ), and the number of the initial group is a preset group number M, M is a positive integer. The processor 101 also sets the aforementioned initial group as the current group. The processor 101 also receives a fitness function from the outside to evolve the current population. It is worth noting that the processor 101 can use the function of simulating random provided by general software (such as the functions of the random module in Python) to randomly generate each value in the node gene and the link gene (as shown in FIG. 3 shown) to generate a network in a random combination mode.

6-1 and 6-2 are schematic diagrams of a Bayesian network according to an embodiment of the present invention. Before describing step S902, please refer to FIG. 6-1 and FIG. 6-2. The Bayesian network is a probabilistic graphical model. A set of random variables (random variables) {X ₁ ,X ₂ ,...,X _n } and Its n groups of conditional probability distributions (conditional probability distributions, CPDs) properties. For example, a Bayesian network can be used to represent the probability relationship between a disease and its related symptoms. Given a certain symptom, the Bayesian network can be used to calculate the probability of various possible diseases. In general, the nodes of a Bayesian network represent random variables, which can be observable variables, latent variables, unknown parameters, etc. Arrows connecting two nodes indicate whether the two random variables are causal or unconditionally independent. If there is no arrow connecting two nodes, the random variables are said to be conditionally independent from each other. If two nodes are connected by a single arrow, it means that one node is a parent node and the other node is a child node, and the child node will generate a conditional probability value for the parent node.

Figure 6-1 shows a Bayesian network, including node 601 (represented by random variable _X1 ), node 602 (represented by random variable _X2 ), node 603 (represented by random variable _X3 ) and node 604 ( represented by the random variable _X4 ). There is a connecting arrow between the two nodes of node 601 (represented by random variable X ₁ ) and node 602 (represented by random variable X ₂ ), which means that the two random variables X ₁ and X ₂ are causal or unconditionally independent . And, node 601 (represented by random variable X ₁ ) is a parent node, another node 602 (represented by random variable X ₂ ) is a child node, node 601 (represented by random variable X ₁ ) is related to node 602 (represented by random variable X 1 ) X ₂ ) generates a conditional probability P(X ₂ |X ₁ ), node 601 (represented by random variable X ₁ ) and node 603 (represented by random variable X ₃ ) generate a conditional probability P(X ₃ |X ₁ ) , node 604 (represented by random variable X ₄ ), node 602 (represented by random variable X ₂ ) and node 603 (represented by random variable X ₃ ) generate a conditional probability P(X ₄ |X _{2 ,} X ₃ ). P(X ₁ ) represents the probability of node 601 (represented by random variable X ₁ ).

In an embodiment of the present invention, the properties of the nodes in the Bayesian network are discrete. In this embodiment, the conditional probability can be represented by a conditional probability table (conditional probability table, CPT). The sum of the probabilities in any column of the conditional probability table must be 1. Taking the Bayesian network shown in Figure 6-2 as an example, the node 605 represents the random variable C of cloudy, the node 606 represents the random variable S of watering, the node 607 represents the random variable R of rain, and the node 608 represents the wet grass The random variable W of , where the values of the random variables C, S, R, and W are all set {T, F}, T, F are the elements of the aforementioned set, T represents occurrence, and F represents non-occurrence. P(C) is represented by the probability table 609 , P(S|C) is represented by the conditional probability table 610 , P(R|C) is represented by the conditional probability table 611 , and P(W|S,R) is represented by the conditional probability table 612 .

FIG. 7 is a schematic diagram of a recurrent Bayesian network (Recurrent Neural Networks) according to an embodiment of the present invention. As shown in FIG. 7 , the recurrent Bayesian network 700 includes an input layer 701 , a hidden layer 702 and an output layer 703 . The hidden layer 702 includes a Bayesian network 7021 and a recursive layer 7022. The recursive layer 7022 is used to record the state of the Bayesian network, and add the state recorded at the current moment to the input at the next moment middle. The lower part of FIG. 7 shows the operation of the recurrent Bayesian network 700 for time expansion, wherein the output layer 706 at time t-1, the output layer 709 at time t, and the output layer 712 at time t+1 correspond to the output layer 703 ; Input layer 704 at time t-1, input layer 707 at time t and input layer 710 at time t+1 correspond to input layer 701; hidden layer 705 at time t-1, hidden layer 708 at time t and time t The hidden layer 711 of +1 corresponds to the hidden layer 702; the Bayesian network 7051 and the recurrent layer 7052 of the hidden layer 705 at time t-1, the Bayesian network 7081 and the recurrent layer 7082 of the hidden layer 708 at time t and the time t +1 The Bayesian network 7111 and the recurrent layer 7112 of the hidden layer 711 correspond to the Bayesian network 7021 and the recurrent layer 7022 of the hidden layer 702 . At time t−1, the hidden layer 705 stores the node state M1 in the recurrent layer 7052 . At time t, node state M1 is input into hidden layer 708 . Similarly, at time t, the hidden layer 705 stores the node state M2 in the recurrent layer 7082 . At time t+1, the node state M2 is input into the hidden layer 711 . Similarly, at time t+1, the hidden layer 711 stores the node state M3 in the recurrent layer 7112, and so on.

FIG. 8 is a schematic diagram illustrating the construction of a recurrent Bayesian network according to an embodiment of the present invention. Please refer to FIG. 8 and FIG. 9 at the same time. In step S902, the processor 101 generates a network for each combined pattern in the current group, and establishes a corresponding recurrent Bayesian network. The following uses FIG. 8 as an example to illustrate the foregoing method for establishing a corresponding recurrent Bayesian network. For the combination pattern generation network 805, for all time t, the combination pattern generation network 805 takes all possible two nodes of the Bayesian network 800 at time t as input (as shown in FIG. 8 The node 801 (represented by the random variable X ₁ ) and the node 802 (represented by the random variable X ₂ )) are shown, and the conditional probability P(X ₂ |X ₁ ) is obtained. It is worth noting that if the two input nodes have no causal relationship or are conditionally independent, the output of the combined pattern generation network 805 is 0. In this way, the Bayer network 800 of the recurrent Bayesian network at time t can be established via the combined pattern generation network 805, wherein the Bayesian network 800 of the recurrent Bayesian network at time t includes nodes 801 (by random Variable X ₁ represents), node 802 (represented by random variable X ₂ ), node 803 (represented by random variable X ₃ ) and node 804 (represented by random variable X ₄ ); and, node 802 (represented by random variable X ₂ ) to the node 801 (represented by the random variable X ₁ ) of its parent node is P(X ₂ |X ₁ ), and the node 803 (represented by the random variable X ₃ ) to the node 801 of its parent node (represented by The conditional probability of the random variable X ₁ ) is P(X ₃ |X ₁ ), the node 804 (represented by the random variable X ₄ ) and the node 802 (represented by the random variable X ₂ ) of its parent node and the node 803 (represented by the random variable X 2 ) are P(X 3 |X 1 ). The conditional probability of random variable X ₃ ) is P(X ₄ |X ₂ ,X ₃ ). P(X ₁ ) represents the probability of node 801 (represented by random variable X ₁ ).

Through the above establishment method, the processor 101 generates a network for each combined pattern in the current group, and can establish a corresponding recurrent Bayesian network. Therefore, the final processor 101 can obtain the recurrent Bayesian network set corresponding to the current group.

In step S903, the processor 101 uses an evolutionary algorithm and a fitness function to evolve the current group to obtain a next group, and sets the next group as the current group. In an embodiment of the present invention, the processor 101 uses the aforementioned fitness function to calculate the fitness value of each combination mode generation network in the current group, and uses the fitness value of each combination mode generation network in the current group to compare the current Each combined pattern in the population generates a network for sorting. The processor 101 uses the sorting result to select N combined patterns in the current group with higher fitness to generate a network, where N is a positive integer and N is smaller than the aforementioned number M of groups. The processor 101 then utilizes the mutation process and the exchange process in the neural network evolution operation shown in Fig. 4 and Fig. 5 to operate on the selected N combination pattern generation networks, so that the aforementioned N combination pattern generation networks The path is extended to the number of populations M to obtain the next population. The processor 101 then sets the next group as the current group to proceed to the next step.

In an embodiment of the present invention, the processor 101 uses the mutation process and the exchange process in the neural network evolution operation shown in Figure 4 and Figure 5 to operate the selected N combination mode generation network It includes that the processor 101 randomly executes a mutation process and an exchange process on the aforementioned N combined pattern generation networks according to an exchange rate. The aforementioned switching ratio is, for example, 0.8. Then, the processor 101 has an 80% probability of performing the exchange process on the aforementioned N combined pattern generating networks, and has an 80% probability of performing a mutation process on the aforementioned N combined pattern generating networks.

It is worth noting that, as long as the neuroevolution algorithm for enhancing the topology of the neural network and the connection weights shown in Figures 3 to 5 are used to evolve the aforementioned current population, the present invention It is not limited to the foregoing embodiments.

In step S904, the processor 101 determines whether the termination condition is satisfied according to the fitness function and the recurrent Bayesian network set corresponding to the current group. In this embodiment, the aforementioned termination condition is that the number of times of repeating steps S902 and S903 has reached a preset value, or according to the aforementioned fitness function, the current cluster already includes a candidate network so that the fitness value of the candidate network is greater than one Default fitness value. If the termination condition is not satisfied, the processor 101 executes step S902 and step S903 again. If the termination condition is met, the processor 101 executes step S905.

In step S905, since the current group contains at least one candidate network, the fitness value of the candidate network is greater than a preset fitness value. The processor 101 then selects a solution network with a fitness value greater than the preset fitness value in the current group as the task network according to the fitness function.

FIG. 10 is a flowchart of a method for constructing a recurrent Bayesian network according to an embodiment of the present invention. In the embodiment shown in FIG. 10 , the properties of the nodes in the Bayesian network are discrete, and the aforementioned step S902 includes step S1001 and step S1002 . In step S1001, the processor 101 selects the current combination mode in the current group to generate a network. In step S1002, the processor 101 uses the current combination mode to generate the output of the network about the child node in the corresponding recursive Bayesian network and the parent node corresponding to the child node, and establishes the corresponding Conditional probability tables of parent nodes to build corresponding recurrent Bayesian networks.

FIG. 11 is a flowchart of a method for constructing a recurrent Bayesian network according to an embodiment of the present invention. In the embodiment shown in FIG. 11 , the aforementioned step S901 further includes step S1101 . In step S1101, the processor 101 randomly generates an initial group according to a parameter. In this embodiment, the aforementioned parameter is a value generated by the processor 101 based on the time when the step S901 is started, and the processor 101 initializes a general software simulation random function based on this value (for example, the random( of the random module in Python) ) function) to randomly generate the initial population. The processor 101 resets the initial group as the current group.

FIG. 12 is a structural block diagram of an electronic device according to an embodiment of the present specification. As shown in FIG. 12 , at the hardware level, the electronic device 1200 includes a processor 1201 , an internal memory 1202 and a non-volatile memory 1203 . The internal memory 1202 is, for example, random access memory (Random-Access Memory, RAM). The non-volatile memory (non-volatile memory) is, for example, at least one disk memory or the like. Of course, the electronic device 1200 may also include hardware required by other functions.

The internal memory 1202 and the non-volatile memory 1203 are used to store programs. The programs may include program codes, and the program codes include computer operation instructions. Internal memory 1202 and non-volatile memory 1203 provide instructions and data to processor 1201 . The processor 1201 reads the corresponding computer program from the non-volatile memory 1203 into the internal memory 1202 and then runs it. The processor 1201 is specifically configured to execute the steps described in FIG. 9 to FIG. 13 .

The processor 1201 may be an integrated circuit chip with signal processing capabilities. During implementation, the methods and steps disclosed in the above-mentioned embodiments may be implemented through hardware integrated logic circuits in the processor 801 or instructions in the form of software. The processor 1201 may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), field programmable A gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices can implement or execute the methods and steps disclosed in the foregoing embodiments.

The embodiment of this specification also provides a computer-readable storage medium. The computer-readable storage medium stores at least one instruction. When the at least one instruction is executed by the processor 1201 of the electronic device 1200, the processor 1201 of the electronic device 1200 can execute The various methods and steps disclosed in the foregoing embodiments.

Examples of storage media for computers include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM) , read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other internal memory technologies, compact disc read-only memory (CD-ROM), digital multi A compact disc (DVD) or other optical storage, magnetic tape cartridge, magnetic tape storage or other magnetic storage device or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media excludes transitory media, such as modulated data signals and carrier waves.

FIG. 13 is a block diagram of a wireless network control system using the task network generated by the aforementioned recursive Bayesian network construction system. Please refer to FIG. 13 , the wireless network control system 1300 is externally coupled to a radio access network (Radio Access Network, RAN) 1301 . The wireless network control system 1300 includes the recurrent Bayesian network construction system 100 , an optimizer 1302 and a control unit 1303 . The optimizer 1302 is configured to receive the network status of the RAN 1301 and receive a target policy from the outside. The optimizer 1302 is configured to output the optimal configuration to the RAN 1301 to configure and optimize the RAN 1301 . The control unit 1303 is configured to receive and store the task network 1304 from the aforementioned recursive Bayesian network construction system 100 . The recurrent Bayesian network construction system 100 generates a task network 1304 by using pre-collected data related to the RAN 1301 and a fitness function related to the RAN 1301 . The task network 1304 supports the second level (intervention) and the third level (counterfactual) of the causal reasoning model. Thus, the task network 1304 can be used to answer the question "given a time series of RAN configuration eigenvectors, what KPI will the RAN 1301 achieve?".

Wherein the optimizer 1302 inputs multiple configuration feature vectors of each solution to the control unit 1303 based on multiple solutions in the solution space. The task network 1304 then evaluates the input key performance indicators (Key Performance Indicators, KPI) of each configuration feature vector to obtain the KPI value of each configuration feature vector. The optimizer 1302 selects the best solution in the solution space based on the KPI value selection for each configured feature vector. The optimizer 1302 outputs the optimal configuration to the RAN 1301 based on the optimal solution.

In some embodiments of the present invention, the components of each configuration eigenvector include a topological structure of the RAN, a routing table, a traffic intensity matrix, and a power intensity matrix.

In some embodiments of the present invention, the aforementioned key performance indicators are selected from one or a combination of delay (delay) indicators, jitter (jitter) indicators and loss (lose) indicators.

In some embodiments of the present invention, the radio access network 1301 is a 5G radio access network.

Based on the above, some embodiments of the present invention provide a recursive Bayesian network construction system, a construction method, a computer-readable recording medium, and a non-transitory computer program product. The group formed by the network can be generated by an evolutionary combination model. Efficiently construct a suitable recurrent Bayesian network.

Some embodiments of the present invention provide a wireless network control system, which can automatically complete a self-organizing network solution by applying the task network generated by the aforementioned recursive Bayesian network construction system.

Although the technical content of the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any modification and modification made by those skilled in the art without departing from the spirit of the present invention should be covered by the present invention. Therefore, the scope of protection of the present invention should be defined by the scope of the appended patent application.

100: Recurrent Bayesian network construction system 101, 1201: Processor 102: Memory 205, 303, 805: Combined pattern generation network 2051, 701, 704, 707, 710: Input layer 2052, 702, 705, 708 , 711: hidden layer 2053, 703, 706, 709, 712: output layer w, w1, w2: output 201, 202, 203, 204, 801, 802, 803, 804: node 301: node gene 302, 401, 403 . 5056: connecting gene elements 601~608: node 609: probability table 610~612: conditional probability table 700: recursive Bayesian network 7021, 7051, 7081, 7111, 800: Bayesian network 7022, 7052, 7082, 7112 : Recursive layer 1200: Electronic equipment 1202: Internal memory 1203: Non-volatile memory 1300: Wireless network control system 1301: Wireless access network 1302: Optimizer 1303: Control unit 1304: Task network P(. ): probability P(X ₁ ): probability of random variable X ₁ P(X ₂ |X ₁ ), P(X ₃ |X ₁ ), P(X ₄ |X ₂ ,X ₃ ): conditional probability t: time X ₁ , X ₂ , X ₃ , X ₄ , C, S, R, W: random variables T, F: elements of a collection M1, M2, M3: node states S901~S905, S1001~S1002, S1101: steps

FIG. 1 is a block diagram of a recurrent Bayesian network construction system according to an embodiment of the present invention. FIG. 2 is a schematic diagram of a combined mode generation network according to an embodiment of the present invention. FIG. 3 is a schematic diagram of neural network coding according to an embodiment of the present invention. FIG. 4 is a schematic diagram of a neural network mutation process according to an embodiment of the present invention. FIG. 5 is a schematic diagram of a neural network switching process according to an embodiment of the present invention. FIG. 6-1 is a schematic diagram of a Bayesian network according to an embodiment of the present invention. FIG. 6-2 is a schematic diagram of a Bayesian network according to an embodiment of the present invention. FIG. 7 is a schematic diagram of a recurrent Bayesian network according to an embodiment of the present invention. FIG. 8 is a schematic diagram illustrating the construction of a recurrent Bayesian network according to an embodiment of the present invention. FIG. 9 is a flowchart of a method for constructing a recurrent Bayesian network according to an embodiment of the present invention. FIG. 10 is a flowchart of a method for constructing a recurrent Bayesian network according to an embodiment of the present invention. FIG. 11 is a flowchart of a method for constructing a recurrent Bayesian network according to an embodiment of the present invention. FIG. 12 is a structural block diagram of an electronic device according to an embodiment of the present specification. Fig. 13 is a block diagram of a wireless network control system for a task network generated by applying a recursive Bayesian network construction system.

S901~S905: steps

Claims (19)

A recurrent Bayesian network construction system includes a processor, wherein the processor is configured to perform the following steps to generate a task network: (a) setting a current group, wherein the current group includes a plurality of combination mode generation networks; (b) generating a network for each combination pattern in the current group, and establishing a corresponding recurrent Bayesian network to obtain a set of recurrent Bayesian networks corresponding to the current group; (c) using an evolutionary algorithm and a fitness function to evolve the current group to obtain the next group, and the next group is set as the current group; (d) judging whether a termination condition is satisfied according to the fitness function and the recursive Bayesian network set corresponding to the current population; and (e) in response to the termination condition not being met, repeating steps (b), (c) and (d), in response to the termination condition being satisfied, based on the fitness function, select one of the current population solution network road as the task network. The recurrent Bayesian network construction system as described in claim 1, wherein step (b) includes: selecting a current combination pattern in the current population to generate the network; and Using the current combination mode to generate an output corresponding to a child node in the corresponding recurrent Bayesian network and a parent node corresponding to the child node, to establish the child in the corresponding recurrent Bayesian network A conditional probability table of the parent node to which the node corresponds. The recurrent Bayesian network construction system as described in Claim 2, wherein the evolutionary algorithm is a neuroevolutionary algorithm of enhanced topology. The recursive Bayesian network construction system as described in claim 1, wherein the termination condition is that the number of repeated steps (b), (c) and (d) has reached a preset value or according to the fitness function, The current group includes a candidate network such that a fitness value of the candidate network is greater than a predetermined fitness value. The recursive Bayesian network construction system as described in Claim 1, wherein step (a) includes: randomly generating an initial group according to a parameter; and setting the initial group as the current group, wherein the initial group includes the Composite mode generates networks. A method for constructing a recursive Bayesian network is executed by a processor to generate a task network. The method for constructing a recursive Bayesian network includes: (a) setting a current group, wherein the current group includes a plurality of combination mode generation networks; (b) generating a network for each combination pattern in the current group, and establishing a corresponding recurrent Bayesian network to obtain a set of recurrent Bayesian networks corresponding to the current group; (c) using an evolutionary algorithm and a fitness function to evolve the current group to obtain the next group, and the next group is set as the current group; (d) judging whether a termination condition is satisfied according to the fitness function and the recursive Bayesian network set corresponding to the current population; and (e) in response to the termination condition not being met, repeating steps (b), (c) and (d), in response to the termination condition being satisfied, based on the fitness function, select one of the current population solution network road as the task network. The recursive Bayesian network construction method as described in claim item 6, wherein step (b) includes: selecting a current combination pattern in the current population to generate the network; and Using the current combination mode to generate an output corresponding to a child node in the corresponding recurrent Bayesian network and a parent node corresponding to the child node, to establish the child in the corresponding recurrent Bayesian network A conditional probability table of the parent node to which the node corresponds. The recurrent Bayesian network construction method as described in claim 7, wherein the evolutionary algorithm is a topology-enhanced neuroevolutionary algorithm. The recursive Bayesian network construction method as described in claim item 6, wherein the termination condition is that the number of times of repeating steps (b), (c) and (d) has reached a preset value or according to the fitness function, The current group includes a candidate network such that a fitness value of the candidate network is greater than a predetermined fitness value. The recursive Bayesian network construction method as described in claim 6, wherein step (a) includes: randomly generating an initial group according to a parameter; and setting the initial group as the current group, wherein the initial group includes the Composite mode generates networks. A computer-readable recording medium with a built-in program. When a processing unit loads and executes the built-in program, the method described in any one of claims 6 to 10 is completed. A non-transitory computer program product, which stores at least one instruction, and when the at least one instruction is executed by a processor, causes the processor to perform the method described in any one of claims 6-10. A wireless network control system, the wireless network control system is coupled to a wireless access network and includes: A recursive Bayesian network construction system includes a processor, wherein the processor is configured to perform the following steps to generate a task network: (a) setting a current group, wherein the current group includes a plurality of combination mode generation networks; (b) generating a network for each combination pattern in the current group, and establishing a corresponding recurrent Bayesian network to obtain a set of recurrent Bayesian networks corresponding to the current group; (c) using an evolutionary algorithm and a fitness function to evolve the current group to obtain the next group, and the next group is set as the current group; (d) judging whether a termination condition is satisfied according to the fitness function and the recursive Bayesian network set corresponding to the current population; and (e) in response to the termination condition not being met, repeating steps (b), (c) and (d), in response to the termination condition being satisfied, based on the fitness function, select one of the current population solution network road as the task network; an optimizer configured to receive a network state and a target policy of the radio access network, and output an optimal configuration to the radio access network to configure and optimize the radio access network; and a control unit configured to receive and store the task network; wherein the optimizer inputs to the control unit configuration feature vectors for each of the solutions based on a plurality of solutions in a solution space, the The task network evaluates at least one KPI value for each configuration feature vector to obtain at least one KPI value for each configuration feature vector, the optimizer based on the at least one KPI value for each configuration feature vector An optimal solution is selected in the solution space, and the optimal configuration is output to the radio access network based on the optimal solution. The wireless network control system as claimed in claim 13, wherein each component of the configuration feature vector includes a topology, a routing table, a traffic intensity matrix and a power intensity matrix of the wireless access network. The wireless network control system according to claim 13, wherein the at least one key performance indicator is selected from a group consisting of a free delay indicator, a jitter indicator and a loss indicator. The wireless network control system as described in claim 13, wherein step (b) includes: selecting a current combination pattern in the current population to generate the network; and Using the current combination mode to generate an output corresponding to a child node in the corresponding recurrent Bayesian network and a parent node corresponding to the child node, to establish the child in the corresponding recurrent Bayesian network A conditional probability table of the parent node to which the node corresponds. The wireless network control system as claimed in claim 16, wherein the evolutionary algorithm is a topology-enhanced neuroevolutionary algorithm. The wireless network control system as described in claim 13, wherein the termination condition is that the number of times of repeatedly performing steps (b), (c) and (d) has reached a preset value or according to the fitness function, the current group Including a candidate network makes a fitness value of the candidate network greater than a preset fitness value. The wireless network control system as described in claim 13, the recursive Bayesian network construction system as described in claim 1, wherein step (a) includes: randomly generating an initial group according to a parameter; and setting the initial The group is the current group, wherein the initial group includes the combination mode generation network.

Images