CN113469491B - Flexible workshop operation scheduling method based on reinforcement learning and graph neural network - Google Patents
Flexible workshop operation scheduling method based on reinforcement learning and graph neural network Download PDFInfo
- Publication number
- CN113469491B CN113469491B CN202110529888.8A CN202110529888A CN113469491B CN 113469491 B CN113469491 B CN 113469491B CN 202110529888 A CN202110529888 A CN 202110529888A CN 113469491 B CN113469491 B CN 113469491B
- Authority
- CN
- China
- Prior art keywords
- workshop
- job
- neural network
- simulation environment
- scheduling
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/04—Manufacturing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Abstract
The invention provides a flexible workshop operation scheduling method based on reinforcement learning and a graphic neural network, and belongs to the field of industrial scheduling. The method specifically comprises the following steps: 1. initializing a workshop simulation environment according to workshop live conditions; 2. importing a pre-trained strategy function model based on a graph neural network; 3. starting a simulation task in a workshop simulation environment, and returning to idle equipment if the idle equipment is detected in the workshop from the current moment of the workshop; 4. information in the workshop simulation environment is summarized into a flexible workshop operation diagram model; 5. calculating action probability by using a graph neural network, and selecting an action according to the probability; 6. the workshop simulation environment executes the action, if the workshop simulation environment still has an unfinished order, the step 3 is returned, otherwise, the solution is ended; 7. the result of the scheduling is derived from the plant simulation environment. The method provided by the patent has good solving effect and higher calculating efficiency, and can be truly applied to industrial production.
Description
Technical Field
The invention provides a flexible workshop operation scheduling method based on reinforcement learning and a graphic neural network, and belongs to the field of industrial scheduling.
Background
The method solves the workshop operation scheduling problem by using a strategy gradient method, and is a classical reinforcement learning method. In the strategy gradient algorithm, a strategy function takes a state as an input, and probability distribution corresponding to each action on the current state, namely strategy pi, is output. The strategic gradient method uses state value as an objective function J (θ):
where θ represents a parameter of the policy function. According to the bellman equation, state value is the sum of the value of each action on the state times its probability:
thus, the strategy gradient method optimizes the strategy function by making a gradient directional propagation to this objective function:
in reinforcement learning, G is used t The sum of the prize discounts from time t is represented, also called payback:
wherein R is t Representing the prize at time t, and gamma e 0,1]Representing a discount on the prize. The REINFORCE algorithm employed in this work uses the true returns from state up-sampling to optimize the policy gradient.
The patent uses a graph neural network technique to implement a policy function, a graph model describes the relationship between data elements, GNN is a neural network computation on the graph model that uses message passing to obtain information of neighbor nodes, and processes the message by pooling computation. The method for assembling the messages comprises the steps of splicing, summing, averaging and weighting. Unlike common deep networks, the input and output scales of the graph neural network can vary flexibly with the graph model of the problem. The problem concerned with this work, flexible shop job scheduling, is complex equipment and job relationships. The GNN technology combined with the graph model and the neural network enables the method provided by the patent to be more efficient in calculation and have strong expandability.
Disclosure of Invention
The invention aims to: the patent designs a method based on reinforcement learning and a graph neural network, which is used for solving the workshop operation scheduling problem with high efficiency.
The technical scheme is as follows: a flexible workshop operation scheduling method based on reinforcement learning and a graph neural network mainly comprises the following steps:
step 1, initializing a workshop simulation environment according to workshop live conditions;
step 2, importing a pre-trained strategy function model based on a graph neural network;
step 3, starting a simulation task in a workshop simulation environment, and returning to the idle equipment if the idle equipment is detected in the workshop from the current moment of the workshop;
step 4, information in the workshop simulation environment is summarized into a flexible workshop operation diagram model;
step 5, calculating action probability by using the graph neural network, and selecting an action according to the probability;
step 6, the workshop simulation environment executes the action, if the workshop simulation environment still has an unfinished order, the step 3 is returned, otherwise, the solution is ended;
and 7, exporting a scheduling result from the workshop simulation environment.
Still further, in step 1, the shop simulation environment is a simulator of a generic flexible job shop scheduling process. In the simulator, the solution process of flexible job shop scheduling is modeled as a multi-agent Markov decision process, where the device is taken as an agent, the state is a flexible shop job graph model of the shop time, the job is an action, and the cost of scheduling is a negative benefit. The plant simulation environment support used in this patent is defined personalized because of the large differences in the manufacturing process between the different plants. In personalizing definitions, a user may specify particular properties of the individual devices, including the processes that they may perform, the efficiency with which they perform a process, and other particular constraints. Furthermore, the plant simulation environment can be loaded into the actual situation of the plant at the time, including the order to be scheduled, the stock of each material, and the scheduling plan that has been scheduled for non-alterability.
Furthermore, the neural network model imported in the step 2 needs to be trained in advance, so that the network model can adapt to a specific manufacturing environment, and each cost can be reduced to be below a threshold value; the primary pre-training process of the graph neural network comprises two parts, namely iterative sampling and optimizing, wherein the sampling is that in the steps 3-6, and a strategy network is optimized by utilizing a REINFORCE algorithm after sampling; the training is terminated after the cost of the scheduling solution obtained by sampling is continuously 10 times lower than the expected cost of each pre-training, the pre-training of the graph neural network can be repeated for a plurality of times, and an optimal model is selected to store or each model is stored.
Still further, in step 4, the flexible shop operation graph model describes specific attributes such as relationships between operations and devices, and real-time inventory of workshops, where the graph model has three types of entities including devices, processes and operations, the entities are nodes of the graph model, and the relationships between the entities correspond to edges of the graph model.
The plant has a set of equipment and a set of processes, each equipment being capable of completing only a portion of the processes therein. The workshops are required to produce orders, each order has a job set, the jobs have a front-to-back constraint relationship, a tree diagram is finally formed, and all the jobs form a set of workshops job sets. Each job corresponds to a process of the plant, so the apparatus indirectly forms a corresponding set of viable jobs from its own set of sub-processes.
The manner in which different types of entities characterize attributes is different. In the patent, the attribute of the operation entity corresponds to the stock and consumption record of the target product corresponding to the operation; for unified representation and calculation, raw materials appear as pseudo-work entities at leaf nodes of the job dendrogram, also with inventory and consumption records. The original attribute of the process entity is a one-hot vector of length equivalent to the shop process set. The attribute of the equipment comprises a static attribute and a dynamic attribute, wherein the static attribute is a sub-process set of the equipment, and is also represented by a vector with the length equal to that of a workshop process set, and 1 in the vector marks the equipment in the sub-process set of the equipment; the dynamic attribute of the device is a real-time production effect estimate for each job corresponding to the current device.
Further, based on the strategy function of the graph neural network, the graph model is operated for a plurality of times and multiple layers in the flexible workshop. In the first-level computation, the original attribute and the collected process information on the node are firstly transferred from bottom to top on the job tree, then order embedding is carried out based on the whole job tree, and in addition, static and dynamic information of equipment needs to be acquired. The second tier of computation performs message passing based on the device and its operational relationships, aggregates messages on the corresponding operational nodes, and computes the logarithm of the corresponding action probabilities. Finally, the output of all the job nodes is collected, and the strategy is obtained through a softmax function.
The design of the specific calculations of the graph neural network is as follows:
step 1, job embedding is to transfer original attribute and collected process information on nodes from bottom to top on a job tree. Due to the material and product constraints of the job, a job node has multiple sub-nodes that provide its production material. The communication channels contemplated by this patent are unidirectional and ordered. Each node accepts information only from its child nodes and passes information up after the local computation is completed. The process of transferring information completes the bottom-up traversal of the order's job tree. To simplify the computation, the graph model will determine the order of traversal after the first order is obtained, so in iterative interactive solutions the order of traversing the order tree for one job at a time is consistent. Let xi (v) be the set of child nodes of node v, jop υ For the representation of the process corresponding to the job node, the information of the node when the job is embedded into each step of calculation can be represented as:
wherein g (|w) g ) Representing an encoder forThe information calculated at this time is encoded. In the formula f (|w) f ) A decoder, representing one, is used to release information in the compressed representation. In fig. 3, the left side is a job tree, and the right side is an operation performed on a frame line portion on the job tree. Gray job nodes accept messages for a white pseudo job node, i.e., leaf node, and a black child job node, i.e., non-leaf node. In the calculation process on the right, where dark rectangles of the hand-drawn style represent the original information, and black rectangles of the non-hand-drawn style represent the processed information. Wherein, when computing dark gray nodes, black nodes have already been computed. The overall style can be generalized as "encode-decode", with the node feature dimension of the new output being the same as the original data dimension. In the calculation process, two collections of information occur together, one is the average operation Mean and one is the splice operation Concat.
And 2, embedding the order into information on the gathering job tree to calculate the representation of the whole order. This representation treats the order as a high-level node, treats all jobs in the order as child nodes thereof, and ignores leaf nodes as pseudo job nodes. Since the order tree no longer needs to maintain vector dimensions for upper nodes, compression coding is not needed, and the calculation process is similar to node embedding but is simpler. Note that the set of non-leaf nodes for one order o is ζ (o). The order is thus expressed as:
wherein h (|w) h ) Is a decoder.
And 3, embedding the process set characteristics of the collecting equipment into the equipment and expanding the dynamic characteristics of the corresponding feasible operation. The attribute of the equipment on a certain job node v is recorded asThe process set code is mop:
and 4, the second layer of calculation obtains a new representation of the job node by collecting the job information and the equipment information. Unlike the first-tier calculation, the current equipment can produce jobs under consideration in this calculation. These particular job nodes first aggregate messages from the first tier of post-computation job nodes, order nodes, and equipment nodes, spliced into a vector form:
e′ 2υ ＝[e υ ，y o ，m υ ]
then, the logarithm of the motion probability from the current equipment to each job is calculated, and the logarithm of the motion probability from the current equipment to each job is calculated' υ Using a multi-layer neural network N (|w) N ) Solving:
drawings
FIG. 1 is a graphical model relationship of the entirety of a flexible shop operation.
Fig. 2 is a diagram of a network architecture based on a graph model.
FIG. 3 is an example diagram of messaging and computation between job nodes.
FIG. 4 is a process diagram of aggregating order tree information.
Fig. 5a is a gater graph of the solution result using the method proposed in this patent.
Fig. 5b is an enlarged illustration of the lower left detail of the pentagat diagram.
Detailed Description
The invention will be further illustrated with reference to specific examples. It is to be understood that these examples are illustrative of the present invention and are not intended to limit the scope of the present invention. The embodiment of the invention relates to a flexible workshop operation scheduling method based on reinforcement learning and a graphic neural network, and the specific implementation process is divided into two steps of model adaptation and model application.
In a further embodiment, the model adaptation specifically comprises the steps of:
step 1, leading equipment and process of a current factory into a workshop simulation environment, and leading a past order into an order generator in the workshop simulation environment;
step 2, selecting a graph model type and defining a graph neural network model;
and 3, resetting the simulation environment of the workshop. Resetting the track of the intelligent body;
step 4, starting a simulation task in a workshop simulation environment, and returning to the idle equipment if the idle equipment is detected in the workshop from the current moment of the workshop;
step 5, information in the workshop simulation environment is summarized into a flexible workshop operation diagram model, then a diagram neural network is used for calculating the action probability, and an action is selected according to the probability;
step 6, the workshop simulation environment executes the action and returns rewards;
step 7, the agent stores the logarithm of the action probability and rewards to the current track;
step 8, if the incomplete order still exists in the workshop simulation environment, returning to the step 4;
step 9, optimizing a strategy network according to the track and REINFORCE algorithm;
and step 10, calculating the cost, if all the costs of the continuous 10-round scheduling plans are lower than the expected value, storing the engineering prototype and the strategy network model to finish operation, otherwise, returning to the step 3.
Model adaptation requires first defining the prototype of the plant and randomly combining the production tasks of the plant with past orders. The network is gradually optimized during the course of multiple rounds of training. And finally, saving the workshop prototype and the pre-trained strategy network.
In a further embodiment, the model application specifically comprises the steps of:
step 1, preparing a calculation task: initializing a workshop simulation environment according to workshop live conditions, and importing a pre-trained strategy function model based on a graph neural network;
step 2, starting a simulation task in a workshop simulation environment, and returning to idle equipment if the idle equipment is detected in the workshop;
step 3, inducing each information in the workshop into a flexible workshop operation diagram model, and calculating by using a diagram neural network;
step 4, the workshop simulation environment executes the action, and if an unfinished order still exists in the workshop simulation environment, the step 3 is returned;
and 5, completing order scheduling planning, and leading out a scheduling result from the workshop simulation environment.
After the model is applied, a workshop scheduling plan is obtained, and the patent visualizes a certain obtained plan into a Gantt chart form. Gantt chart is one of the most common methods of visualizing a shop scheduling plan, where the vertical axis corresponds to the devices in the shop and the horizontal axis corresponds to time. Each of the rectangles corresponds to a production job, the start and end points of the job production are the start time and the end time of the job production, and the row corresponding to the vertical axis is the job for executing the production. FIG. 5 is an example of the result of operation, and it can be seen that the scheduling plan generated by the present patent prioritizes the completion of shorter production tasks for reduced latency and balances the continuous production and switching process of the process. This patent is designed to face the complex manufacturing scheduling problem, so the example Gantt chart is very large in size, for which purpose FIG. 5b is enlarged in the lower right hand corner of the Gantt chart. The ordinate axis marks the equipment number, and an order and a job number are marked on each job patch.
Claims (1)
1. The workshop operation scheduling method based on reinforcement learning and the graph neural network is characterized by comprising the following steps of:
step 1, initializing a workshop simulation environment according to workshop live conditions, wherein the workshop simulation environment is a simulator of a general flexible job workshop scheduling process, in the simulator, a solution process of flexible workshop job scheduling is modeled as a multi-agent Markov decision process, wherein equipment is used as an agent, the state is a flexible workshop job graph model of workshop time, the job is action, and the scheduling cost is negative income;
step 2, importing a pre-trained strategy function model based on a graph neural network, wherein the imported graph neural network model needs to be trained in advance, so that the network model can adapt to a specific manufacturing environment, and each cost can be reduced to be below a threshold value; the primary pre-training process of the graph neural network comprises two parts, namely iterative sampling and optimizing, wherein the sampling is that the strategy network is optimized by utilizing REINFORCE algorithm after the sampling, namely the steps 3 to 6; the training is terminated after the cost of the scheduling solution obtained by sampling is continuously 10 times lower than the expected cost after each pre-training; the pre-training of the graph neural network can be repeated for a plurality of times, and an optimal model is selected to store or each model is stored;
step 3, starting a simulation task in a workshop simulation environment, and returning to the idle equipment if the idle equipment is detected in the workshop from the current moment of the workshop;
step 4, information in the simulation environment of the workshop is summarized into a flexible workshop operation graph model, the flexible workshop operation graph model describes specific attributes such as relationships among operations and between operations and devices, real-time inventory of the workshop and the like, the graph model is provided with three types of entities including the devices, the processes and the operations, the entities are nodes of the graph model, the relationship among the entities corresponds to a set of child nodes with a side record xi (v) of the graph model as the node v, and jop υ For the representation of the process corresponding to the job node, the information of the node when the job is embedded into each step of calculation is represented as:
step 5, calculating action probability by using the graph neural network, selecting an action according to the probability, and performing multiple and multi-layer operation on the flexible workshop operation graph model based on a strategy function of the graph neural network: in the first-layer calculation, firstly, the original attribute and the collected process information on the node are transmitted from bottom to top on the job tree; then, order embedding is carried out based on the whole job tree, and static and dynamic information of equipment is required to be acquired; the second layer of calculation is based on the information transmission of the equipment and the feasible operation relation, the information is gathered on the corresponding operation node, the logarithmic value of the corresponding strategy pi is calculated, then a certain operation is randomly selected according to the strategy pi, and the operation is carried out on e' υ Using a multi-layer neural network N (|w) N ) Solving:
step 6, the workshop simulation environment executes the action, if the workshop simulation environment still has an unfinished order, the step 3 is returned, otherwise, the solution is ended, the action is the flexible workshop job scheduling plan, each element of the scheduling plan comprises a job, equipment for producing the job, and the start-stop time of the equipment for executing the job;
and 7, exporting a scheduling result from the workshop simulation environment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110529888.8A CN113469491B (en) | 2021-05-14 | 2021-05-14 | Flexible workshop operation scheduling method based on reinforcement learning and graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110529888.8A CN113469491B (en) | 2021-05-14 | 2021-05-14 | Flexible workshop operation scheduling method based on reinforcement learning and graph neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113469491A CN113469491A (en) | 2021-10-01 |
| CN113469491B true CN113469491B (en) | 2023-09-01 |
Family
ID=77870828
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110529888.8A Active CN113469491B (en) | 2021-05-14 | 2021-05-14 | Flexible workshop operation scheduling method based on reinforcement learning and graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113469491B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107767022A (en) * | 2017-09-12 | 2018-03-06 | 重庆邮电大学 | A kind of Dynamic Job-shop Scheduling rule intelligent selecting method of creation data driving |
| WO2020182989A1 (en) * | 2019-03-13 | 2020-09-17 | Deepmind Technologies Limited | Scheduling computation graphs using neural networks |
| CN111756653A (en) * | 2020-06-04 | 2020-10-09 | 北京理工大学 | Multi-coflow scheduling method based on deep reinforcement learning of graph neural network |
| CN112116129A (en) * | 2020-08-24 | 2020-12-22 | 中山大学 | Dynamic path optimization problem solving method based on deep reinforcement learning |
-
2021
- 2021-05-14 CN CN202110529888.8A patent/CN113469491B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107767022A (en) * | 2017-09-12 | 2018-03-06 | 重庆邮电大学 | A kind of Dynamic Job-shop Scheduling rule intelligent selecting method of creation data driving |
| WO2020182989A1 (en) * | 2019-03-13 | 2020-09-17 | Deepmind Technologies Limited | Scheduling computation graphs using neural networks |
| CN111756653A (en) * | 2020-06-04 | 2020-10-09 | 北京理工大学 | Multi-coflow scheduling method based on deep reinforcement learning of graph neural network |
| CN112116129A (en) * | 2020-08-24 | 2020-12-22 | 中山大学 | Dynamic path optimization problem solving method based on deep reinforcement learning |
Non-Patent Citations (1)
| Title |
|---|
| 面向多目标柔性作业车间调度的强化学习NSGA-II算法;尹爱军等;《重庆大学学报》;第45卷(第10期);第113-123页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113469491A (en) | 2021-10-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Learning scheduling policies for multi-robot coordination with graph attention networks | |
| US20210278825A1 (en) | Real-Time Production Scheduling with Deep Reinforcement Learning and Monte Carlo Tree Research | |
| CN111104522A (en) | Regional industry association effect trend prediction method based on knowledge graph | |
| CN112187554B (en) | Operation and maintenance system fault positioning method and system based on Monte Carlo tree search | |
| CN112907150B (en) | Production scheduling method based on genetic algorithm | |
| CN111967972A (en) | Financial product recommendation method and device | |
| CN103324954A (en) | Image classification method based on tree structure and system using same | |
| CN113487165A (en) | Intelligent factory production operation scheduling method and system based on deep reinforcement learning | |
| CN111079989A (en) | Water supply company water supply amount prediction device based on DWT-PCA-LSTM | |
| Oxman | Morphogenesis in the theory and methodology of digital tectonics | |
| CN113780667A (en) | Enterprise operation prediction method and system based on AI artificial intelligence | |
| Estrada et al. | A method of stepwise benchmarking for inefficient DMUs based on the proximity-based target selection | |
| CN115409563A (en) | Multi-factor-influenced agricultural equipment inventory demand prediction method | |
| CN112990584B (en) | Automatic production decision system and method based on deep reinforcement learning | |
| CN113469491B (en) | Flexible workshop operation scheduling method based on reinforcement learning and graph neural network | |
| CN116258356B (en) | Work order dispatching method and device based on WaveNet and related medium | |
| CN116157812A (en) | Automatic generation of system-wide factory-optimized optimization models | |
| CN112184007A (en) | Workshop equipment remote diagnosis method based on digital twins | |
| CN117197643A (en) | Visit display guide suggestion generation method and system based on AIGC technology | |
| Noorul Haq et al. | Effect of forecasting on the multi-echelon distribution inventory supply chain cost using neural network, genetic algorithm and particle swarm optimisation | |
| Perkins III et al. | Stochastic optimization with parametric cost function approximations | |
| Chien et al. | Production-level artificial intelligence applications in semiconductor supply chains | |
| CN114186799A (en) | Enterprise valuation method and system based on heterogeneous graph neural network | |
| CN112712251B (en) | Ship intelligent scheduling method applied to barge management system | |
| CN113435927B (en) | User willingness prediction method, device, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |