Cloud edge end adaptive scheduling method and device based on cognitive decision, and medium

Abstract

The invention discloses a cloud side end adaptive scheduling method and device based on cognitive decision, and a medium, and relates to the technical field of data processing. The method comprises the following steps: carrying out high-fidelity simulation on a cloud side end system by utilizing a digital twin technology, simulating resource use conditions in various service scenes before actual deployment, and finding potential bottlenecks in advance. According to the method, resource configuration can be dynamically adjusted through preset hard constraints in combination with real-time monitoring and simulation deduction, and it is ensured that the service quality requirement can be met even in peak hours. The MADDPG algorithm allows a plurality of agents to carry out collaborative optimization for different scheduling dimensions, a Pareto optimal solution is searched, and the operation cost and the energy consumption are reduced to the maximum extent. By continuously monitoring the deviation between the actual operation data and the simulation predicted value, once a significant difference is found, a model fine tuning mechanism is triggered. In combination with the mechanism, the efficiency of service scheduling under the cloud side-end collaborative architecture is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a cloud-edge-device adaptive scheduling method, device and medium based on cognitive decision-making. Background Technology

[0002] With the popularization of 5G and IoT technologies, cloud-edge-device collaborative architecture has become the core architecture supporting massive terminal devices and high real-time services. Currently, operation and maintenance and computing tasks exhibit characteristics of "diversification (such as video analytics, device control, and data storage), dynamism (business load fluctuates over time), and high concurrency (terminal device scale reaches hundreds of thousands or more)." However, existing scheduling systems still have three core defects, making it difficult to meet the needs of complex scenarios: The first core defect is that most systems (such as Kubernetes native schedulers and traditional load balancers) only trigger scheduling adjustments when resources are overloaded (such as CPU utilization exceeding 90%) or tasks time out (such as service response latency exceeding 500ms), resulting in untimely scheduling mechanism responses and potential business interruptions. For example, during the morning peak hours, when video analytics tasks surge in a smart park, the existing system must wait for the edge node load to exceed the threshold before migrating tasks. During this period, approximately 30% of video stream analysis latency exceeds 200ms, affecting security monitoring effectiveness. The second core defect is that existing technologies often optimize for single indicators, such as "only improving CPU utilization" or "only reducing task latency," leading to overly simplistic scheduling mechanism decisions and poor adaptability to complex scheduling environments. For example, while the automatic scheduling tool for Elastic Compute Service (ECS) can shut down idle instances based on cost thresholds, this can lead to insufficient cloud resources during sudden surges in tasks, increasing task latency. The third core flaw is that scheduling rules and algorithm parameters rely on manual presets (e.g., "migrate tasks when edge node load exceeds 80%)," resulting in rigid performance of the scheduling rules. For instance, after a car factory added an AI quality inspection task, operations personnel needed to manually adjust the scheduling priority. During this period, the false positive rate of the quality inspection task increased by 5% for approximately two hours due to insufficient resource allocation. These three existing technological flaws result in poor efficiency for current load balancing business scheduling.

[0003] Therefore, how to provide a cloud-edge-device adaptive scheduling method based on cognitive decision-making to improve the efficiency of business scheduling under the cloud-edge-device collaborative architecture is an urgent problem to be solved. Summary of the Invention

[0004] This application provides a cloud-edge-device adaptive scheduling method, device, and medium based on cognitive decision-making, with the main purpose of improving the efficiency of service scheduling under the cloud-edge-device collaborative architecture.

[0005] In a first aspect, embodiments of this application provide a cloud-edge-device adaptive scheduling method based on cognitive decision-making, the method comprising: Periodically collect multi-dimensional data from the cloud, edge, and terminal, and fuse the multi-dimensional data from the cloud, edge, and terminal to generate a runtime situation map. The collection frequency of the multi-dimensional data from the cloud, edge, and terminal is set according to the business priority. Based on a large language model, the received natural language scheduling instructions are parsed into a structured target set, which includes hard constraints; The digital twin and simulation simulator is initialized based on the runtime situation map, and simulations are performed according to the first preset number of candidate scheduling strategies. The first preset number of strategy prediction index combinations are output, and effective candidate strategies that meet the hard constraints are selected from the first preset number of strategy prediction index combinations. Based on the feasible region defined by the effective candidate strategies, and combined with the reward mechanism constructed according to the structured target set, Pareto optimal scheduling strategies are dynamically generated in the digital twin and simulation simulator through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model. The Pareto optimal scheduling strategy is executed, and the actual effect values ​​are collected. The simulation prediction values ​​are compared with the actual effect values. If the difference exceeds a preset ratio, the large language model and the MADDPG model are fine-tuned to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and MADDPG model. The simulation prediction values ​​are obtained by simulating the Pareto optimal scheduling strategy in the digital twin and simulation simulation simulator.

[0006] In one implementation of this application, the periodic collection of cloud-edge-device multidimensional data and the fusion of the cloud-edge-device multidimensional data to generate a runtime situation map includes: The cloud-edge-device multidimensional data is collected periodically based on a lightweight collector. The cloud-edge-device multidimensional data includes resource indicators, application performance indicators, business context data, and infrastructure topology relationship data. The cloud-edge-device multidimensional data is stored in a time-series database, and the runtime situation map with spatiotemporal labels is generated by associating the cloud-edge-device multidimensional data corresponding to the same type of device using a graph neural network.

[0007] In one implementation of this application, the step of parsing the received natural language scheduling instructions into a structured target set based on a large language model includes: Receive the natural language scheduling instructions input by the user through the management platform; Based on the large language model, the natural language scheduling instructions are broken down into business scenarios, hard constraints, and optimization objectives; The business scenario, the hard constraints, and the optimization objectives are quantified to obtain the structured objective set.

[0008] In one implementation of this application, a digital twin and simulation simulator is initialized based on the runtime situation map, and simulations are performed according to a first preset number of candidate scheduling strategies. The first preset number of strategy prediction index combinations are then output, and effective candidate strategies that satisfy the hard constraints are selected from the first preset number of strategy prediction index combinations. The runtime situation map is injected as an initial state parameter into the complete state snapshot of the simulated cloud-edge-device system. The simulated cloud-edge-device system is set in the digital twin and simulation simulation simulator. The simulation time window of the simulated cloud-edge-device system is set according to the business scenario to be simulated, and the business traffic model corresponding to the business scenario is loaded to initialize the digital twin and simulation simulation simulator. For the first preset number of candidate scheduling strategies, after initialization, each complete simulation is run independently once in the digital twin and simulation simulation simulator, and the performance indicators during the simulation process are calculated in real time, and the combination of the first preset number of strategy prediction indicators is output. Iterate through all the combinations of the aforementioned strategy prediction indicators and check whether each one satisfies the aforementioned hard constraints; If any combination of strategy prediction indicators satisfies the hard constraint, then the candidate scheduling strategy corresponding to any combination of strategy prediction indicators is taken as the effective candidate strategy. If any combination of strategy prediction indicators fails to meet the hard constraints, the candidate scheduling strategy corresponding to that combination of strategy prediction indicators will be marked as invalid and removed.

[0009] In one implementation of this application, the step of dynamically generating a Pareto optimal scheduling strategy in the digital twin and simulation simulator through multi-agent collaborative optimization, based on the feasible region defined by the effective candidate strategy and combined with the reward mechanism constructed according to the structured target set, includes: Calculate the maximum and minimum values ​​of each prediction index in the strategy prediction index combination corresponding to the effective candidate strategy, and construct the feasible region based on the maximum and minimum values ​​of each prediction index. The feasible region is a hyperrectangular search space. Based on the aforementioned hard constraints and optimization objectives, a multi-dimensional reward mechanism is constructed. In the digital twin and simulation simulator, multiple agents constructed by the MADDPG model are initialized, and the multi-dimensional reward mechanism is matched with each of the multiple agents one by one, with each agent corresponding to a scheduling dimension; Under the constraints of the feasible domain, the first single-dimensional scheduling strategy under different scheduling dimensions is output based on the multiple agents, and the first single-dimensional scheduling strategies are concatenated to obtain the initial scheduling strategy. The initial scheduling strategy is executed in the digital twin and simulation simulator to obtain the initial strategy prediction index combination; Based on the multi-dimensional reward mechanism, a reward value is calculated for each agent according to the combination of the initial policy prediction indicators. The reward value is a weighted combination of hard constraint positive reward terms and optimization objective penalty terms. The network parameters of the plurality of agents are updated based on the reward value to obtain the updated plurality of agents; Under the constraints of the feasible region, based on the updated outputs of multiple agents, a second single-dimensional scheduling strategy under different scheduling dimensions is generated, and the second unit scheduling strategies are concatenated to obtain the updated scheduling strategy. This process is repeated for a preset number of iterations to generate the scheduling strategy and obtain the Pareto front solution set. Based on the requirements of the business scenario, the Pareto scheduling strategy with the best overall performance is selected from the Pareto front set as the Pareto optimal scheduling strategy.

[0010] In one implementation of this application, the lightweight collector supports the MQTT protocol and / or the gRPC protocol.

[0011] In one implementation of this application, the multi-dimensional reward mechanism includes: If the initial policy prediction metric combination violates any of the hard constraints, then a strong negative reward is assigned to all the agents. For each optimization objective, it is converted into a continuous positive reward proportional to its achievement, and the weight of each optimization objective in the overall reward value is dynamically adjusted according to the business scenario. The strong negative reward and the continuous positive reward are weighted and fused to form the comprehensive reward value received by each agent in each round of simulation.

[0012] In one implementation of this application, the step of fine-tuning the large language model and the MADDPG model if the comparison difference exceeds a preset proportion includes: If the relative deviation between the actual effect value and the simulation prediction value on any prediction index exceeds a preset ratio, the difference data between the actual effect value and the simulation prediction value will be used as a monitoring signal. The instruction parsing module of the large language model is fine-tuned using the supervision signal to improve its quantification accuracy of the hard constraints and optimization objectives in the natural language scheduling instructions. Furthermore, the parameters of the Critic network and the multi-dimensional reward mechanism of the MADDPG model are fine-tuned using the difference data to improve the prediction fidelity of digital twins and simulation simulators.

[0013] Secondly, embodiments of this application also provide a course adaptability assessment device, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: periodically collect cloud-edge-device multidimensional data, fuse the cloud-edge-device multidimensional data to generate a runtime situation map, wherein the collection frequency of the cloud-edge-device multidimensional data is set according to business priority; parse received natural language scheduling instructions into a structured target set based on a large language model, the structured target set including hard constraints; initialize a digital twin and simulation simulation simulator based on the runtime situation map, and perform simulations according to a first preset number of candidate scheduling strategies, outputting the first preset number of strategy prediction index combinations, and from the first preset... Effective candidate strategies that satisfy the hard constraints are selected from a combination of policy prediction indicators. Based on the feasible region defined by the effective candidate strategies and combined with the reward mechanism constructed according to the structured target set, a Pareto optimal scheduling strategy is dynamically generated in the digital twin and simulation simulation simulator through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model. The Pareto optimal scheduling strategy is executed, and the actual effect value is collected. The simulation prediction value is compared with the actual effect value. If the difference exceeds a preset proportion, the large language model and the MADDPG model are fine-tuned to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and the MADDPG model. The simulation prediction value is obtained by simulation in the digital twin and simulation simulation simulator by executing the Pareto optimal scheduling strategy.

[0014] Thirdly, embodiments of this application also provide a non-volatile computer storage medium corresponding to the method of the first aspect described above, storing computer-executable instructions, which are configured to: periodically collect multi-dimensional data from the cloud, edge, and endpoint; fuse the multi-dimensional data from the cloud, edge, and endpoint to generate a runtime situation map, wherein the collection frequency of the multi-dimensional data from the cloud, edge, and endpoint is set according to business priority; parse the received natural language scheduling instructions into a structured target set based on a large language model, wherein the structured target set includes hard constraints; initialize a digital twin and simulation simulation simulator based on the runtime situation map, and perform simulations according to a first preset number of candidate scheduling strategies, outputting the first preset number of strategy prediction index combinations, and selecting from the first preset number of strategy prediction index combinations that meet the hard constraints. Effective candidate strategies with sexual constraints are identified. Based on the feasible region defined by the effective candidate strategies and combined with the reward mechanism constructed according to the structured objective set, a Pareto optimal scheduling strategy is dynamically generated in the digital twin and simulation simulation simulator through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model. The Pareto optimal scheduling strategy is executed, and the actual effect value is collected. The simulation prediction value is compared with the actual effect value. If the difference exceeds a preset proportion, the large language model and the MADDPG model are fine-tuned to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and the MADDPG model. The simulation prediction value is obtained by simulation in the digital twin and simulation simulation simulator by executing the Pareto optimal scheduling strategy.

[0015] This application provides a cloud-edge-device adaptive scheduling method, device, and medium based on cognitive decision-making, which has the following beneficial effects: Utilizing digital twin technology to perform high-fidelity simulation of the cloud-edge-device system allows for the simulation of resource usage under various business scenarios before actual deployment, identifying potential bottlenecks in advance. This predictive mechanism enables scheduling strategies to be adjusted before problems occur, avoiding service interruptions or performance degradation due to sudden loads. Through pre-set hard constraints, combined with real-time monitoring and simulation, resource configuration can be dynamically adjusted to ensure service quality requirements are met even during peak periods. Especially in scenarios with extremely high real-time requirements, such as video surveillance, this mechanism effectively guarantees system stability and reliability. The MADDPG algorithm allows multiple agents to collaboratively optimize different scheduling dimensions to find Pareto optimal solutions. This not only ensures that business indicators for any scheduling dimension are strictly adhered to but also minimizes operating costs and energy consumption without affecting service quality. By continuously monitoring the deviation between actual operating data and simulation predictions, a model fine-tuning mechanism is triggered once a significant difference is detected. This self-evolving capability ensures that the scheduling strategy can adapt to environmental changes and maintain long-term effectiveness and accuracy. Combined with the aforementioned mechanism, this improves the efficiency of business scheduling under the cloud-edge-device collaborative architecture. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of a cloud-edge-device adaptive scheduling method based on cognitive decision-making is provided for embodiments of this application; Figure 2 This is a schematic diagram of the internal structure of a cloud-edge-device adaptive scheduling device based on cognitive decision-making, provided as an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] This application provides a cloud-edge-device adaptive scheduling method, device, and medium based on cognitive decision-making, which aims to improve the efficiency of service scheduling under a cloud-edge-device collaborative architecture.

[0019] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0020] Figure 1 A flowchart illustrating a cloud-edge-device adaptive scheduling method based on cognitive decision-making, provided in an embodiment of this application. Figure 1 As shown in the figure, the cloud-edge-device adaptive scheduling method based on cognitive decision-making provided in this application embodiment specifically includes the following steps: Step 101: Periodically collect multi-dimensional data from the cloud, edge, and terminal, and fuse the multi-dimensional data from the cloud, edge, and terminal to generate a runtime situation map. The collection frequency of the multi-dimensional data from the cloud, edge, and terminal is set according to the service priority.

[0021] Step 102: Based on the large language model, parse the received natural language scheduling instructions into a structured target set, which includes hard constraints.

[0022] Step 103: Initialize the digital twin and simulation simulator based on the runtime situation map, and perform simulations according to the first preset number of candidate scheduling strategies, output the first preset number of strategy prediction index combinations, and select effective candidate strategies that meet the hard constraints from the first preset number of strategy prediction index combinations.

[0023] Step 104: Based on the feasible region defined by the effective candidate strategies, and in conjunction with the reward mechanism constructed according to the structured target set, a Pareto optimal scheduling strategy is dynamically generated in the digital twin and simulation simulator through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model.

[0024] Step 105: Execute the Pareto optimal scheduling strategy, collect the actual effect value, compare the simulation prediction value with the actual effect value, and if the difference exceeds a preset ratio, fine-tune the large language model and the MADDPG model to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and the MADDPG model. The simulation prediction value is obtained by simulating the Pareto optimal scheduling strategy in the digital twin and simulation simulation simulator.

[0025] This application provides a cloud-edge-device adaptive scheduling method based on cognitive decision-making, comprising: using digital twin technology to perform high-fidelity simulation of the cloud-edge-device system, which can simulate resource usage under various business scenarios before actual deployment and identify potential bottlenecks in advance. This predictive mechanism enables the scheduling strategy to be adjusted before problems occur, avoiding service interruptions or performance degradation caused by sudden loads. Through pre-set hard constraints, combined with real-time monitoring and simulation, resource configuration can be dynamically adjusted to ensure that service quality requirements are met even during peak periods. Especially in scenarios with extremely high real-time requirements, such as video surveillance, this mechanism effectively guarantees the stability and reliability of the system. The MADDPG algorithm allows multiple agents to collaboratively optimize different scheduling dimensions to find Pareto optimal solutions. This not only ensures that business indicators for any scheduling dimension are strictly adhered to, but also minimizes operating costs and energy consumption without affecting service quality. By continuously monitoring the deviation between actual operating data and simulation predictions, a model fine-tuning mechanism is triggered once a significant difference is detected. This self-evolutionary capability ensures that the scheduling strategy can adapt to environmental changes, maintain long-term effectiveness and accuracy, and, combined with the aforementioned mechanism, improve the efficiency of business scheduling under the cloud-edge-device collaborative architecture.

[0026] As a refinement of the embodiments of this application, when performing step 101, periodically collecting cloud-edge-device multidimensional data and fusing the cloud-edge-device multidimensional data to generate a runtime situation map, the following implementation methods can also be adopted, but are not limited to: periodically collecting cloud-edge-device multidimensional data based on a lightweight collector, wherein the cloud-edge-device multidimensional data includes resource indicators, application performance indicators, business context data, and infrastructure topology relationship data; storing the cloud-edge-device multidimensional data based on a time-series database, and associating the cloud-edge-device multidimensional data corresponding to the same device type based on a graph neural network to generate the runtime situation map with spatiotemporal labels.

[0027] It is important to understand that the cloud-edge-device multidimensional data is collected from a system composed of various cloud-edge-device devices. These devices include cloud devices, edge computing devices, and terminal devices. Cloud devices include, but are not limited to, cloud servers and distributed storage systems; edge computing devices include, but are not limited to, edge servers and industrial network management systems; and terminal devices include, but are not limited to, mobile terminals, cameras, and sensors.

[0028] In some embodiments, the lightweight collector includes, but is not limited to, Prometheus Exporter; the resource metrics include, but are not limited to, CPU utilization, memory, GPU utilization, and network bandwidth; the application performance metrics include, but are not limited to, service response latency and throughput; the business context data includes, but is not limited to, work order type and fault alarm level; the collection frequency range of the cloud-edge-device multidimensional data is 50 to 200 ms, and the collection frequency setting is determined according to the business type, with the highest priority business (such as device control) defaulting to 50 ms.

[0029] In some embodiments, the lightweight collector supports the MQTT protocol and / or the gRPC protocol.

[0030] To facilitate understanding of the runtime situation map generation process, this embodiment provides an exemplary description, including a lightweight collector on cloud-edge devices collecting multi-dimensional data in real time, such as edge node CPU utilization, cloud instance cost, and video analytics task latency; storing the data through a time-series database (such as InfluxDB), and constructing a situation map that associates "devices, tasks, and metrics" based on a graph neural network (GNN), such as "edge node A, video analytics task B, current latency 120ms"; and generating a runtime situation map with spatiotemporal labels (such as "May 20, 2024, 08:30, edge node of Building 1 in the park").

[0031] As a refinement of the above embodiments, when performing step 102, which involves parsing the received natural language scheduling instructions into a structured target set based on a large language model, the following implementation methods can also be adopted, but are not limited to: receiving the natural language scheduling instructions input by the user through a management platform; splitting the natural language scheduling instructions into business scenarios, hard constraints, and optimization goals based on the large language model; and quantifying the business scenarios, hard constraints, and optimization goals respectively to obtain the structured target set.

[0032] In some embodiments, the process of parsing the natural language scheduling instructions includes, but is not limited to: using a large language model (LLM) fine-tuned based on Llama 2-7B to convert fuzzy semantic intent into mathematical expressions, such as converting "reduce nighttime costs" into "from 10 PM to 6 AM the next day, cloud instance costs are ≤ 80% of daytime costs". In conjunction with specific application scenarios, the implementation of parsing the natural language scheduling instructions can also include: Users input natural language instructions through the management platform, such as "An exhibition will be held in the park next week. Security video P99 latency must be ≤100ms, and cloud costs during the exhibition (9:00-18:00) must be reduced by 25%"; LLM breaks down the instruction into "business scenario (exhibition), hard constraint (P99 latency ≤100ms), optimization goal (cost reduction of 25%)"; and outputs a structured set of goals, such as {Hard constraint: P99_Latency ≤ 100ms; Optimization goal: Cloud_Cost ≤ original cost × 75%; Time range: 2024-05-25 to 2024-05-28 9:00-18:00}. The structured set of goals provided in this example only shows one hard constraint and one optimization goal. It should be understood that the number of hard constraints and optimization goals is not limited to one type and can be added and adjusted according to the actual scenario.

[0033] As a refinement of the above embodiments, when performing step 103, which involves initializing the digital twin and simulation simulation simulator based on the runtime situation map, performing simulations according to a first preset number of candidate scheduling strategies, outputting the first preset number of strategy prediction index combinations, and selecting effective candidate strategies that meet the hard constraints from the first preset number of strategy prediction index combinations, the following implementation methods can also be adopted, but are not limited to: injecting the runtime situation map as an initial state parameter into a complete state snapshot of the simulated cloud-edge-device system, wherein the simulated cloud-edge-device system is set in the digital twin and simulation simulation simulator, and setting the simulation time window of the simulated cloud-edge-device system and loading the corresponding business scenario according to the business scenario to be simulated. The system initializes the digital twin and simulation simulator using a traffic model. For the first preset number of candidate scheduling strategies, it independently runs a complete simulation once in the simulator after initialization, calculating performance metrics in real time and outputting the first preset number of strategy prediction metric combinations. It iterates through all the strategy prediction metric combinations, checking each one to see if it meets the hard constraints. If any strategy prediction metric combination meets the hard constraints, the candidate scheduling strategy corresponding to that combination is considered a valid candidate strategy. If any strategy prediction metric combination does not meet the hard constraints, the candidate scheduling strategy corresponding to that combination is marked as invalid and removed. The strategy prediction metric combinations include, but are not limited to, the first prediction delay and the first prediction cost predicted based on the first candidate strategy, where the first candidate strategy is any one of the first preset number of candidate scheduling strategies.

[0034] In some embodiments, the digital twin and simulation simulator includes, but is not limited to, the use of a discrete event simulation engine, such as AnyLogic, which is used to synchronize the device status and task load of the physical environment; supports simulation with time windows of 1 to 60 minutes, outputs the predicted values ​​of the strategy for "latency, cost, and energy consumption"; and achieves a prediction error of ≤10% through calibration with historical data.

[0035] To facilitate understanding of the implementation process of step 103, this application provides an exemplary description in conjunction with a real-world scenario. For example: the current situation map, the set of quantified targets, and 3 to 5 candidate scheduling strategies are input into the digital twin and simulation simulator (e.g., "Strategy 1: Migrate tasks to the cloud when edge load exceeds 70%; Strategy 2: Pre-allocate 20% of cloud Spot instances for exhibition tasks"). The digital twin model simulates the task execution process for the next hour and outputs the predicted indicators for each strategy, such as "Strategy 1 predicts a latency of 95ms and a cost reduction of 22%; Strategy 2 predicts a latency of 92ms and a cost reduction of 28%". Strategies that do not meet the hard constraints are eliminated (e.g., a strategy predicting a latency of 110ms is directly excluded), and 2 to 3 valid candidate strategies are retained.

[0036] As a refinement of the above embodiments, when performing step 104, which involves defining the feasible region based on the effective candidate strategy and combining it with the reward mechanism constructed according to the structured objective set, and dynamically generating a Pareto optimal scheduling strategy through multi-agent collaborative optimization in the digital twin and simulation simulator, the following implementation methods can also be adopted, but are not limited to: calculating the maximum and minimum values ​​of each prediction index in the strategy prediction index combination corresponding to the effective candidate strategy; constructing the feasible region based on the maximum and minimum values ​​of each prediction index, wherein the feasible region is a hyperrectangular search space; constructing a multi-dimensional reward mechanism based on the hard constraints and the optimization objective; initializing multiple agents constructed by the MADDPG model in the digital twin and simulation simulator, and matching the multi-dimensional reward mechanism with each of the multiple agents one by one, wherein each agent corresponds to a scheduling dimension; and under the constraints of the feasible region, outputting different values ​​based on the multiple agents. The system firstly implements a single-dimensional scheduling strategy under the scheduling dimension, and concatenates these strategies to obtain an initial scheduling strategy. This initial scheduling strategy is then executed in the digital twin and simulation simulator to obtain a combination of initial strategy prediction indicators. Based on the multi-dimensional reward mechanism, a reward value is calculated for each agent according to the initial strategy prediction indicator combination. This reward value is a weighted combination of a hard constraint positive reward term and an optimization objective penalty term. The network parameters of the agents are updated based on the reward values ​​to obtain updated agents. Under the constraints of the feasible region, second-dimensional scheduling strategies under different scheduling dimensions are output based on the updated agents, and these second-dimensional scheduling strategies are concatenated to obtain an updated scheduling strategy. This process is repeated iteratively for a preset number of rounds to generate the scheduling strategy, resulting in a Pareto front solution set. Based on the requirements of the business scenario, the Pareto scheduling strategy with the best overall performance is selected from the Pareto front set as the Pareto optimal scheduling strategy. The types of scheduling dimensions include, but are not limited to, scheduling based on delay, scheduling based on operating cost, and scheduling based on computing resource load.

[0037] In some embodiments, the MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm is used to construct the multiple agents, with one agent corresponding to each optimization objective (performance, cost, energy consumption); under the premise of satisfying hard constraints (such as "latency ≤ 100ms"), the comprehensive optimal policy is selected from the Pareto front (a set of optimal solutions for multiple objectives) to obtain the Pareto optimal policy.

[0038] To facilitate understanding of the above embodiments, this application provides an exemplary description of generating a Pareto optimal scheduling strategy in conjunction with an application scenario, including: inputting effective candidate strategies and a structured target set into the simulation environment provided by the twin model; setting up three agents of the MADDPG algorithm in the simulation environment (corresponding to "latency", "cost", and "energy consumption" respectively), and iteratively training the scheduling strategy for 500-1000 rounds, adjusting the strategy through a reward function (such as "10 points for meeting the latency target, 5 points deducted for exceeding the cost target"); generating a scheduling strategy set on the Pareto front through polling adjustment, and selecting the comprehensive optimal strategy from it, such as "pre-allocating 25% of cloud Spot instances, migrating tasks when the edge load exceeds 65%, predicting a latency of 90ms and a cost reduction of 27%".

[0039] As a refinement of the above embodiments, the multi-dimensional reward mechanism includes: if the initial policy prediction index combination violates any of the hard constraints, then a strong negative reward is assigned to all the agents; for each optimization objective, it is converted into a continuous positive reward proportional to its achievement degree, and the weight of each optimization objective in the comprehensive reward value is dynamically adjusted according to the business scenario; the strong negative reward and the continuous positive reward are weighted and fused to form the comprehensive reward value received by each agent in each round of simulation.

[0040] As a refinement of the above embodiments, when performing step 105, if the difference exceeds a preset proportion, fine-tuning the large language model and the MADDPG model can also be carried out in the following ways, but not limited to: if the relative deviation between the actual effect value and the simulation prediction value on any prediction index exceeds a preset proportion, the difference data between the actual effect value and the simulation prediction value is used as a supervision signal; the supervision signal is used to fine-tune the instruction parsing module of the large language model to improve its quantification accuracy of the hard constraints and optimization objectives in the natural language scheduling instructions; and the difference data is used to fine-tune the Critic network and the multi-dimensional reward mechanism parameters of the MADDPG model to improve the prediction fidelity of the digital twin and simulation simulation simulator.

[0041] In some embodiments, commands are issued via the Kubernetes API or EdgeX Foundry to execute the Pareto optimal scheduling strategy, such as "expand the cloud Spot instance to 5 and configure the load threshold of edge node A to 65%"; execution effect data is collected in real time, such as "actual latency 88ms, cost reduction 29%", and compared with the simulation prediction value (latency 90ms, cost reduction 27%); if the difference is ≤10%, the current model parameters are retained; if the difference is >10% (e.g., actual cost reduction is only 20%), the difference data is used as a reward signal to fine-tune the LLM and MADDPG models to ensure more accurate decisions in the next iteration. During execution, a Grafana dashboard is integrated to display the strategy execution effect in real time.

[0042] To facilitate understanding of the embodiments of this application, this embodiment provides a first application description. The application background includes: a smart park in a commercial complex, comprising one central cloud cluster (10 8-core 16G cloud servers) and 15 edge nodes (distributed in buildings 1-15, each node with 4 cores and 8G). Services include security video analytics (200 cameras, requiring P99 latency ≤100ms) and environmental monitoring (temperature and humidity data collection, requiring latency ≤500ms). An exhibition will be held from May 25-28, 2024, which is expected to increase the video analytics workload by 50%. It is necessary to ensure security services while controlling cloud costs. The deployment process includes: deploying a Prometheus Server in the central cloud and a Node Exporter on the edge nodes, with a data collection frequency of 50ms. Data is uploaded via the MQTT protocol and stored in InfluxDB. The LLM is fine-tuned based on Llama 2-7B, with the training corpus containing historical scheduling instructions from the park (such as "ensure video latency during the exhibition"), and deployed on a 2-core 4G cloud server. A twin model is built using AnyLogic to synchronize edge node load, number of cameras, and cloud instance specifications, with a simulation time window of 1 hour. The MADDPG algorithm is implemented based on PyTorch, with three agents optimizing "latency," "cost," and "energy consumption" respectively, undergoing 1000 training iterations, and deployed on a 4-core 8G cloud server. Interaction with the cloud-edge-device cluster is achieved through the Kubernetes API, with Grafana used to monitor the execution effect, and a feedback cycle of 1 minute. The operation process included: One hour before the exhibition (May 25th, 8:00 AM), the situation map showed "Current edge node average load 60%, video latency 80ms, 5 cloud instances idle"; the operations and maintenance personnel entered the command "During the exhibition (9:00-18:00), ensure security video P99 latency ≤100ms, cloud cost reduced by 25%", LLM quantification was: {Constraint 1: P99_Latency ≤100ms; Objective 1: Cloud_Cost ≤ 100ms; Target 2: Cloud_Cost ≤ 100ms; Target 3: Cloud_Cost ≤ 100ms; Target 4: Cloud_Cost ≤ 100ms; Target 5: Cloud_Cost ≤ 100ms; Target 6: Cloud_Cost ≤ 100ms; Target 7: Cloud_Cost ≤ 100ms; Target 8: Cloud On May 24th, the same period last year was 75%; the twin model simulated three candidate strategies and selected two effective strategies (Strategy A: for migration tasks with edge load exceeding 70%, pre-allocate 3 cloud Spot instances; Strategy B: for migration tasks with edge load exceeding 65%, pre-allocate 5 cloud Spot instances); after training the MADDPG algorithm, strategy B was selected, with a prediction latency of 92ms and a cost reduction of 28%; when strategy B was implemented, the actual video P99 latency was 88ms and the cloud cost was reduced by 29%, with a difference from the prediction of ≤5%, and the model did not require significant fine-tuning.

[0043] To facilitate understanding of the embodiments of this application, this embodiment provides a first application description. The application background includes: an automotive parts factory, comprising one central cloud (5 servers with 16 cores and 32GB of RAM) and eight edge nodes (distributed on the production line, each node with 8 cores and 16GB of RAM). The business includes equipment control (robotic arm movement commands, requiring latency ≤50ms), AI quality inspection (part defect detection, requiring accuracy ≥99%), and data storage (production logs, requiring storage latency ≤1s). It is necessary to ensure the smooth operation of equipment control while reducing the energy consumption of edge nodes and the cost of cloud storage. The deployment and operation of the method include: inputting the scheduling intent, such as "ensuring that the control latency of the robotic arm is ≤50ms, the quality inspection accuracy is ≥99%, the energy consumption of edge nodes is reduced by 10%, and the cloud storage cost is reduced by 20%"; setting the global situational awareness acquisition frequency to 30ms, the digital twin simulation window to 30 minutes, and adding the "quality inspection accuracy" optimization target to the MADDPG agent; after the digital twin simulation, selecting "the robotic arm tasks are preferentially allocated to edge nodes, the quality inspection tasks are migrated to the cloud when the edge load exceeds 75%, and the production logs are compressed and uploaded to the cloud at night (22:00-6:00)".

[0044] In summary, the embodiments of this application can achieve the following effects, including: 1. Utilizing digital twin technology for high-fidelity simulation of cloud-edge-device systems allows for the simulation of resource usage under various business scenarios before actual deployment, identifying potential bottlenecks in advance. This predictive mechanism enables scheduling strategies to be adjusted before problems occur, avoiding service interruptions or performance degradation due to sudden load increases. Through pre-defined hard constraints, combined with real-time monitoring and simulation, resource configuration can be dynamically adjusted to ensure service quality requirements are met even during peak periods. This mechanism effectively guarantees system stability and reliability, especially in scenarios with extremely high real-time requirements such as video surveillance. The MADDPG algorithm allows multiple agents to collaboratively optimize different scheduling dimensions to find Pareto optimal solutions. This not only ensures that business metrics for any scheduling dimension are strictly adhered to but also minimizes operating costs and energy consumption without affecting service quality. By continuously monitoring the deviation between actual operating data and simulation predictions, a model fine-tuning mechanism is triggered once a significant difference is detected. This self-evolving capability ensures that scheduling strategies can adapt to environmental changes, maintaining long-term effectiveness and accuracy, thereby improving the efficiency of business scheduling under the cloud-edge-device collaborative architecture.

[0045] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a cloud-edge-device adaptive scheduling device based on cognitive decision-making, the structure of which is as follows: Figure 2 As shown.

[0046] Figure 2This is a schematic diagram of the internal structure of a cloud-edge-device adaptive scheduling device based on cognitive decision-making, provided as an embodiment of this application. Figure 2 As shown, the device includes: At least one processor 201; And a memory 202 that is communicatively connected to at least one processor; The memory 202 stores instructions executable by at least one processor. These instructions are executed by at least one processor 201 to enable the processor 201 to: periodically collect multi-dimensional data from the cloud, edge, and endpoint; fuse the multi-dimensional data to generate a runtime situation map, wherein the collection frequency of the multi-dimensional data is set according to service priority; parse received natural language scheduling instructions into a structured target set based on a large language model, the structured target set including hard constraints; initialize a digital twin and simulation simulator based on the runtime situation map, and perform simulations according to a first preset number of candidate scheduling strategies, outputting the first preset number of strategy prediction index combinations, and selecting strategies from the first preset number of strategy prediction index combinations that satisfy the hard constraints. Effective candidate strategies; based on the feasible region defined by the effective candidate strategies, and combined with the reward mechanism constructed according to the structured target set, a Pareto optimal scheduling strategy is dynamically generated in the digital twin and simulation simulation simulator through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model. The Pareto optimal scheduling strategy is executed, and the actual effect value is collected. The simulation prediction value is compared with the actual effect value. If the difference exceeds a preset proportion, the large language model and the MADDPG model are fine-tuned to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and the MADDPG model. The simulation prediction value is obtained by simulation in the digital twin and simulation simulation simulator by executing the Pareto optimal scheduling strategy.

[0047] Some embodiments of this application provide corresponding to Figure 1A non-volatile computer storage medium stores computer-executable instructions, which are configured to: periodically collect multi-dimensional data from cloud, edge, and endpoint; fuse the multi-dimensional data from cloud, edge, and endpoint to generate a runtime situation map, wherein the collection frequency of the multi-dimensional data from cloud, edge, and endpoint is set according to business priority; parse received natural language scheduling instructions into a structured target set based on a large language model, wherein the structured target set includes hard constraints; initialize a digital twin and simulation simulation simulator based on the runtime situation map, and perform simulations according to a first preset number of candidate scheduling strategies, outputting the first preset number of strategy prediction index combinations, and selecting effective candidate strategies that meet the hard constraints from the first preset number of strategy prediction index combinations; based on the... The feasible domain defined by the effective candidate strategies is combined with the reward mechanism constructed based on the structured target set. In the digital twin and simulation simulation simulator, a Pareto optimal scheduling strategy is dynamically generated through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model. The Pareto optimal scheduling strategy is executed, and actual performance values ​​are collected. The simulated predicted values ​​are compared with the actual performance values. If the difference exceeds a preset proportion, the large language model and the MADDPG model are fine-tuned to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and the MADDPG model. The simulated predicted values ​​are obtained through simulation by executing the Pareto optimal scheduling strategy in the digital twin and simulation simulation simulator.

[0048] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0049] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0050] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0055] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0056] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media 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 memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0057] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0058] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A cloud-edge-device adaptive scheduling method based on cognitive decision-making, characterized in that, The method includes: Periodically collect multi-dimensional data from the cloud, edge, and terminal, and fuse the multi-dimensional data from the cloud, edge, and terminal to generate a runtime situation map. The collection frequency of the multi-dimensional data from the cloud, edge, and terminal is set according to the business priority. Based on a large language model, the received natural language scheduling instructions are parsed into a structured target set, which includes hard constraints; The digital twin and simulation simulator is initialized based on the runtime situation map, and simulations are performed according to the first preset number of candidate scheduling strategies. The first preset number of strategy prediction index combinations are output, and effective candidate strategies that meet the hard constraints are selected from the first preset number of strategy prediction index combinations. Based on the feasible region defined by the effective candidate strategies, and combined with the reward mechanism constructed according to the structured target set, Pareto optimal scheduling strategies are dynamically generated in the digital twin and simulation simulator through multi-agent collaborative optimization. The multiple agents are constructed using the MADDPG model. The Pareto optimal scheduling strategy is executed, and the actual effect values ​​are collected. The simulation prediction values ​​are compared with the actual effect values. If the difference exceeds a preset ratio, the large language model and the MADDPG model are fine-tuned to generate a fine-tuned Pareto optimal scheduling strategy based on the fine-tuned large language model and MADDPG model. The simulation prediction values ​​are obtained by simulating the Pareto optimal scheduling strategy in the digital twin and simulation simulation simulator.

2. The cloud-edge-device adaptive scheduling method based on cognitive decision-making according to claim 1, characterized in that, The periodic collection of multi-dimensional data from the cloud, edge, and endpoint, and the fusion of this multi-dimensional data to generate a runtime situation map, includes: The cloud-edge-device multidimensional data is collected periodically based on a lightweight collector. The cloud-edge-device multidimensional data includes resource indicators, application performance indicators, business context data, and infrastructure topology relationship data. The cloud-edge-device multidimensional data is stored in a time-series database, and the runtime situation map with spatiotemporal labels is generated by associating the cloud-edge-device multidimensional data corresponding to the same type of device using a graph neural network.

3. The cloud-edge-device adaptive scheduling method based on cognitive decision-making according to claim 2, characterized in that, The process of parsing the received natural language scheduling instructions into a structured target set based on a large language model includes: Receive the natural language scheduling instructions input by the user through the management platform; Based on the large language model, the natural language scheduling instructions are broken down into business scenarios, hard constraints, and optimization objectives; The business scenario, the hard constraints, and the optimization objectives are quantified to obtain the structured objective set.

4. The cloud-edge-device adaptive scheduling method based on cognitive decision-making according to claim 3, characterized in that, The digital twin and simulation simulator is initialized based on the runtime situation map, and simulations are performed according to a first preset number of candidate scheduling strategies. The first preset number of strategy prediction index combinations are output, and effective candidate strategies that meet the hard constraints are selected from the first preset number of strategy prediction index combinations. The runtime situation map is injected as an initial state parameter into the complete state snapshot of the simulated cloud-edge-device system. The simulated cloud-edge-device system is set in the digital twin and simulation simulation simulator. The simulation time window of the simulated cloud-edge-device system is set according to the business scenario to be simulated, and the business traffic model corresponding to the business scenario is loaded to initialize the digital twin and simulation simulation simulator. For the first preset number of candidate scheduling strategies, after initialization, each complete simulation is run independently once in the digital twin and simulation simulation simulator, and the performance indicators during the simulation process are calculated in real time, and the combination of the first preset number of strategy prediction indicators is output. Iterate through all the combinations of the aforementioned strategy prediction indicators and check whether each one satisfies the aforementioned hard constraints; If any combination of strategy prediction indicators satisfies the hard constraint, then the candidate scheduling strategy corresponding to any combination of strategy prediction indicators is taken as the effective candidate strategy. If any combination of strategy prediction indicators fails to meet the hard constraints, the candidate scheduling strategy corresponding to that combination of strategy prediction indicators will be marked as invalid and removed.

5. The cloud-edge-device adaptive scheduling method based on cognitive decision-making according to claim 4, characterized in that, The feasible region defined based on the effective candidate strategies, combined with the reward mechanism constructed according to the structured target set, dynamically generates Pareto optimal scheduling strategies in the digital twin and simulation simulator through multi-agent collaborative optimization, including: Calculate the maximum and minimum values ​​of each prediction index in the strategy prediction index combination corresponding to the effective candidate strategy, and construct the feasible region based on the maximum and minimum values ​​of each prediction index. The feasible region is a hyperrectangular search space. Based on the aforementioned hard constraints and optimization objectives, a multi-dimensional reward mechanism is constructed. In the digital twin and simulation simulator, multiple agents constructed by the MADDPG model are initialized, and the multi-dimensional reward mechanism is matched with each of the multiple agents one by one, with each agent corresponding to a scheduling dimension; Under the constraints of the feasible domain, the first single-dimensional scheduling strategy under different scheduling dimensions is output based on the multiple agents, and the first single-dimensional scheduling strategies are concatenated to obtain the initial scheduling strategy. The initial scheduling strategy is executed in the digital twin and simulation simulator to obtain the initial strategy prediction index combination; Based on the multi-dimensional reward mechanism, a reward value is calculated for each agent according to the combination of the initial policy prediction indicators. The reward value is a weighted combination of hard constraint positive reward terms and optimization objective penalty terms. The network parameters of the plurality of agents are updated based on the reward value to obtain the updated plurality of agents; Under the constraints of the feasible region, based on the updated outputs of multiple agents, a second single-dimensional scheduling strategy under different scheduling dimensions is generated, and the second unit scheduling strategies are concatenated to obtain the updated scheduling strategy. This process is repeated for a preset number of iterations to generate the scheduling strategy and obtain the Pareto front solution set. Based on the requirements of the business scenario, the Pareto scheduling strategy with the best overall performance is selected from the Pareto front set as the Pareto optimal scheduling strategy.

6. The cloud-edge-device adaptive scheduling method based on cognitive decision-making according to claim 5, characterized in that, The lightweight collector supports the MQTT protocol and / or gRPC protocol.

7. The cloud-edge-device adaptive scheduling method based on cognitive decision-making according to claim 5, characterized in that, The multi-dimensional reward mechanism includes: If the initial policy prediction metric combination violates any of the hard constraints, then a strong negative reward is assigned to all the agents. For each optimization objective, it is converted into a continuous positive reward proportional to its achievement, and the weight of each optimization objective in the overall reward value is dynamically adjusted according to the business scenario. The strong negative reward and the continuous positive reward are weighted and fused to form the comprehensive reward value received by each agent in each round of simulation.

8. A cloud-edge-device adaptive scheduling method based on cognitive decision-making according to any one of claims 1-7, characterized in that, If the comparison difference exceeds a preset proportion, the fine-tuning of the large language model and the MADDPG model includes: If the relative deviation between the actual effect value and the simulation prediction value on any prediction index exceeds a preset ratio, the difference data between the actual effect value and the simulation prediction value will be used as a monitoring signal. The instruction parsing module of the large language model is fine-tuned using the supervision signal to improve its quantification accuracy of the hard constraints and optimization objectives in the natural language scheduling instructions. Furthermore, the parameters of the Critic network and the multi-dimensional reward mechanism of the MADDPG model are fine-tuned using the difference data to improve the prediction fidelity of digital twins and simulation simulation simulators.

9. A cloud-edge-device adaptive scheduling device based on cognitive decision-making, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a cloud-edge-device adaptive scheduling method based on cognitive decision-making as described in any one of claims 1-8.

10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, the cloud-edge-device adaptive scheduling method based on cognitive decision-making as described in any one of claims 1-8 is implemented.

Images