Communication-limited multi-agent distributed optimization control method
By initializing the distributed optimization model and data computation, generating an initial set of control parameters and performing hardware status checks, the problems of poor data constraints and insufficient security in communication-constrained multi-agent optimization control are solved, achieving efficient data constraints and improved security.
Patent Information
- Application Number
- CN202511158165.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-19
- Publication Date
- 2025-11-14
AI Technical Summary
Existing technologies lack the application of gradient descent algorithms under communication constraints and fail to incorporate regularization constraints into the cost function, resulting in poor data constraint effects, an inability to accurately understand the agent's situation, and a lack of dynamic integration of policy parameters and hardware status, affecting the reference value and security of the analysis results.
A communication-constrained multi-agent distributed optimization control method is adopted. By initializing the distributed optimization model, an initial set of control parameters is generated and the hardware status is checked. The information weight matrix and policy parameters are calculated by combining the pre-acquired data, and the relative entropy constraint is updated. Finally, the optimized control instruction set is generated and the adjustment parameters are fed back.
It achieves efficient local optimization, improves the effectiveness of data constraints and overall fault tolerance, ensures the accuracy and security of analysis results, and maintains the availability of core functions.
Smart Images
Figure CN120949573A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of multi-agent communication technology, specifically to a distributed optimization control method for communication-constrained multi-agent systems. Background Technology
[0002] With the development of technology, multi-agent systems are being applied to detection in various industries, replacing manual labor and improving detection efficiency. In situations where communication is limited, it is crucial to study distributed decision-making and control strategies so that agents can self-adjust and optimize based on various data.
[0003] Existing technologies, such as the invention application patent with announcement number CN116862417A, disclose a distributed dynamic coverage control algorithm based on multiple agents. The obstacle detection module identifies obstacles based on environmental data collected by the environmental perception module. The environmental perception data is shared and transmitted through the dynamic communication network maintenance module. When an agent fails, the dynamic communication network maintenance module immediately detects this change and quickly reconstructs the network, responding to the agent failure handling module. The adaptive control module adjusts the behavioral parameters of each agent according to the new agent state and number to maintain effective coverage and perform tasks. The distributed decision-making module considers various state parameters provided by the adaptive control and coordinates among different agents to formulate a globally optimal coverage scheme and control strategy.
[0004] Regarding the above-mentioned solutions, the inventors of this application have discovered that the above-mentioned technologies have at least the following technical problems: 1. Currently, there is a lack of efficient local optimization through gradient descent algorithms. Regularization constraints are not integrated into the cost function; instead, constraints are directly applied to the data. The lack of mathematical optimization to smooth the constraints fails to improve the constraint effect on the data, thus making it impossible to more accurately understand the actual situation of the agent. Furthermore, it cannot guarantee the comprehensiveness and authenticity of the analysis process when the agent's communication is limited, thereby failing to guarantee the reference value and accuracy of the analysis results. 2. Currently, there is a lack of dynamic integration of strategy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results. This fails to provide a basis for optimization while suppressing hardware risks and cannot ensure implementation security, resulting in the inability to maintain the availability of core functions and reducing overall fault tolerance. Summary of the Invention
[0005] To address the aforementioned technical shortcomings, the purpose of this application is to provide a communication-constrained multi-agent distributed optimization control method.
[0006] To solve the above technical problems, this application adopts the following technical solution: This application provides a communication-constrained multi-agent distributed optimization control method, including: S1, based on the pre-acquisition of local state data of each agent and a preset communication topology map, initializing the distributed optimization model, generating the initial control parameter set of each agent, and performing hardware status checks on each agent.
[0007] S2. When the hardware is in normal condition, based on the initial control parameter set and the pre-acquired local constraint data, the parameters are integrated to generate preliminary control parameters.
[0008] S3. Based on the preliminary control parameters and the pre-acquired restricted communication data, calculate the fusion information weight matrix using a function.
[0009] S4. Calculate the strategy parameters based on the information weight matrix and the pre-acquired system target data, and perform relative entropy constraint updates.
[0010] S5. Based on the updated policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results, optimize the control instruction set and generate an optimized control instruction set.
[0011] S6. Based on the control instruction set and the pre-acquired execution feedback data, calculate the evaluation index and feed it back to S3 and S4.
[0012] Preferably, generating the initial control parameter set for each agent includes: merging the data in the agent's local state data into a state vector, and denoting it as... , This is represented by the ID number corresponding to each agent. , This is represented by the total number of agents, and calculated according to the formula. Generate the initial control parameter set for each agent. , The initial control parameters for each agent are denoted as , where and These are represented by the historical mean and standard deviation of the state vector, respectively. Represented as a weight matrix of a preset communication topology. It is represented as a bias term in the weight matrix.
[0013] Preferably, the step of checking the hardware status of each agent includes: checking the hardware status of each agent, generating hardware evaluation coefficients for each agent, and generating a global judgment flag based on the decision comparison of the hardware evaluation coefficients. When the hardware evaluation coefficient is greater than the hardware evaluation coefficient threshold, the global judgment flag is 1, and S2 is executed; otherwise, the global judgment flag is 0, and the fault handling stage is entered.
[0014] Preferably, the step of generating preliminary control parameters by integrating the parameters based on the initial control parameter set and the pre-acquired local constraint data includes: using calculation formulas. Generate preliminary control parameters , Let be the gradient of the cost function, where Represented as control parameters, This is represented as local optimization constraint data. Represented as the preset regularization coefficient, It is expressed as the square of the norm of the control parameters and the initial control parameters.
[0015] Preferably, the step of calculating the fusion information weight matrix based on the preliminary control parameters and the pre-acquired restricted communication data includes: using a fractional attention mechanism to fuse the preliminary control parameters and the pre-acquired restricted communication data; extracting a query vector from the preliminary control parameter set and extracting key and value vectors from the restricted communication data; simultaneously calculating a decay factor and dynamically adjusting the weights corresponding to the restricted communication data based on the decay factor; using the product of the normalized decay factor and the query-key dot product as the attention score to generate weight matrix elements, and integrating and outputting them as the fusion information weight matrix.
[0016] Preferably, the step of using the product of the normalized decay factor and the query-key dot product as the attention score, generating weight matrix elements, and integrating and outputting them as a fusion information weight matrix includes: denoting the query vector as... Restricted communication data is denoted as The key vector is denoted as The value vector is denoted as Through the calculation formula: Generate weight matrix elements and information weight matrix ,in This is represented as the attenuation factor for restricted communication data. Represented as query-key dot product, This is expressed as the dimension of the key vector. It is represented as a normalization function.
[0017] Preferably, the step of calculating strategy parameters based on the information weight matrix and pre-acquired system target data, and performing relative entropy constraint updates, includes: according to the calculation formula... Derive strategy parameters ,in This represents the current policy parameter. This represents the preset update rate. This is represented as the output value of the reward function. Represented as reward weight, Represented as a state-action value function, This is represented as the agent's state. This represents the number corresponding to the agent's state. , This represents the total number of agent states. This represents the action to be performed corresponding to the agent's state. Represented as the first The maximum output value of the reward function for each agent state.
[0018] Preferably, the output value of the reward function is calculated as follows: using the calculation formula... Obtain the output value of the reward function , This represents the actual output value. This is represented as input data. Represented as a consumption item, Represented as quantum relative entropy, , and These are respectively represented as the weighting factor corresponding to the actual output value, the weighting factor corresponding to the consumption term, and the weighting factor corresponding to the quantum relative entropy.
[0019] Preferably, the step of integrating the optimized control instruction set based on the updated strategy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from the hardware status check results to generate the optimized control instruction set includes: according to the calculation formula group Generate optimized control instructions and optimized control instruction set ,in This is expressed as a hardware evaluation coefficient. This is represented as the preset gain coefficient. This is represented as real-time environmental data.
[0020] Preferably, the step of calculating the evaluation index based on the control instruction set and the pre-acquired execution feedback data, and feeding it back to S3 and S4, includes: inputting the optimized control instruction set, driving the virtual actuator in the digital twin model, calculating the evaluation index, and acquiring the execution feedback data; updating the decay factor of the restricted communication data in S3 through gradient descent; and adjusting the reward weight in S4 to strengthen the quantum game strategy.
[0021] The beneficial effects of this application are as follows: 1. The communication-constrained multi-agent distributed optimization control method provided in this application initializes the distributed optimization model based on the pre-acquired local state data of each agent and the preset communication topology map, generates the initial control parameter set of each agent, and checks the hardware status of each agent; when the hardware status is normal, the parameters are integrated to generate preliminary control parameters; combined with the pre-acquired restricted communication data, the fusion information weight matrix is calculated through a function; combined with the pre-acquired system target data, the strategy parameters are calculated and the relative entropy constraint is updated; the optimization control instruction set is integrated to generate the optimization control instruction set; finally, based on the control instruction set and the pre-acquired execution feedback data, the evaluation index is calculated and the adjustment parameters are fed back; this solves the limitations existing in the current communication-constrained multi-agent optimization control process.
[0022] 2. This application achieves efficient local optimization through the gradient descent algorithm. Its innovation lies in integrating regularization constraints into the cost function, avoiding simple technical stacking. Instead of directly constraining the data, it smooths out the constraints through mathematical optimization, thereby improving the constraint effect on the data to a certain extent.
[0023] 3. This application integrates policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results in a triple dynamic manner. This provides a basis for optimization while suppressing hardware risks, ensuring implementation security, maintaining the availability of core functions, and improving overall fault tolerance. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a schematic diagram of the method flow of this application.
[0026] Figure 2 This is a flowchart of the data calculation process. Detailed Implementation
[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0028] Please see Figure 1 As shown, this application provides a communication-constrained multi-agent distributed optimization control method, including: S1, based on the pre-acquired local state data of each agent and the preset communication topology, initializing the distributed optimization model, generating the initial control parameter set of each agent, and performing hardware state checks on each agent.
[0029] In a specific example, generating the initial control parameter set for each agent includes: merging the data in the agent's local state data into a state vector, and denoting it as... , This is represented by the ID number corresponding to each agent. , This is represented by the total number of agents, and calculated according to the formula. Generate the initial control parameter set for each agent. , The initial control parameters for each agent are denoted as , where and These are represented by the historical mean and standard deviation of the state vector, respectively. Represented as a weight matrix of a preset communication topology. It is represented as a bias term in the weight matrix.
[0030] In a specific example, the hardware status check of each agent includes: checking the hardware status of each agent, generating hardware evaluation coefficients for each agent, and generating a global judgment flag based on the decision comparison of the hardware evaluation coefficients. When the hardware evaluation coefficient is greater than the hardware evaluation coefficient threshold, the global judgment flag is 1, and S2 is executed; otherwise, the global judgment flag is 0, and the fault handling stage is entered.
[0031] It should be noted that the hardware check is based on the pre-acquired communication restriction indication data and hardware status data of the intelligent agent. When the communication restriction indication data is greater than the communication restriction indication data threshold, the hardware check is triggered, and the sum of the products of each hardware status data and the corresponding weight factor is used as the hardware evaluation coefficient, and a global judgment flag is generated.
[0032] It should be noted that the communication limitation indication data is the packet loss rate; the hardware status data includes temperature, pressure, and flow rate.
[0033] S2. When the hardware is in normal condition, based on the initial control parameter set and the pre-acquired local constraint data, the parameters are integrated to generate preliminary control parameters.
[0034] It should be noted that the local constraint data are pre-acquired constants, such as maximum speed.
[0035] In a specific example, the step of generating preliminary control parameters by integrating the initial control parameter set and pre-acquired local constraint data includes: generating preliminary control parameters through calculation formulas. Generate preliminary control parameters , Let be the gradient of the cost function, where Represented as control parameters, This is represented as local optimization constraint data. Represented as the preset regularization coefficient, It is expressed as the square of the norm of the control parameters and the initial control parameters.
[0036] It should be noted that the squared norm of the control parameters and the initial control parameters is used to calculate the deviation between the control parameters and the initial control parameters, minimizing this part on the local constraint data; the regularization coefficient is used to process the constraint conditions, and is only applied when... It will take effect at that time.
[0037] It should be noted that, Represented as a maximum value function, when Outputting 0 at the time incurs no penalty. The output correction deviation is linearly penalized.
[0038] It should be noted that, , This is represented as obtaining the result through gradient descent iterations. smallest The iterative formula for gradient descent is: ,in Represented as the first The parameter values for the next iteration. It is expressed as the learning rate, and calculated using the formula: Derive the gradient of the cost function The descent gradient from the initial Start by iterating repeatedly until convergence, such as when the gradient approaches zero or the maximum number of iterations is reached.
[0039] It should be noted that the learning rate is used to control the update magnitude.
[0040] This application achieves efficient local optimization through the gradient descent algorithm. Its innovation lies in incorporating regularization constraints into the cost function, avoiding mere technical stacking. Instead of directly constraining the data, it smooths out the constraints through mathematical optimization, thereby improving the constraint effect on the data to a certain extent.
[0041] S3. Based on the preliminary control parameters and the pre-acquired restricted communication data, calculate the fusion information weight matrix using a function.
[0042] It should be noted that restricted communication data includes packet loss rate, etc. For example, the packet loss rate records 7 days of historical packet loss rate data, and is collected every minute.
[0043] In a specific example, the step of calculating the fused information weight matrix based on the preliminary control parameters and the pre-acquired restricted communication data includes: using a fractional attention mechanism to fuse the preliminary control parameters and the pre-acquired restricted communication data; extracting a query vector from the preliminary control parameter set and extracting key and value vectors from the restricted communication data; simultaneously calculating a decay factor and dynamically adjusting the weights corresponding to the restricted communication data based on the decay factor; using the product of the normalized decay factor and the query-key dot product as the attention score to generate weight matrix elements, and integrating and outputting them as the fused information weight matrix.
[0044] In a specific example, the step of using the product of the normalized decay factor and the query-key dot product as the attention score, generating weight matrix elements, and integrating and outputting them as a fused information weight matrix includes: denoting the query vector as... Restricted communication data is denoted as The key vector is denoted as The value vector is denoted as Through the calculation formula: Generate weight matrix elements and information weight matrix ,in This is represented as the attenuation factor for restricted communication data. Represented as query-key dot product, This is expressed as the dimension of the key vector. It is represented as a normalization function.
[0045] It should be noted that the query vector is represented as a local control parameter feature vector, with dimensions of [dimension number missing]. The dimension of the key vector is Used to scale the query-key dot product and prevent gradient explosion; the dimension of the value vector is... .
[0046] It should be noted that the Mittag-Leffler function is used to define the attenuation factor of constrained communication data, and the formula is as follows: ,in This is represented as the preset fractional order. Control the shape of the attenuation curve; This is represented as the preset attenuation coefficient. To control the attenuation intensity and reduce the impact on communication data; This is represented as the Mittag-Leffler function.
[0047] It should be noted that the attenuation factor replaces the hard threshold, improving the detection effect of communication fluctuations.
[0048] S4. Calculate the strategy parameters based on the information weight matrix and the pre-acquired system target data, and perform relative entropy constraint updates.
[0049] It should be noted that system target data, such as reference trajectories, etc.
[0050] In a specific example, the step of calculating strategy parameters based on the information weight matrix and pre-acquired system target data, and performing relative entropy constraint updates, includes: according to the calculation formula... Derive strategy parameters ,in This represents the current policy parameter. This represents the preset update rate. This is represented as the output value of the reward function. Represented as reward weight, Represented as a state-action value function, This is represented as the agent's state. This represents the number corresponding to the agent's state. , This represents the total number of agent states. This represents the action to be performed corresponding to the agent's state. Represented as the first The maximum output value of the reward function for each agent state.
[0051] It should be noted that the current strategy parameters represent the control strategy, such as PID gain, where PID stands for control algorithm principle; the preset update rate controls the update step size. A small update rate ensures stable convergence; reward weights control the adjustment range; the state-action value function evaluates the long-term benefits of performing actions in an agent state; agent states include things like position and velocity; the actions corresponding to agent states are control inputs, such as acceleration commands.
[0052] In a specific example, the output value of the reward function is calculated as follows: using the calculation formula... Obtain the output value of the reward function , This represents the actual output value. This is represented as input data. Represented as a consumption item, Represented as quantum relative entropy, , and These are respectively represented as the weighting factor corresponding to the actual output value, the weighting factor corresponding to the consumption term, and the weighting factor corresponding to the quantum relative entropy.
[0053] It should be noted that the actual output values include the current position or speed, etc.; the consumption items include the agent's energy consumption or wear and tear, etc., used to calculate the agent's losses. The weighting factor corresponding to the actual output value penalizes the output deviation, which is large. Enhance tracking accuracy; The weight factor corresponding to the consumption item penalizes high-consumption operations, while small-consumption operations are penalized. Allow high-consumption strategies; The weighting factor corresponding to quantum relative entropy controls the strength of the quantum consistency penalty. Reduce the impact of quantum confinement.
[0054] It should be noted that the formula for calculating quantum relative entropy is: ,in Represented as the agent's policy quantum state, Represented as the reference state of neighboring agents; it measures the difference between the agent's policy quantum state and the reference states of neighboring agents, with a larger value indicating higher policy inconsistency; where the calculation formula is used... The quantum distribution probability of the policies of neighboring agents is obtained, and then... Let be the reference state of the neighboring agent, where Represented as a quantum vector state; the elements of the weight matrix are taken as input data and mapped to a quantum state vector through quantum angle encoding, as shown in the formula: ,in Expressed as rotation angle, It is along A revolving door with an axle.
[0055] S5. Based on the updated policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results, optimize the control instruction set and generate an optimized control instruction set.
[0056] It should be noted that the pre-acquisition of real-time environmental data refers to the difference between the acquired environmental data and the standard value of the corresponding environmental data. The environmental data includes signal strength, obstacle distance, temperature, and humidity.
[0057] In a specific example, the process of integrating optimized control instruction sets based on updated policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results to generate optimized control instruction sets includes: calculating according to a set of formulas... Generate optimized control instructions and optimized control instruction set ,in This is expressed as a hardware evaluation coefficient. This is represented as the preset gain coefficient. This is represented as real-time environmental data.
[0058] It should be noted that the gain coefficient represents the default gain when there is no risk; real-time environmental data represents environmental constraints; and the optimized control command is positively correlated with the environmental constraints.
[0059] S6. Based on the control instruction set and the pre-acquired execution feedback data, calculate the evaluation index and feed it back to S3 and S4.
[0060] In a specific example, the step of calculating the evaluation index based on the control instruction set and the pre-acquired execution feedback data, and feeding it back to S3 and S4, includes: inputting the optimized control instruction set, driving the virtual actuator in the digital twin model, calculating the evaluation index, and acquiring the execution feedback data; updating the decay factor of the restricted communication data in S3 through gradient descent; and adjusting the reward weight in S4 to strengthen the quantum game strategy.
[0061] It should be noted that the execution feedback data represents the deviation between the actual system output and the expected output.
[0062] It should be noted that, according to the calculation formula Determine the evaluation indicators ,in This is represented as the feedback data error value. , The number corresponding to the control cycle. To control the total number of cycles.
[0063] It should be noted that the decay factor of the restricted communication data in S3 is updated using gradient descent, and the gradient descent formula is as follows: The updated attenuation factor is obtained. , Represented as the learning rate, it controls the update step size. It is expressed as the partial derivative of the performance index with respect to the attenuation factor, calculated through automatic differentiation.
[0064] It should be noted that the reward weights in S4 are adjusted to strengthen the quantum game strategy. The specific adjustment formula is as follows: Get updated reward weights ,in This is represented as an adjustment coefficient, used to control the adjustment range.
[0065] This application employs a triple dynamic integration of policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results. This provides a foundation for optimization while mitigating hardware risks, ensuring implementation security, maintaining the availability of core functions, and enhancing overall fault tolerance.
[0066] This application discloses a distributed optimization control method for communication-constrained multi-agent systems, relating to the field of multi-agent communication technology. The method includes: initializing a distributed optimization model based on pre-acquired local state data of each agent and a preset communication topology; generating initial control parameter sets for each agent and checking the hardware status of each agent; when the hardware status is normal, integrating parameters to generate preliminary control parameters; calculating a fusion information weight matrix using a function based on pre-acquired constrained communication data; calculating strategy parameters based on pre-acquired system target data and updating relative entropy constraints; integrating an optimized control instruction set to generate an optimized control instruction set; and finally calculating evaluation indicators based on the control instruction set and pre-acquired execution feedback data, and feeding back adjustment parameters. This method solves the limitations existing in current communication-constrained multi-agent optimization control processes.
[0067] The above content is merely an example and illustration of the concept of this application. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in this specification, they should all fall within the protection scope of this application.
Claims
1. A communication-constrained multi-agent distributed optimization control method, characterized in that, include: S1. Based on the pre-acquired local state data of each agent and the preset communication topology, initialize the distributed optimization model, generate the initial control parameter set of each agent, and perform hardware status checks on each agent. S2. When the hardware is in normal condition, based on the initial control parameter set and the pre-acquired local constraint data, the parameters are integrated to generate preliminary control parameters; S3. Based on the preliminary control parameters and the pre-acquired restricted communication data, calculate the fusion information weight matrix using a function; S4. Calculate the strategy parameters based on the information weight matrix and the pre-acquired system target data, and perform relative entropy constraint updates; S5. Based on the updated policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients in the hardware status check results, optimize the control instruction set and generate an optimized control instruction set. S6. Based on the set of control instructions and the pre-acquired execution feedback data, calculate the evaluation index and feed it back to S3 and S4.
2. The communication-constrained multi-agent distributed optimization control method according to claim 1, characterized in that, The initial control parameter set for each agent includes: The data in the agent's local state data are merged into a state vector, and denoted as... , This is represented by the ID number corresponding to each agent. , This is represented by the total number of agents, and calculated according to the formula. Generate the initial control parameter set for each agent. , The initial control parameters for each agent are denoted as , where and These are represented by the historical mean and standard deviation of the state vector, respectively. Represented as a weight matrix of a preset communication topology. It is represented as the bias term in the weight matrix.
3. The communication-constrained multi-agent distributed optimization control method according to claim 1, characterized in that, The hardware status check of each agent includes: The hardware status of each agent is checked, and hardware evaluation coefficients are generated for each agent. A global judgment flag is generated based on the decision comparison of the hardware evaluation coefficients. When the hardware evaluation coefficient is greater than the hardware evaluation coefficient threshold, the global judgment flag is 1 and S2 is executed; otherwise, the global judgment flag is 0 and the fault handling stage is entered.
4. The communication-constrained multi-agent distributed optimization control method according to claim 1, characterized in that, The step of integrating parameters based on the initial control parameter set and pre-acquired local constraint data to generate preliminary control parameters includes: Through calculation formula Generate preliminary control parameters , Let be the gradient of the cost function, where Represented as control parameters, This is represented as local optimization constraint data. Represented as the preset regularization coefficient, It is expressed as the square of the norm of the control parameters and the initial control parameters.
5. The communication-constrained multi-agent distributed optimization control method according to claim 1, characterized in that, The step of calculating the fusion information weight matrix based on the preliminary control parameters and pre-acquired restricted communication data includes: A fractional attention mechanism is used to fuse information from preliminary control parameters and pre-acquired restricted communication data. Query vectors are extracted from the preliminary control parameter set, and key vectors and value vectors are extracted from the restricted communication data. At the same time, a decay factor is calculated, and the weights corresponding to the restricted communication data are dynamically adjusted based on the decay factor. The product of the normalized decay factor and the query-key dot product is used as the attention score to generate weight matrix elements, and the results are integrated and output as a fused information weight matrix.
6. The communication-constrained multi-agent distributed optimization control method according to claim 5, characterized in that, The step of using the product of the normalized decay factor and the query-key dot product as the attention score to generate weight matrix elements, and integrating and outputting them as a fused information weight matrix, includes: Let the query vector be denoted as Restricted communication data is denoted as The key vector is denoted as The value vector is denoted as Through the calculation formula: Generate weight matrix elements and information weight matrix ,in This is represented as the attenuation factor for restricted communication data. Represented as query-key dot product, This is expressed as the dimension of the key vector. It is represented as a normalization function.
7. The communication-constrained multi-agent distributed optimization control method according to claim 1, characterized in that, The calculation of strategy parameters based on the information weight matrix and pre-acquired system target data, and the subsequent relative entropy constraint update, includes: According to the calculation formula Derive strategy parameters ,in This represents the current policy parameter. This represents the preset update rate. This is represented as the output value of the reward function. Represented as reward weight, Represented as a state-action value function, This is represented as the agent's state. This represents the number corresponding to the agent's state. , This represents the total number of agent states. This represents the action to be performed corresponding to the agent's state. Represented as the first The maximum output value of the reward function for each agent state.
8. The communication-constrained multi-agent distributed optimization control method according to claim 7, characterized in that, The specific calculation process for the output value of the reward function is as follows: Through calculation formula Obtain the output value of the reward function , This represents the actual output value. This is represented as input data. Represented as a consumption item, Represented as quantum relative entropy, , and These are respectively represented as the weighting factor corresponding to the actual output value, the weighting factor corresponding to the consumption term, and the weighting factor corresponding to the quantum relative entropy.
9. The communication-constrained multi-agent distributed optimization control method according to claim 8, characterized in that, The optimized control instruction set is generated by integrating the updated policy parameters, pre-acquired real-time environmental data, and hardware evaluation coefficients from hardware status check results, including: According to the calculation formula group Generate optimized control instructions and optimized control instruction set ,in This is expressed as a hardware evaluation coefficient. This is represented as the preset gain coefficient. This is represented as real-time environmental data.
10. The communication-constrained multi-agent distributed optimization control method according to claim 9, characterized in that, The step of calculating evaluation metrics based on the control instruction set and pre-acquired execution feedback data, and feeding them back to S3 and S4, includes: The system inputs an optimized control instruction set to drive the virtual actuator in the digital twin model, calculates evaluation metrics, and obtains execution feedback data. It then updates the decay factor of the restricted communication data in S3 using gradient descent and adjusts the reward weights in S4 to strengthen the quantum game strategy.
Citation Information
Patent Citations
Water quality supervision method and system for rural sewage treatment facility
CN116862417A