Multi-objective particle swarm optimization parameter adjustment method, system, device and storage medium

Abstract

Embodiments of the present application provide a multi-objective particle swarm optimization parameter adjustment method, system, device and storage medium. The method comprises: for a current multi-objective optimization problem to be solved, extracting a state vector; inputting the state vector into a trained parameter adjustment module to output an adjustment result of an optimization control parameter; based on the adjustment result of the optimization control parameter, determining a current candidate solution group of the current multi-objective optimization problem to be solved; based on the current candidate solution group, iteratively updating a non-dominated solution set of the current multi-objective optimization problem to be solved until a latest non-dominated solution set is output after a preset termination condition is met. Thus, the optimization control parameter can be dynamically adjusted according to the actual state of the current search process, breaking free from the limitation of fixed optimization strategy, effectively improving the adaptation ability and optimization flexibility of the multi-objective particle swarm optimization algorithm to complex multi-objective problems, and ensuring the dynamic adaptability of the iterative optimization process.

Description

Technical Field

[0001] This application relates to the field of intelligent optimization algorithm technology, specifically to a method for adjusting parameters in multi-objective particle swarm optimization, a system for adjusting parameters in multi-objective particle swarm optimization, an electronic device, and a non-volatile storage medium. Background Technology

[0002] Multi-objective optimization is a technique for solving complex problems with multiple conflicting and mutually constraining optimization objectives. Its goal is to search for a set of compromise optimal solutions that balance the various optimization objectives under multiple constraints. Currently, mainstream algorithms for solving multi-objective optimization problems include various swarm intelligence optimization algorithms, such as multi-objective particle swarm optimization, multi-objective gray wolf optimization, and multi-objective ant colony optimization.

[0003] The aforementioned algorithm primarily relies on a population-based iterative optimization mechanism to solve the problem. The search behavior, evolutionary rhythm, and update method of the frontier solution set determine the overall optimization performance of the algorithm. Existing mainstream multi-objective optimization algorithms have rigid and singular search evolution logics, resulting in limited environmental adaptability and optimization robustness. They struggle to achieve a dynamic balance between global exploration and local mining, and are prone to problems such as insufficient population search vitality and inadequate frontier solution set updates during iteration. Consequently, the overall quality of the resulting non-dominated solution sets varies significantly, making it difficult to simultaneously meet the multiple requirements of convergence accuracy, distribution diversity, and solution stability, and thus unsuitable for complex and high-precision engineering optimization scenarios. Summary of the Invention

[0004] This application is made in consideration of the above-mentioned problems. According to a first aspect of this application, a method for adjusting parameters in multi-objective particle swarm optimization is provided, comprising:

[0005] For the current multi-objective optimization problem to be solved, a state vector is extracted, which is used to represent the multi-objective search process;

[0006] The state vector is input into the trained parameter tuning module to output the tuning result of the optimization control parameters, where the optimization control parameters represent the search control parameters used to solve the current multi-objective optimization problem.

[0007] Based on the adjustment results of the optimization control parameters, the current candidate solution group for the multi-objective optimization problem to be solved is determined;

[0008] Based on the current candidate solution group, iteratively update the non-dominated solution set of the current multi-objective optimization problem to be solved until the preset termination condition is met, and then output the latest non-dominated solution set.

[0009] For example, the parameter tuning module is implemented using the soft actor-critic algorithm; the training process of the parameter tuning module includes one or more training rounds, and each training round includes multiple iterations of the multi-objective optimization algorithm;

[0010] The training steps for the parameter tuning module include:

[0011] For each training round, the reward value for the current iteration is calculated based on the quality change of the non-dominated solution set of the current multi-objective optimization problem before and after the update.

[0012] Based on the state transition samples of each iteration cycle, the network parameters of the parameter adjustment module are iteratively updated until the training termination condition is met. The state transition samples are generated based on the reward value of the corresponding iteration cycle.

[0013] For example, the reward value is calculated using the following formula: ,in, Indicates the number of training rounds in the current round. The reward value per iteration cycle. Indicates the number of training rounds in the current round. The change in distance characteristics between the non-dominated solution set and the theoretical optimal solution set of the current multi-objective optimization problem over several iterations. Indicates the number of training rounds in the current round. The change in the distribution characteristics of all non-dominated solutions in the set of non-dominated solutions over one iteration period. Indicates the number of training rounds in the current round. The change in the quantity of all non-dominated solutions in the non-dominated solution set over one iteration cycle. , as well as All are weighting coefficients.

[0014] For example, the state vector includes particle swarm state information, non-dominated solution set state information, and current iteration progress information. The particle swarm state information includes particle position distribution characteristics, and the non-dominated solution set state information includes at least one of the following: non-dominated solution set size information, non-dominated solution set front distribution span information, and distance information between the particle swarm and the non-dominated solution set.

[0015] For example, based on the adjustment results of the optimization control parameters, the current candidate solution group for the multi-objective optimization problem to be solved is determined, including:

[0016] When the adjustment result of the optimization control parameter is expressed as an incremental parameter, the optimization control parameter for the current iteration period is determined based on the adjustment result and optimization control parameter of the previous iteration period.

[0017] Based on the updated optimization control parameters, the particle positions of each particle in the particle swarm are updated, and the updated particle positions are used as at least a part of the current candidate solution group.

[0018] For example, based on the updated optimization control parameters, the particle positions of each particle in the particle swarm are updated, including:

[0019] Based on the updated optimization control parameters, the velocity and position of each particle in the particle swarm in the previous iteration cycle, the global guiding solution in the updated non-dominated solution set of the previous iteration cycle, and the optimal position of each individual particle, the update velocity of each particle in the particle swarm is calculated; wherein, the global guiding solution is determined based on a preset guiding strategy.

[0020] Based on the update rate, the particle positions of each particle in their respective particle swarm are updated.

[0021] For example, before updating the particle positions of each particle in the particle swarm, the method further includes:

[0022] Boundary constraints are applied to the updated optimization control parameters to limit them to a preset range.

[0023] According to a second aspect of this application, a multi-objective particle swarm optimization parameter tuning system is also provided, comprising:

[0024] The state extraction module is used to extract the state vector for the current multi-objective optimization problem to be solved, where the state vector is used to represent the multi-objective search process;

[0025] The trained parameter tuning module is used to receive the state vector and output the tuning results of the optimization control parameters, where the optimization control parameters represent the search control parameters used to solve the current multi-objective optimization problem.

[0026] The solution determination module is used to determine the current candidate solution group for the multi-objective optimization problem to be solved based on the adjustment results of the optimization control parameters.

[0027] The solution set update module is used to iteratively update the non-dominated solution set of the current multi-objective optimization problem based on the current candidate solution group, until the preset termination condition is met and the latest non-dominated solution set is output.

[0028] According to a third aspect of this application, an electronic device is also provided, including a processor and a memory, wherein the memory stores computer program instructions, which are executed by the processor to perform the multi-objective particle swarm optimization parameter adjustment method as described above.

[0029] According to a fourth aspect of this application, a non-volatile storage medium is also provided, on which program instructions are stored, which, when executed, are used to perform the multi-objective particle swarm optimization parameter adjustment method as described above.

[0030] According to the above technical solution, a state vector representing the multi-objective search process is extracted and input into a trained parameter adjustment module to output the adjustment result of the optimization control parameters. Based on this adjustment result, candidate solutions are determined and the non-dominated solution set is iteratively updated. This allows the optimization control parameters to be dynamically adjusted according to the actual state of the current search process, overcoming the limitations of fixed optimization strategies and effectively improving the adaptability and optimization flexibility of the multi-objective particle swarm optimization algorithm for complex multi-objective problems, ensuring the dynamic adaptability of the iterative optimization process. Simultaneously, the above solution continuously optimizes the non-dominated solution set through an iterative update mechanism, reducing reliance on manually preset parameter adjustment experience and effectively improving the adaptability of parameter adjustment and the quality of the solution set during multi-objective optimization.

[0031] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0032] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof.

[0033] Figure 1 A schematic flowchart of a multi-objective particle swarm optimization parameter adjustment method according to an embodiment of this application is shown;

[0034] Figure 2 A schematic diagram of a parameter adaptive adjustment structure combining a parameter adjustment module with a multi-objective optimization algorithm according to an embodiment of this application is shown.

[0035] Figure 3 This illustration shows a schematic diagram of the offline training and online deployment process of a parameter adjustment module and a multi-objective optimization algorithm according to an embodiment of this application;

[0036] Figure 4 A schematic block diagram of a multi-objective particle swarm optimization parameter adjustment system according to an embodiment of this application is shown;

[0037] Figure 5A schematic block diagram of an electronic device according to one embodiment of this application is shown. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application more apparent, exemplary embodiments according to this application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein. Based on the embodiments of this application described herein, all other embodiments obtained by those skilled in the art without inventive effort should fall within the protection scope of this application.

[0039] As mentioned earlier, the search evolution logic of existing mainstream multi-objective optimization algorithms is rigid and singular. They generally employ fixed parameter configurations, pre-defined change patterns, or parameter adjustment strategies based on human experience, making it difficult to achieve adaptive dynamic adjustments based on real-time search progress, population distribution, and dynamic changes in the non-dominated front solution set. Parameter matching largely relies on manual debugging experience, making it difficult to adapt to the optimization needs of different iteration stages. This static or simplistic parameter adjustment mode leads to poor stability in the optimization process of multi-objective particle swarm optimization algorithms, easily resulting in problems such as premature decay of population diversity, insufficient front convergence accuracy of the Pareto archive used to store the non-dominated solution set, and poor uniformity of the non-dominated solution set distribution. Ultimately, the overall quality of the output non-dominated solution set is low, making it difficult to balance solution set convergence, diversity, and overall quality, significantly limiting the application effectiveness of multi-objective particle swarm optimization algorithms in high-precision, high-complexity multi-objective engineering optimization scenarios.

[0040] To at least partially solve the above-mentioned technical problems, an embodiment of this application provides a method for adjusting parameters in multi-objective particle swarm optimization. Figure 1 A schematic flowchart of a multi-objective particle swarm optimization parameter adjustment method 100 according to an embodiment of this application is shown. Figure 1 As shown, method 100 may include, but is not limited to, the following steps S110-S140.

[0041] Step S110: For the current multi-objective optimization problem to be solved, extract the state vector, whereby the state vector is used to represent the multi-objective search process.

[0042] Optionally, the multi-objective optimization problem to be solved can be constructed first. The optimization objective, the range of values ​​for optimization variables, and various constraints can be defined according to the actual engineering optimization needs, thus fully constructing a multi-objective optimization mathematical model adapted to the current solution scenario. Here, F(x) can represent the objective function vector, which... Each component in Each can represent a target function to be optimized. It can represent the target quantity. The value of can be 2, 3, or a positive integer greater than 3. When When this multi-objective optimization problem can contain two objective functions; when In this context, the multi-objective optimization problem can contain three objective functions. Specifically, the decision variables of the current multi-objective optimization problem to be solved can be determined, and these decision variables can constitute a candidate solution vector. Multiple objective functions can be determined according to the actual application requirements. The constraints of the multi-objective optimization problem can also be determined. The dimension of the candidate solution vector can be determined by the number of decision variables. For example, the candidate solution vector can be represented as... .in, It can represent the number of iterations within the t-th iteration period. The position of each particle, D, can represent the number of decision variables. The number of objective functions can be two, three, or more. Objective functions can conflict with each other. Each objective function can be set to a minimization form, a maximization form, or unified into a minimization form after normalization or sign transformation, depending on the specific application scenario. Constraints can include equality constraints, inequality constraints, or both. This multi-objective optimization problem can originate from microgrid capacity planning scenarios, integrated energy system planning scenarios, energy storage capacity configuration scenarios, or other multi-objective optimization application scenarios.

[0043] Based on this, the fundamental data for the multi-objective particle swarm optimization algorithm can be initialized. Specifically, this includes initializing the particle population size, initial position information of each particle, initial velocity information, initial optimization control parameters, and an empty non-dominated solution set. Specifically, the population can be initialized, containing a preset number of individuals. Each individual can represent a candidate solution to the multi-objective optimization problem. The position of each individual can be initialized randomly or based on prior knowledge. The velocity of each individual can be initialized, with initial values ​​set to a zero vector or a random vector. The optimal position of each individual can be initialized, with initial values ​​that can be the individual's own initial position. The non-dominated solution set can be initialized, initially as an empty set or as a set containing all non-dominated individuals in the population. This non-dominated solution set can be stored in an external Pareto archive or in other data structures. An upper limit on the capacity of the non-dominated solution set can be set. Algorithm operating parameters such as population size, maximum number of iterations, and convergence threshold can be set.

[0044] Subsequently, for the multi-objective optimization problem currently undergoing iterative optimization, the corresponding state vector can be extracted in real time. This state vector represents the overall search progress of the multi-objective optimization algorithm and reflects its real-time optimization status. It can be understood that the dimensions and specific composition of the state vector can be flexibly set according to the actual optimization scenario and optimization requirements. The multi-objective optimization problem may include at least two conflicting objective functions, or it may include multiple decision variables and corresponding constraints. This state vector can be extracted in each iteration cycle, or it can be extracted after every preset number of iterations.

[0045] Step S120: Input the state vector into the trained parameter adjustment module to output the adjustment result of the optimization control parameters, wherein the optimization control parameters represent the search control parameters used to solve the current multi-objective optimization problem.

[0046] The extracted state vector can be input into a pre-trained parameter tuning module. The parameter tuning module can employ a neural network model, a reinforcement learning model, or other function fitting structures with parameter mapping capabilities to achieve the corresponding function. The training logic and network structure within the parameter tuning module can be configured independently according to the actual application scenario. After receiving the state vector, the parameter tuning module can output the adjustment result of the corresponding optimization control parameters. These optimization control parameters can be various search control parameters used within the current iteration cycle to drive the multi-objective particle swarm optimization algorithm to iteratively optimize the multi-objective optimization problem to be solved. The specific type of parameter can be selected as needed. Optionally, the output of the parameter tuning module can be the adjustment increment of the optimization control parameter or the direct value of the optimization control parameter. When using the multi-objective particle swarm optimization (MOPSO) algorithm, the corresponding optimization control parameters can include at least one of inertia weight parameters, individual learning factors, and social learning factors, or other algorithm parameters that control the search behavior of the population. The adjustment result can be a numerical vector or a numerical scalar. The parameter adjustment module can output the adjustment results in each iteration cycle, or it can output the adjustment results at a preset interval of iteration cycles.

[0047] Step S130: Based on the adjustment results of the optimization control parameters, determine the current candidate solution group for the current multi-objective optimization problem to be solved.

[0048] Based on the adjustment results of the optimization control parameters output by the parameter adjustment module, the current optimization control parameters of the algorithm can be updated and adjusted accordingly. Then, the multi-objective particle swarm optimization algorithm is run based on the updated optimization control parameters to iteratively generate the current candidate solution group corresponding to the current multi-objective optimization problem to be solved. The method of generating candidate solutions can be flexibly implemented by combining the basic iterative rules of the multi-objective particle swarm optimization algorithm.

[0049] Step S140: Based on the current candidate solution group, iteratively update the non-dominated solution set of the current multi-objective optimization problem to be solved until the preset termination condition is met, and then output the latest non-dominated solution set.

[0050] Based on the current candidate solution group generated in the aforementioned steps, the non-dominated solution set corresponding to the current multi-objective optimization problem can be iteratively updated according to the non-dominated selection rules of multi-objective optimization. The iterative steps of extracting the state vector, outputting parameter adjustment results, determining candidate solutions, and updating the non-dominated solution set are continuously executed in a loop. An iteration termination condition can be preset; when the iteration process meets the preset termination condition, the iteration loop stops, and the latest non-dominated solution set is output as the final optimization result. It can be understood that the preset termination condition can serve as the termination condition for solving the current multi-objective optimization problem. Specifically, when updating the non-dominated solution set based on the current candidate solution group, the objective function value of each candidate solution in the current candidate solution group can be calculated separately, and the dominance relationship of each candidate solution can be compared with the solutions in the existing non-dominated solution set. Candidate solutions not dominated by existing non-dominated solutions can be added to the non-dominated solution set. Existing non-dominated solutions dominated by candidate solutions can be removed from the non-dominated solution set. Candidate solutions dominated by existing non-dominated solutions can be discarded. When the number of solutions in the non-dominated solution set exceeds a preset capacity limit, solutions in the non-dominated solution set can be filtered according to a preset screening strategy to control the size of the non-dominated solution set. Preset termination conditions may include reaching a preset maximum number of iterations, the non-dominated solution set remaining unchanged for a preset number of consecutive iterations, the improvement in the quality evaluation index of the non-dominated solution set being less than a preset threshold, or any combination of the above conditions. If the preset termination conditions are not met, the iteration can return to step S110 to continue. If the preset termination conditions are met, the current non-dominated solution set can be output as the final optimization result.

[0051] Alternatively, the preset termination condition can also be used as the training termination condition for the current training round during the offline training phase of the parameter adjustment module.

[0052] According to the above technical solution, a state vector representing the multi-objective search process is extracted and input into a trained parameter adjustment module to output the adjustment result of the optimization control parameters. Based on this adjustment result, candidate solutions are determined and the non-dominated solution set is iteratively updated. This allows the optimization control parameters to be dynamically adjusted according to the actual state of the current search process, overcoming the limitations of fixed optimization strategies and effectively improving the adaptability and optimization flexibility of the multi-objective particle swarm optimization algorithm for complex multi-objective problems, ensuring the dynamic adaptability of the iterative optimization process. Simultaneously, the above solution continuously optimizes the non-dominated solution set through an iterative update mechanism, reducing reliance on manually preset parameter adjustment experience and effectively improving the adaptability of parameter adjustment and the quality of the solution set during multi-objective optimization.

[0053] For example, the state vector may include particle swarm state information, non-dominated solution set state information, and current iteration progress information. The particle swarm state information may include particle position distribution characteristics. The non-dominated solution set state information may include at least one of the following: non-dominated solution set size information, non-dominated solution set front distribution span information, and distance information between the particle swarm and the non-dominated solution set.

[0054] Specifically, the state vector can contain information in three dimensions: particle swarm state information, non-dominated solution set state information, and current iteration progress information. The data normalization method and information fusion weights for these three types of information can be flexibly configured according to the actual solution scenario, and the data collection frequency for each dimension can be consistent with the algorithm iteration cycle.

[0055] The particle swarm state information can include particle position distribution characteristics. These characteristics can be obtained by statistically analyzing data such as the coordinate dispersion, distribution range, and clustering degree of all particles within the swarm in the optimization variable space. The optimization variable space can be understood as the multi-dimensional numerical space formed by all adjustable optimization variables within a preset value range in the multi-objective optimization problem to be solved; it is the effective solution space for particle positions and iterative search. Specifically, it can be obtained by calculating the standard deviation of all individual positions in the particle swarm, the average distance between all pairwise positions, or the average offset of all individual positions relative to the swarm's centroid. The specific calculation dimension of the particle position distribution characteristics can be adaptively matched to the variable dimensions of the multi-objective optimization problem to be solved.

[0056] The state information of the non-dominated solution set can be selected from at least one of the following: the size of the non-dominated solution set, the front distribution span of the non-dominated solution set, and the distance between the particle swarm and the non-dominated solution set. Alternatively, two or more types can be combined. The specific combination can be flexibly adjusted according to the solution requirements before inputting into the parameter tuning module. For the same trained parameter tuning module, the input dimension, feature composition, and arrangement order of the state vector should remain consistent during the training and online deployment phases. The size information of the non-dominated solution set can be the actual number of non-dominated solutions in the set, or the normalized value of that actual number. This normalization can be based on the upper limit of the capacity of the non-dominated solution set or on the historical maximum size. The front distribution span information of the non-dominated solution set can be the distribution range of the non-dominated solutions in the target space. The target space can refer to a multi-dimensional evaluation space composed of the various optimization objectives of the multi-objective optimization problem, used to quantitatively evaluate the optimization performance of each set of candidate solutions. For example, this information can be obtained by calculating the difference between the maximum and minimum values ​​of each objective function in the non-dominated solution set, by calculating the Euclidean distance between the extreme points of each objective function in the non-dominated solution set, or by calculating the hypervolume covered by the non-dominated solution set. The distance information between the particle swarm and the non-dominated solution set can be obtained by calculating the distance between the centroid of the particle swarm and the centroid of the non-dominated solution set, by calculating the average distance between each particle and its nearest non-dominated solution, or by calculating the average distance between each non-dominated solution and its nearest particle.

[0057] The current iteration progress information can be the ratio of the current iteration count to the preset maximum iteration count, or it can be the absolute value of the current iteration count. The current iteration count can be represented using the current iteration cycle number. The current iteration progress information can also be the ratio of the current number of objective function evaluations to the preset maximum number of evaluations, which can be used to mark the stage of the algorithm's overall optimization process. The quantification form of the iteration progress information can be set as needed.

[0058] Optionally, the state vector may also include parameter adjustment action information determined in the previous iteration cycle. For example, it may include the output inertia weight increment, individual learning factor increment, and social learning factor increment from the previous iteration cycle. Setting parameter adjustment action information determined in the previous iteration cycle in the state vector can enhance the temporal continuity of the optimization control parameter adjustment.

[0059] In a specific embodiment, the state vector can be represented as: .in, It can represent the first The state vector extracted in each iteration cycle It can represent the characteristic information of particle position distribution. It can represent the size information of the non-dominated solution set. It can represent the distance information between the particle swarm and the non-dominated solution set. It can represent the frontier distribution span information of the non-dominated solution set. It can represent the current iteration progress information. It can represent the parameter adjustment action information determined in the previous iteration cycle.

[0060] It is understandable that the components in the above state vector can be normalized and then input into the parameter adjustment module to reduce the impact of the differences in the dimensions of each component on the learning process, which is beneficial to improving the learning efficiency and output stability of the parameter adjustment module.

[0061] After extracting the multi-dimensional state vectors, the state vectors can be input into a pre-trained parameter adjustment module. This module combines the fused multi-dimensional state information to output the adjustment results of the optimization control parameters for the current iteration cycle. Subsequently, the optimization control parameters can be updated based on the parameter adjustment results, candidate solutions can be generated, and the non-dominated solution set can be iteratively updated. This process is repeated until a preset termination condition is met, at which point the final non-dominated solution set is output.

[0062] The aforementioned scheme specifically defines the dimensions of the state vector used to characterize the multi-objective search process, integrating three core types of information: particle swarm distribution state, non-dominated solution set state, and iteration progress. This allows for precise matching of the optimization characteristics of the multi-objective particle swarm optimization algorithm. This multi-dimensional state vector provides complete and comprehensive feedback on the algorithm's real-time search state, offering accurate and comprehensive input for the parameter adjustment module's adaptive parameter control, thus ensuring the alignment between parameter adjustment and the algorithm's optimization process.

[0063] For example, the parameter adjustment module can be implemented using the Soft Actor-Critic (SAC) algorithm. The SAC algorithm can be a reinforcement learning algorithm applicable to continuous action spaces. It can include a policy network. The policy network is used to optimize the adjustment results of control parameters based on the input state vector. The SAC algorithm can also include a value network and a temperature parameter. The value network is used to evaluate the value of state-action pairs. The temperature parameter is used to adjust the balance between exploration and exploitation. The policy network, value network, and temperature parameter can all be iteratively updated during training. Optionally, the policy network can be a deterministic policy network or a stochastic policy network. When using a stochastic policy network, its output can be the mean and variance of the action distribution, or the adjustment result can be directly sampled from the action distribution. The value network can include one or two value networks. When using two value networks, the smaller value of the two value network outputs can be taken as the target value, or the average value of the two value network outputs can be taken as the target value. The network structure of the soft actor-critic algorithm can include fully connected layers, as well as convolutional layers or recurrent neural network layers.

[0064] The training process of the parameter tuning module can include one or more training rounds, and each training round can include multiple iterations of the multi-objective optimization algorithm. At the beginning of each training round, a multi-objective particle swarm optimization environment can be initialized for the same multi-objective optimization problem to be solved. Each iteration can represent one interaction between the parameter tuning module and the multi-objective particle swarm optimization environment.

[0065] In one embodiment, step S210, the training parameter adjustment module, may include, but is not limited to, the following steps S211 and S212.

[0066] Step S211: For each training round, calculate the reward value for the current iteration cycle based on the quality change of the non-dominated solution set of the multi-objective optimization problem in the training samples before and after the update.

[0067] For example, the training samples can be the aforementioned multi-objective optimization problem to be solved. Its objective function, decision variables, and constraints can be determined before training in the parameter tuning module. At the beginning of each training round, the particle swarm, particle velocity, optimal position of individual particles, optimization control parameters, and non-dominated solution set of the multi-objective particle swarm optimization algorithm can be reinitialized for the current multi-objective optimization problem to be solved.

[0068] In each iteration cycle, the current state vector can be extracted and input into the parameter adjustment module to output the adjustment result of the optimization control parameters. Subsequently, the optimization control parameters can be updated based on the adjustment result, and a multi-objective particle swarm optimization iteration can be performed to update the particle swarm and non-dominated solution set. After completing the current iteration cycle, the next state vector can be extracted. Simultaneously, the state data and solution set data before and after the update of the non-dominated solution set in the current iteration cycle can be recorded.

[0069] During training, the reward value for the current iteration cycle can be quantitatively calculated based on the overall quality change of the non-dominated solution set before and after updating within the current training round. The quantification rules for the reward value can be flexibly set according to the quality evaluation criteria of the multi-objective optimization solution set. Specifically, in each iteration cycle of the current training round, the first state of the non-dominated solution set can be recorded before updating it, and the second state can be recorded after updating it. Based on the first and second states, the quality change of the non-dominated solution set before and after the update can be calculated. The reward value for the current iteration cycle can then be calculated based on the quality change. The reward value can be positively correlated with the quality change, or it can be determined based on the relationship between the quality change and a preset threshold. A positive reward value is given when the quality change exceeds the preset threshold. A negative or zero reward value is given when the quality change does not exceed the preset threshold. The reward value can be calculated after each iteration cycle, or after a preset number of iteration cycles.

[0070] Step S212: Based on the state transition samples of each iteration period, iteratively update the network parameters of the parameter adjustment module until the training termination condition is met. The state transition samples are generated based on the reward value of the corresponding iteration period.

[0071] Based on the reward value of the current iteration cycle, state transition samples for the current iteration cycle can be generated. Specifically, the state vector of the current iteration cycle can be generated first. Adjustment results Reward Value and the next state vector extracted after the current iteration cycle. A state transition sample is stored in the experience replay pool. Then, a batch of state transition samples is randomly sampled from this pool, and the network parameters of the parameter adjustment module are updated based on this batch of state transition samples. It can be understood that each state transition sample can be represented by the following four-tuple form: The update objective of the policy network can be to maximize the sum of the expected cumulative reward and the entropy regularization term. The update objective of the value network can be to minimize the mean squared error between the predicted value and the target value. The update objective of the temperature parameter can be to keep the policy entropy near a preset target entropy. These update operations can be performed after each iteration of each training round, after each preset number of iterations, or after each training round. Multiple training rounds can be repeated to continuously iterate and optimize the network parameters of the parameter adjustment module until a preset training termination condition is met, at which point the training process stops, and the trained parameter adjustment module is solidified. The training termination condition can be any one or more combinations of a preset number of training rounds, network parameter convergence state, and stable solution set quality; specific termination conditions can be configured as needed.

[0072] In one specific embodiment, when the termination condition for the current training round is met, the current training round can be terminated, and the multi-objective particle swarm optimization environment for the next training round can be reinitialized. Multiple training rounds can be repeated until the overall training termination condition is met, at which point training stops, and the trained parameter adjustment module is fixed. The termination condition for the current training round may include reaching a preset maximum number of iterations or meeting a preset convergence condition. The overall training termination condition may include one or more of the following: reaching a preset number of training rounds, network parameters tending to converge, or training performance tending to stabilize.

[0073] The above scheme employs a soft actor-critic algorithm to construct the parameter tuning module, and calculates training rewards and iteratively optimizes the module's network parameters based on the quality changes of the non-dominated solution set. By using interactive reinforcement learning training, the parameter tuning module achieves autonomous iterative optimization, enabling it to adapt to the optimization characteristics of multi-objective optimization problems. This achieves autonomous learning and updating of the optimization parameter adjustment strategy, improving the intelligence level of parameter adaptive adjustment.

[0074] For example, the reward value can be calculated using the following formula: . It can represent the number of training rounds in the current training round. The reward value for each iteration cycle. It can represent the number of training rounds in the current training round. The change in distance characteristics between the non-dominated solution set and the theoretical optimal solution set of the multi-objective optimization problem in the training samples over several iterations. It can represent the number of training rounds in the current training round. The change in the distribution characteristics of all non-dominated solutions in the non-dominated solution set over a set of iterations. It can represent the number of training rounds in the current training round. The change in the quantity of all non-dominated solutions in the non-dominated solution set over a set of iterations. , as well as All are weighting coefficients.

[0075] In the current training round of the parameter tuning module, the first... Within each iteration cycle, after completing the iterative optimization of the multi-objective optimization problem and updating the non-dominated solution set for the current iteration cycle, the reward value for the current iteration cycle can be quantified using a preset reward calculation formula. The specific reward calculation formula can be: . It can represent the number of training rounds in the current training round. The total reward value corresponding to each iteration cycle is used to represent the feedback benefit corresponding to the overall optimization in the current iteration cycle. , as well as These can be weight coefficients for the corresponding reward components. The specific values ​​of each weight coefficient can be flexibly configured according to training needs, and the range of values ​​for each weight coefficient can be predetermined.

[0076] It can represent the number of training rounds in the current training round. The convergent reward component of the nth iteration. Specifically, it can be represented as the nth iteration reward component within the current training round. The change in distance characteristics between the non-dominated solution set and the theoretical optimal solution set of the multi-objective optimization problem in the training samples over a given iteration cycle. This distance characteristic can be selected from any of the average distance, minimum distance, and maximum distance. The calculation method of the distance characteristic can be set as needed, and the change can be calculated from the difference in distance characteristics between adjacent iteration cycles.

[0077] It can represent the number of training rounds in the current training round. The diversity reward component of each iteration cycle. Specifically, it can represent the reward component of the current training round. The change in the distribution characteristics of all non-dominated solutions in the non-dominated solution set over a set of iterations. The distribution characteristics may include any one or more combinations of the uniformity of solution distribution, the span of the target space distribution, and the degree of dispersion. The calculation rules for the distribution characteristics can be flexibly set according to the multi-objective solution set evaluation criteria, and the change can be calculated by the difference in distribution characteristics between adjacent iterations.

[0078] It can represent the number of training rounds in the current training round. The quantity and quality reward component of each iteration cycle can be specifically represented as the number of iterations within the current training round. The change in the quantity characteristic of all non-dominated solutions in the non-dominated solution set over a set of iterations. The quantity characteristic can be the total number of valid solutions in the non-dominated solution set, and the change in the quantity characteristic can be calculated by the difference in the number of solutions in adjacent iterations.

[0079] In one specific embodiment, after the reward value is calculated in each iteration cycle, the multidimensional reward value, which includes convergence, diversity, quantity and quality, can be used as a feedback signal to iteratively update the network parameters of the parameter adjustment module. This process is repeated for multiple training rounds until the preset training termination condition is met, thus completing the training and solidification of the parameter adjustment module.

[0080] Optionally, the three reward components mentioned above can be calculated and represented using a threshold-based binarized form. For example... .in, It can represent the first The amount of improvement in each quality indicator. This can represent the minimum effective improvement threshold. A positive reward is given when the improvement is greater than the minimum effective improvement threshold. A zero or negative reward is given when the improvement is less than or equal to the minimum effective improvement threshold.

[0081] In the above scheme, the reward value is decomposed into three dimensions: convergence, diversity, and solution set quantity. This allows the reward value to comprehensively reflect different aspects of improvement in the multi-objective optimization process. Simultaneously, by integrating information from each dimension through a weighted summation method, the scheme fully aligns with the solution set evaluation characteristics of multi-objective optimization problems. This provides accurate and multi-dimensional quantitative feedback for the iterative training of the parameter adjustment module, ensuring that the module's training direction matches the core needs of multi-objective optimization.

[0082] For example, step S130, based on the adjustment results of the optimization control parameters, determines the current candidate solution group of the current multi-objective optimization problem to be solved, which may include, but is not limited to, the following steps S131 and S132.

[0083] Step S131: For cases where the adjustment result of the optimization control parameter is expressed as an incremental parameter, determine the optimization control parameter for the current iteration based on the adjustment result and optimization control parameter of the previous iteration cycle.

[0084] Optionally, within each iteration cycle, the optimization control parameter adjustment result output by the parameter adjustment module can be an incremental parameter. This incremental parameter can be a correction offset relative to the original optimization control parameters within its respective iteration cycle, i.e., the aforementioned adjustment increment. The sign and magnitude of this adjustment increment can be dynamically determined based on the output result of the parameter adjustment module. Based on the adjustment result and optimization control parameters of the previous iteration cycle, the optimization control parameters for the current iteration cycle can be calculated. Specifically, the correction offset of the previous iteration cycle can be directly superimposed with the optimization control parameters of the previous iteration cycle. It can be understood that the sign of the correction offset is considered during the superposition process, thereby obtaining the optimization control parameters for the current iteration cycle. For example, the optimization control parameters of the previous iteration cycle may include the inertia weight parameter as described above. Individual learning factors Social learning factors Correspondingly, the adjustment result may include an increment in inertia weight. Individual learning factor increment social learning factor increment The result of this incremental adjustment can be expressed as: When the adjustment result includes an inertia weight increment, this inertia weight increment can be superimposed on the inertia weight parameter of the previous iteration cycle, that is, The inertia weight parameters for the current iteration period are obtained. When the adjustment result includes an increase in the individual learning factor, this increase in the individual learning factor can be superimposed on the individual learning factor from the previous iteration. This allows us to obtain the individual learning factor for the current iteration cycle. When the adjustment result includes an increment in the social learning factor, this increment can be superimposed on the social learning factor from the previous iteration cycle. This allows us to obtain the social learning factor for the current iteration cycle. .

[0085] Step S132: Based on the updated optimization control parameters, update the particle positions of each particle in the particle swarm, and use the updated particle positions as at least a part of the current candidate solution group.

[0086] After incrementally updating the optimization control parameters, the particle positions of each particle in the particle swarm can be iteratively updated according to the particle iteration rules of the multi-objective particle swarm optimization algorithm, based on the updated optimization control parameters. The specific calculation formula for particle position updating can be flexibly selected in conjunction with the basic particle swarm algorithm rules, and the update process can be synchronously adapted to the updated optimization control parameters. Then, all particle positions updated in this iteration cycle can be used as candidate solutions to form the current candidate solution group for the multi-objective optimization problem to be solved, completing the determination of candidate solutions for this iteration cycle. Subsequently, the non-dominated solution set can be iteratively updated based on the current candidate solution group, and the iterative process can be executed repeatedly until the preset termination condition is met.

[0087] The control parameters for iterative optimization are updated in real time using an incremental superposition method, and the particle positions are updated based on the updated parameters to obtain candidate solutions. Dynamic parameter updates are achieved through an incremental parameter correction mode, enabling gradual adjustment of the optimization parameters. This effectively avoids abrupt parameter changes and ensures the stability of the particle iterative update process.

[0088] Optionally, before updating the particle positions of each particle in the particle swarm in step S132, the method may further include: performing boundary constraint processing on the updated optimization control parameters to limit the updated optimization control parameters within a preset value range.

[0089] For example, preset value ranges can be set for different types of optimization control parameters in advance. The upper and lower limits of each parameter can be configured and fixed in advance according to the operating rules of the multi-objective particle swarm optimization algorithm and the solution requirements of the optimization problem to be solved. Boundary constraint processing can be to adjust the part of the updated optimization control parameter that exceeds the preset value range to within the preset value range. Specifically, when the value of the updated optimization control parameter is greater than the preset upper limit, the optimization control parameter can be corrected to the corresponding upper limit; when the value of the updated optimization control parameter is less than the preset lower limit, the optimization control parameter can be corrected to the corresponding lower limit; when the value of the updated optimization control parameter is within the preset value range, the original value of the parameter can be retained without modification.

[0090] After boundary constraint processing, all updated optimization control parameters can be restricted to their corresponding preset value ranges, resulting in effective optimization control parameters with complete constraint correction. Subsequently, based on the constraint-corrected optimization control parameters, the particle positions of each particle in the particle swarm can be updated, thereby determining the current candidate solution group for the multi-objective optimization problem to be solved, and iteratively completing the subsequent non-dominated solution set update process.

[0091] In one specific embodiment, the boundary constraint processing of the updated optimization control parameters can be expressed in the following form:

[0092] ;

[0093] ;

[0094] .

[0095] in, It can represent a boundary truncation function, used to ensure that the updated optimization control parameters are within a preset range. Within. For the current iteration cycle, This can represent the inertia weight parameter in the optimization control parameters of the previous iteration cycle. Individual learning factors Social learning factors Any one of them.

[0096] This allows for the standardization of the optimization control parameters' value range in each iteration cycle, preventing out-of-limit parameters from participating in the algorithm iteration. This ensures the effectiveness and feasibility of the particle iteration update rule.

[0097] For example, step S132, which updates the particle position of each particle in the particle swarm based on the updated optimization control parameters, may include, but is not limited to, the following steps S132a and S132b.

[0098] Step S132a: Based on the updated optimization control parameters, the velocity and position of each particle in the particle swarm in the previous iteration cycle, the global guiding solution in the updated non-dominated solution set of the previous iteration cycle, and the optimal position of each individual particle, calculate the update velocity of each particle in the particle swarm. The global guiding solution can be determined based on a preset guiding strategy.

[0099] After completing the current number After incrementally updating the control parameters in the first iteration cycle, the previous iteration cycle can be called back, i.e., the first iteration cycle. The particle swarm optimization data and solution set data retained from each iteration cycle. Specifically, this may include the particle swarm optimization data from the previous iteration cycle. The velocity of each particle Particle position Individual optimal position And the globally guiding solution in the non-dominated solution set updated in the previous iteration cycle. The optimal position of a particle can be the optimal position it found before the current iteration cycle. The global guided solution can be determined by filtering from the non-dominated solution set of the previous iteration cycle based on a preset guided strategy. The specific rules of the preset guided strategy can be flexibly configured according to the multi-objective optimization requirements. Different iteration cycles can use the same or different guided strategies to filter the global guided solution. Alternatively, the global guided solution can also be randomly selected from the non-dominated solution set.

[0100] Specifically, when calculating the update rate, the velocity of the particle in the previous iteration cycle can be used as a starting point. With the updated inertia weight parameters Multiply to obtain the first component The optimal position of the particle in the previous iteration cycle can be determined. With particle position The difference and the updated individual learning factor Multiplying them together yields the second component. The global guiding solution can be found in the non-dominated solution set updated in the previous iteration. The particle position in the previous iteration period The difference and the updated social learning factor Multiplying them together yields the third component. The update rate can be obtained by adding the first, second, and third components. Before adding, the second and third components can each be divided with a random number. and Multiplication. Random numbers. and The random number can be a uniformly distributed random number within the interval [0,1], or it can be a random number following other distributions. Therefore, the particle update rate for the current iteration cycle can be calculated. .

[0101] Step S132b: Update the particle positions of each particle in its corresponding particle swarm according to the update rate.

[0102] After obtaining the aforementioned update rate, the particle positions from the previous iteration can be updated based on the particle update rate of the current iteration. This update can be achieved by adding the particle position from the previous iteration to the particle update rate of the current iteration to obtain the new particle position, i.e., the particle position of the current iteration. Specifically, the updated particle position can be calculated using the following formula: .

[0103] Optionally, the velocity and position update operations described above can be performed on every single particle in the particle swarm, or they can be performed on only a subset of the particles. The updated particle positions can be included as part of the current candidate solution pool. When position updates are performed on multiple particles in the particle swarm, the updated particle positions can be collectively included as the current candidate solution pool.

[0104] In the above technical solution, the particle update rate is calculated by combining historical iterative particle motion data, individual optimal positions, and the global guiding solution of the non-dominated solution set, and then the particle position is updated. This fully utilizes historical optimization information and global high-quality solution information from the iterative process, improving the calculation basis for particle iterative updates. This makes particle position updates more aligned with the optimization logic of multi-objective optimization.

[0105] In a specific embodiment, such as a multi-objective capacity planning scenario, the candidate solution vector can represent the capacity configuration scheme of the equipment to be planned, such as photovoltaic capacity, energy storage capacity, and the number of charging facilities. The objective function can be set to multiple objectives such as economy, low carbon emissions, and reliability according to specific needs. Each particle in MOPSO can correspond to a candidate capacity configuration scheme. An external Pareto archive can be used to store the non-dominated capacity configuration schemes obtained during the current search process. In this application scenario, this application does not limit the specific capacity planning model, equipment type, or objective function form. Its core lies in: extracting the state vector of the multi-objective search state during the MOPSO solution process, inputting the state vector into the SAC network, and having the SAC network output continuous parameter increment actions. These actions are then used to adaptively update the inertia weight parameters, individual learning factors, and social learning factors of MOPSO, thereby improving the quality of the Pareto non-dominated solution set.

[0106] In a specific simulation experiment, this application was applied to a multi-objective capacity planning problem and compared with methods such as fixed-parameter MOPSO, linearly decreasing inertia weight MOPSO, nonlinearly decreasing inertia weight MOPSO, Gaussian decreasing inertia weight MOPSO, and heuristic adaptive MOPSO. Simulation results show that the Pareto solution set obtained by this application performs better in multi-objective optimization algorithm evaluation metrics such as hypervolume, generation distance, reverse generation distance, and number of non-dominated solutions. Specifically, in this simulation experiment, the hypervolume index of SAC-MOPSO reached 0.741, and the reverse generation distance was 0.018; compared with fixed-parameter MOPSO, the hypervolume was improved by 6.0%, the reverse generation distance was reduced by 62.5%, and the number of non-dominated solutions was increased by 108.8%; compared with Gaussian decreasing inertia weight MOPSO, the reverse generation distance was reduced by 40.0%, and the number of non-dominated solutions was increased by 34.1%. In microgrid capacity planning applications, the method improves the quality of the Pareto solution set, providing more stable economic, low-carbon, or reliability trade-offs for capacity configuration schemes.

[0107] The above results are only used to illustrate the technical effects of this application in specific simulation experiments and should not be construed as limiting the scope of protection of this application.

[0108] Figure 2 A schematic diagram of a parameter adaptive adjustment structure combining a parameter adjustment module with a multi-objective optimization algorithm according to an embodiment of this application is shown. Figure 2 As shown, the parameter adjustment module can be implemented using the SAC model, and the multi-objective optimization algorithm uses the MOPSO algorithm. For the SAC model, its input consists of various state features from the particle swarm optimization and Pareto archives. For example, in the current iteration of the current training round, the input to the SAC model can be a state vector. The SAC model can contain multiple layers of neural networks, which can process the input state vector. Perform mapping calculations to output the adjustment increments for each parameter, for example... These adjustment increments can be integrated into continuous parameter increment actions. The continuous parameter increment action output by the SAC model. Input can be sent to MOPSO. MOPSO receives continuously variable parameters and incremental actions. And update its own optimization control parameters. The updated optimization control parameters are obtained. The updated optimization control parameters can be used to perform MOPSO particle search and candidate solution evaluation, completing the iteration of particle velocity and position. Screening non-dominated solutions updates the Pareto archive, maintaining the non-dominated solution set. The new particle swarm state and Pareto archive state are then fed back to the state vector extraction stage, allowing the extraction of the state vector for the next iteration cycle. The data is fed into the SAC model to form a complete closed-loop control between iteration cycles.

[0109] Figure 3 A schematic diagram illustrating the offline training and online deployment process of a parameter adjustment module and a multi-objective optimization algorithm according to an embodiment of this application is shown. Figure 3 As shown, the process can be broken down into two independent phases: an offline training phase and an online deployment phase. For Figure 3 The offline training phase on the left aims to train a usable SAC model. The SAC model can include a policy network, a value network, and temperature parameters, all of which can be iteratively updated during training. Specific steps may include: in each training round, initializing the MOPSO particle swarm optimization and Pareto profile; and in each iteration of the training round, extracting the current optimization state to obtain the state vector. Based on this, the SAC model can output parameter increment actions. This action is used to correct various optimization control parameters of MOPSO, perform particle search and candidate solution evaluation, and then update the Pareto file; the reward value for the current iteration can be calculated based on the changes in the solution set. Furthermore, after completing the current iteration cycle, the state vector for the next iteration cycle is extracted and collected. The state vector of the current iteration period Incremental parameter action Reward Value and the state vector for the next iteration period The samples are encapsulated as state transition samples and stored in the experience replay pool. Batch samples are periodically extracted from the replay pool to update the SAC model. This iterative process continues until the termination condition for the current training round is met. Then, it is determined whether the overall training termination condition is met. If not, the MOPSO particle swarm optimization and Pareto profiles are reinitialized, and training for the next training round begins. If the conditions are met, the policy network parameters of the trained SAC model are stored. It can be understood that the above training process may include multiple training rounds, and each training round may include multiple iteration cycles.

[0110] for Figure 3The online deployment phase on the right aims to solve the actual optimization problem using the trained SAC policy network. Specifically, it may include the following steps: loading the SAC policy network trained in the offline training phase; initializing the particle swarm optimization and Pareto profile corresponding to the problem to be solved; and extracting the real-time state vector in each iteration cycle. The trained SAC policy network outputs incremental actions to adjust the optimization control parameters of MOPSO. Then, one MOPSO iteration is performed, including particle search and candidate solution evaluation, and the Pareto profile is updated. When a preset termination condition is met, the Pareto non-dominated solution set is output, and the solution process ends; otherwise, the next iteration cycle begins. It can be understood that during the online deployment phase, the network parameters of the SAC model are not updated.

[0111] According to a second aspect of this application, a multi-objective particle swarm optimization parameter adjustment system is also provided. Figure 4 A schematic block diagram of a multi-objective particle swarm optimization parameter adjustment system 400 according to one embodiment of this application is shown. Figure 4 As shown, system 400 may include:

[0112] The state extraction module 410 is used to extract a state vector for the current multi-objective optimization problem to be solved, wherein the state vector is used to represent the multi-objective search process;

[0113] The trained parameter tuning module 420 is used to receive the state vector and output the tuning result of the optimization control parameters, wherein the optimization control parameters represent the search control parameters used to solve the current multi-objective optimization problem to be solved;

[0114] Solution determination module 430 is used to determine the current candidate solution group for the current multi-objective optimization problem to be solved based on the adjustment results of the optimization control parameters;

[0115] The solution set update module 440 is used to iteratively update the non-dominated solution set of the current multi-objective optimization problem to be solved based on the current candidate solution group, until the preset termination condition is met and the latest non-dominated solution set is output.

[0116] According to a third aspect of this application, an electronic device is also provided. Figure 5 A schematic block diagram of an electronic device 500 according to one embodiment of this application is shown. Figure 5 As shown, the electronic device 500 may include a processor 510 and a memory 520. The memory 520 stores computer program instructions 521, which, when executed by the processor 510, are used to perform the multi-objective particle swarm optimization parameter adjustment method described above.

[0117] According to a fourth aspect of this application, a non-volatile storage medium is also provided, on which program instructions are stored. These program instructions, when executed, are used to perform the multi-objective particle swarm optimization parameter adjustment method described above. The non-volatile storage medium may, for example, include a storage component of a tablet computer, a hard disk of a personal computer, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disc read-only memory (CD-ROM), a USB memory, or any combination of the above storage media. The non-volatile storage medium may be any combination of one or more non-volatile storage media.

[0118] Those skilled in the art can understand the specific implementation scheme and beneficial effects of the above-mentioned multi-objective particle swarm optimization parameter adjustment system, electronic device and non-volatile storage medium by reading the relevant description of the above-mentioned multi-objective particle swarm optimization parameter adjustment method. For the sake of brevity, it will not be described in detail here.

[0119] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of this application. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of this application. All such changes and modifications are intended to be included within the scope of this application as claimed in the appended claims.

[0120] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0122] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0123] Similarly, it should be understood that, in order to simplify this application and aid in understanding one or more aspects of the various applications, features of this application are sometimes grouped together in a single embodiment, figure, or description thereof in the description of exemplary embodiments of this application. However, this approach should not be construed as reflecting an intention that the claimed application requires more features than are expressly recited in each claim. Rather, as reflected in the corresponding claims, the inventive point lies in solving the corresponding technical problem with fewer features than all features of a single disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0124] Those skilled in the art will understand that, apart from the mutual exclusion of features, all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or elements of any method or apparatus so disclosed may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0125] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0126] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules in the system and / or device according to the embodiments of this application. This application can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0127] It should be noted that the above embodiments are illustrative of this application and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0128] The above description is merely a specific embodiment or illustration of the embodiments of this application. The scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. The scope of protection of this application shall be determined by the scope of the claims.

Claims

1. A multi-objective particle swarm optimization parameter adjustment method, characterized by, include: For the current multi-objective optimization problem to be solved, a state vector is extracted, wherein the state vector is used to represent the multi-objective search process; The state vector is input to the trained parameter adjustment module to output the adjustment result of the optimization control parameters, wherein the optimization control parameters represent the search control parameters used to solve the current multi-objective optimization problem to be solved; Based on the adjustment results of the optimization control parameters, the current candidate solution group for the current multi-objective optimization problem to be solved is determined; Based on the current candidate solution group, the non-dominated solution set of the current multi-objective optimization problem to be solved is iteratively updated until a preset termination condition is met, and then the latest non-dominated solution set is output.

2. The multi-objective particle swarm optimization parameter adjustment method according to claim 1, characterized by, The parameter adjustment module is implemented using the soft actor-critic algorithm; the training process of the parameter adjustment module includes one or more training rounds, and each training round includes multiple iteration cycles of the multi-objective optimization algorithm; The training steps of the parameter adjustment module include: For each training round, the reward value for the current iteration period is calculated based on the quality change of the non-dominated solution set of the current multi-objective optimization problem before and after the update. Based on the state transition samples of each iteration cycle, the network parameters of the parameter adjustment module are iteratively updated until the training termination condition is met. The state transition samples are generated based on the reward value of the corresponding iteration cycle.

3. The method for adjusting parameters in multi-objective particle swarm optimization according to claim 2, characterized in that, The reward value is calculated using the following formula: ,in, Indicates the number of training rounds in the current round. The reward value per iteration cycle. Indicates the number of training rounds in the current round. The change in distance characteristics between the non-dominated solution set and the theoretical optimal solution set of the current multi-objective optimization problem over several iterations. Indicates the number of training rounds in the current round. The change in the distribution characteristics of all non-dominated solutions in the set of non-dominated solutions over one iteration period. Indicates the number of training rounds in the current round. The change in the quantity of all non-dominated solutions in the non-dominated solution set over one iteration cycle. , as well as All are weighting coefficients.

4. The method for adjusting parameters in multi-objective particle swarm optimization according to claim 1, characterized in that, The state vector includes particle swarm state information, non-dominated solution set state information, and current iteration progress information. The particle swarm state information includes particle position distribution characteristics. The non-dominated solution set state information includes at least one of the following: non-dominated solution set size information, non-dominated solution set front distribution span information, and distance information between the particle swarm and the non-dominated solution set.

5. The method for adjusting parameters in multi-objective particle swarm optimization according to claim 1, characterized in that, The process of determining the current candidate solution group for the multi-objective optimization problem to be solved based on the adjustment results of the optimization control parameters includes: When the adjustment result of the optimization control parameter is expressed as an incremental parameter, the optimization control parameter for the current iteration period is determined based on the adjustment result and optimization control parameter of the previous iteration period. Based on the updated optimization control parameters, the particle positions of each particle in the particle swarm are updated, and the updated particle positions are used as at least a part of the current candidate solution group.

6. The method for adjusting parameters in multi-objective particle swarm optimization according to claim 5, characterized in that, The process of updating the particle positions of each particle in the particle swarm based on the updated optimization control parameters includes: Based on the updated optimization control parameters, the velocity and position of each particle in the particle swarm in the previous iteration cycle, the global guiding solution in the non-dominated solution set updated in the previous iteration cycle, and the optimal position of each individual particle, the update velocity of each particle in the particle swarm is calculated; wherein, the global guiding solution is determined based on a preset guiding strategy. The particle positions of each particle in the corresponding particle swarm are updated according to the update rate.

7. The method for adjusting parameters in multi-objective particle swarm optimization according to claim 5, characterized in that, Before updating the particle positions of each particle in the particle swarm, the method further includes: Boundary constraint processing is applied to the updated optimization control parameters to limit them to a preset value range.

8. A multi-objective particle swarm optimization parameter adjustment system, characterized in that, include: The state extraction module is used to extract a state vector for the current multi-objective optimization problem to be solved, wherein the state vector is used to represent the multi-objective search process; The trained parameter adjustment module is used to receive the state vector and output the adjustment result of the optimization control parameters, wherein the optimization control parameters represent the search control parameters used to solve the current multi-objective optimization problem to be solved; The solution determination module is used to determine the current candidate solution group of the current multi-objective optimization problem to be solved based on the adjustment results of the optimization control parameters. The solution set update module is used to iteratively update the non-dominated solution set of the current multi-objective optimization problem to be solved based on the current candidate solution group, until a preset termination condition is met and the latest non-dominated solution set is output.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, which, when executed by the processor, are used to perform the multi-objective particle swarm optimization parameter adjustment method as described in any one of claims 1 to 7.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores program instructions that, when executed, are used to perform the multi-objective particle swarm optimization parameter adjustment method as described in any one of claims 1 to 7.

Images