Dynamic multi-objective path optimization method based on long short-term memory network information accumulation

Abstract

The invention discloses a dynamic multi-target path optimization method based on long short-term memory network information accumulation, which relates to the technical field of dynamic multi-target path optimization, and comprises the following steps: obtaining a dynamic multi-target path optimization problem, and initializing a population and parameters; an environment change detection strategy is executed in the initial stage of each generation of evolution, if the environment changes, a nonlinear prediction strategy of long and short-term memory network information accumulation is executed, and then a population auxiliary adjustment strategy is executed; and a static multi-objective evolutionary algorithm is used to optimize the dynamic multi-objective path optimization problem, iteration is carried out until the maximum number of iterations is reached, and a final optimization result is output as an optimal path. According to the method, the accuracy of dynamic multi-target path optimization is ensured while the environmental adaptability is improved by constructing and combining a nonlinear prediction strategy and a population auxiliary adjustment strategy of long and short-term memory network information accumulation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dynamic multi-objective path optimization, and particularly relates to a dynamic multi-objective path optimization method based on long short-term memory network information accumulation. BACKGROUND

[0002] In the field of dynamic path planning, due to the influence of dynamic change factors such as sudden conditions and resource requirements, the problem target function expression, decision variables, target number, and environmental parameters will change over time or environment. Such problems are called dynamic multi-objective path optimization problems. In recent years, evolutionary algorithms have been considered an effective method for solving dynamic multi-objective path optimization problems due to their wide applicability, good global search ability, and outstanding performance in multi-objective optimization problems. However, evolutionary algorithms converge to the optimal front through population convergence. In the late evolution, due to severe loss of diversity, the population's adaptability to the environment gradually weakens, and it cannot adapt to changes in the new environment. Therefore, how to maintain the diversity of the population is a major challenge in solving dynamic multi-objective path optimization problems. At the same time, the time interval of dynamic changes is very short, and how the algorithm quickly tracks the optimal solution and optimal front in the new environment is another major challenge in solving dynamic multi-objective path optimization problems. Therefore, it is necessary to study dynamic multi-objective path optimization methods.

[0003] In the prior art, Chinese patent CN117689092A discloses a constraint multi-modal multi-objective path optimization method based on dynamic sorting, generates a first path population and a second path population including N connectable path individuals, determines a first weight and a second weight for balancing the conflict between constraint quality and convergence quality at the current iteration number, and iterates the first path population and the second path population according to a preset co-evolution strategy, the first weight and the second weight. The strategy is set according to the objective function and the constraint condition of the model to be solved, so as to iteratively output the first path population that more satisfies the optimization target of the objective function on the basis of satisfying the constraint condition. When the preset iteration termination condition is reached, the N path individuals in the first path population at this time are determined as N equivalent optimal paths.

[0004] However, the above prior art does not fully utilize historical information, and it is difficult to capture nonlinear and irregular change patterns that may exist in the dynamic environment, and it is difficult to balance environmental adaptability and the accuracy of dynamic multi-objective path optimization. SUMMARY

[0005] The present application provides a dynamic multi-objective path optimization method based on long short-term memory network information accumulation, to solve the problem that existing dynamic multi-objective path optimization techniques cannot balance environmental adaptability and the accuracy of dynamic multi-objective path optimization.

[0006] In one aspect, the application provides a dynamic multi-objective path optimization method based on long short-term memory network information accumulation, comprising the following steps: Step one, obtaining a dynamic multi-objective path optimization problem, initializing a population and parameters.

[0007] Step two, performing an environment change detection strategy in the initial stage of each generation evolution. If the environment changes, the next step is performed. If the environment does not change, go to step four.

[0008] Step three, performing a long short-term memory network information accumulation nonlinear prediction strategy, and then performing a population auxiliary adjustment strategy.

[0009] Step four, using a static multi-objective evolutionary algorithm to optimize the dynamic multi-objective path optimization problem, returning to step two for iteration until the maximum number of iterations is reached, and outputting the final optimization result as the optimal path.

[0010] In one possible implementation, in step two, the environment change detection strategy comprises: A change detection operator is used to randomly extract a certain proportion of individuals in the population as environment detectors. The environment detectors are re-evaluated in the initial stage of each generation evolution. If the current objective value of the environment detector is different from the objective value of the last generation, it is considered that the environment has changed, otherwise it is considered that the environment has not changed.

[0011] In one possible implementation, in step three, the long short-term memory network information accumulation nonlinear prediction strategy comprises: Based on the forget gate, the input gate and the candidate memory cell state calculation information accumulation set of the long short-term memory network.

[0012] Based on the output gate of the long short-term memory network and the information accumulation set, the hidden state output of the long short-term memory network is calculated.

[0013] The hidden state output is mapped back to the decision variable value range by a mapping function to obtain a mapping result set.

[0014] The mapping weight matrix and the mapping result set are used to predict the population to obtain a predicted population.

[0015] A number of individuals are selected from the union of the initial population and the predicted population to form a new initial population.

[0016] In one possible implementation, in step three, when time The information accumulation set starts to accumulate historical information, and is updated after each environment change; when time The information accumulation set at the current time is determined by the input at the current time and the information accumulation set at the previous time.

[0017] In a possible implementation, in step three, after each environment change, a parameter correction strategy based on the least square method is used to correct the mapping weight matrix.

[0018] In a possible implementation, in step three, the population auxiliary adjustment strategy includes: Calculate the loss vector and loss value at the current time.

[0019] Count the number of elements in the loss set that are greater than the loss value, and calculate the parameter probability.

[0020] According to the parameter probability, generate a mutation solution and a random solution to join the initial population.

[0021] Save the loss value to the loss set, and after fast non-dominated sorting of the initial population, use the truncation operation to select a number of individuals as the new initial population.

[0022] In a possible implementation, in step three, according to the parameter probability, generating a mutation solution and a random solution to join the initial population includes: When the parameter probability is less than a preset probability threshold, generate a mutation solution and a random solution to join the initial population.

[0023] The number of mutation solutions is determined by one minus the parameter probability, and the number of random solutions is determined by the parameter probability.

[0024] In a possible implementation, in step four, the static multi-objective evolutionary algorithm uses the MOEA / D-DE algorithm.

[0025] The dynamic multi-objective path optimization method based on long short-term memory network information accumulation in the application has the following advantages: By constructing and combining the nonlinear prediction strategy and the population auxiliary adjustment strategy based on long short-term memory network information accumulation, the environmental adaptability is improved while the accuracy of dynamic multi-objective path optimization is ensured.

[0026] When the time , the information accumulation set starts to accumulate historical information, and after each environment change, the information accumulation set is updated, when the time , the information accumulation set at the current time is determined by the input at the current time and the information accumulation set at the previous time, and the information is accumulated by learning the information of all time steps, rather than only from the information of the solution in the last one or two time steps.

[0027] After each environment change, the parameter correction strategy based on the least square method is used to correct the mapping weight matrix, which reduces the error accumulation.

[0028] The proposed population auxiliary adjustment strategy includes: calculating a loss vector and a loss value at a current time, counting the number of elements in the loss set greater than the loss value, calculating a parameter probability, generating a mutation solution and a random solution according to the parameter probability, adding the mutation solution and the random solution to the initial population, saving the loss value to the loss set, performing fast non-dominated sorting on the initial population, and selecting a plurality of individuals as a new initial population using a truncation operation, which improves the inaccuracy of prediction.

[0029] When the parameter probability is less than a preset probability threshold, the mutation solution and the random solution are generated and added to the initial population, the number of mutation solutions is determined by one minus the parameter probability, and the number of random solutions is determined by the parameter probability. This adaptive adjustment of mutation solutions and random solutions increases the diversity of the population, thereby maintaining convergence while better exploring new solution spaces. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0031] Figure 1 A flowchart of a dynamic multi-objective path optimization method based on long short-term memory network information accumulation is provided for the embodiments of the present application. Figure 2 A framework diagram of a dynamic multi-objective path optimization method based on long short-term memory network information accumulation is provided for the embodiments of the present application. Figure 3 A process diagram of a nonlinear prediction strategy of long short-term memory network information accumulation is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0033] Without loss of generality, an unconstrained dynamic multi-objective path optimization problem is shown in formula (1):

[0034]

[0035] wherein, is a time instance, denotes a time space, is a D-dimensional decision vector, is a decision space. denotes a target vector containing targets over time.

[0036] Definition 1 (Pareto dominance): For two decision variables at time , if and satisfy formula (2).

[0037]

[0038] is considered to dominate , denoted as .

[0039] Definition 2 (Pareto optimal solution set, DPOS): For a solution , if it is not dominated by other solutions, it is called a Pareto optimal solution. The DPOS is the set of all Pareto optimal solutions, which is defined as formula (3).

[0040]

[0041] Definition 3 (Pareto optimal front, DPOF): At time t, the dynamic Pareto optimal front (DPOF) is the mapping vector of the dynamic Pareto optimal solution set (DPOS) in the target space, which is expressed as formula (4).

[0042]

[0043] The long short-term memory network is a special recurrent neural network. Due to its special gate structure and storage structure, it can selectively add new information and selectively forget previously stored information, thereby making up for the problems of RNN network in solving short-term memory and gradient disappearance / gradient explosion, and better learning the relevant characteristics contained in time series data.

[0044] The properties of the long short-term memory network are related to the dynamic multi-objective path optimization method based on long short-term memory network information accumulation of the present application as follows: The long short-term memory network has the property of recurrence of sequence data (such as language, speech and time series). There is a feedback loop in the structure of the long short-term memory network, which can learn and remember the information observed in the past by forming an information loop structure. Based on the recursive property of the long short-term memory network, information accumulation is carried out, which guides the evolution of the population once the environment changes. ​

[0045] As long short-term memory network has good prediction ability, it has been proved to be a universal approximator. Dynamic multi-objective path optimization problem is similar to the change of the environment, and the prediction of the current environment needs the information of the historical environmental change to guide.

[0046] As shown in Figure 1 and Figure 2 , the embodiment of the application provides a dynamic multi-objective path optimization method based on long short-term memory network information accumulation, which comprises the following steps: Step one, obtaining a dynamic multi-objective path optimization problem, initializing a population and parameters.

[0047] Step two, in the initial stage of each generation evolution, an environment change detection strategy is executed, if the environment changes, the next step is executed, if the environment does not change, step four is jumped to.

[0048] Step three, a long short-term memory network information accumulation nonlinear prediction strategy is executed, and then a population auxiliary adjustment strategy is executed.

[0049] Step four, using a static multi-objective evolutionary algorithm to optimize the dynamic multi-objective path optimization problem, returning to step two for iteration until the maximum iteration number is reached, and outputting the final optimization result as the optimal path.

[0050] Specifically, from Figure 2 It can be seen that the framework of the dynamic multi-objective path optimization method based on long short-term memory network information accumulation mainly includes three parts: a static multi-objective evolutionary algorithm as a basic optimization solver, a change detection operator for detecting environmental changes, and a prediction strategy for predicting the Pareto optimal front in the new environment.

[0051] Exemplarily, in step two, the environment change detection strategy comprises: A change detection operator is used to randomly extract a certain proportion of individuals in the population as environment detectors, which are re-evaluated in the initial stage of each generation evolution, if the current target value of the environment detector is different from the target value of the last generation, it is considered that the environment changes, otherwise it is considered that the environment does not change.

[0052] Specifically, in this embodiment, the change detection operator randomly extracts 5% of the individuals in the population as environment detectors.

[0053] As shown in Figure 3 , exemplarily, in step three, the long short-term memory network information accumulation nonlinear prediction strategy comprises: Based on the long short-term memory network, the forgetting gate, the input gate and the candidate memory cell state calculation information accumulation set.

[0054] The hidden state output of the long short-term memory network is calculated based on an output gate of the long short-term memory network and the information accumulation set.

[0055] The hidden state output is mapped back into a decision variable value range through a mapping function to obtain a mapping result set.

[0056] The mapping weight matrix and the mapping result set are used to predict a population to obtain a predicted population.

[0057] A number of individuals are selected from the union of the initial population and the predicted population to form a new initial population.

[0058] Exemplarily, in step three, when time , the information accumulation set starts to accumulate historical information, and the information accumulation set is updated after each environmental change; when time , the information accumulation set at the current time is determined by the input at the current time and the information accumulation set at the previous time.

[0059] Specifically, in the dynamic multi-objective optimization process, at time , the information accumulation set is determined by the unit state, and the unit state is determined by the forget gate and the output gate. When time , the information accumulation set starts to accumulate historical information, and the information accumulation set is updated after each environmental change. When , the information accumulation set at the current time is determined not only by the input at the current time, but also by the information accumulation set at the previous time . From , the cycle is repeated in turn, when , the information accumulation set accumulates the information at time .

[0060] Specifically, the calculation formula of the forget gate of the long short-term memory network is shown in formula (5):

[0061] wherein, is a sigmoid activation function, is a weight matrix of the forget gate, is a bias term of the forget gate.

[0062] The calculation formula of the input gate of the long short-term memory network is shown in formula (6):

[0063] wherein, sigmoid activation function, weight matrix of the input gate, bias term of the input gate.

[0064] candidate memory cell state of the LSTM The calculation formula is shown in equation (7):

[0065] wherein, activation function, weight matrix of the candidate memory cell state, bias term of the candidate memory cell state.

[0066] Therefore, information accumulation set at the moment, i.e. cell state The calculation formula is shown in equation (8):

[0067] wherein, forget gate, input gate, information accumulation set at the moment, candidate memory cell state.

[0068] output gate of the LSTM The calculation formula is shown in equation (9):

[0069] wherein, sigmoid activation function, weight matrix of the output gate, bias term of the output gate.

[0070] hidden state output of the LSTM The calculation formula is shown in equation (10):

[0071] wherein, activation function, output gate, information accumulation set.

[0072] Considering the nonlinear factor, an activation function is added. The value after the activation function is mapped back to the value range of the decision variable, and the mapping function is as shown in equation (11):

[0073] ​wherein, is the mapping result set at time step , i.e. the set of all possible decision variable values after being processed by the mapping function, denotes the mapped decision variable must be within the value range of all decision variables. and are the minimum value and the maximum value of the th decision variable, respectively.

[0074] Finally, the population is predicted using the mapping weight matrix from the output layer to the prediction set and the mapping result set, as shown in equation (12):

[0075] Exemplarily, in step three, after each environmental change, a parameter correction strategy based on the least square method is used to correct the parameters of the mapping weight matrix.

[0076] Specifically, in the dynamic multi-objective path optimization problem, as time changes, errors may accumulate, causing the population to deviate more and more from the PS / PF. To solve this problem, a parameter correction strategy is proposed. After the long short-term memory network unit accumulates information and extracts features, it is transmitted to the output layer through the hidden layer. After each environmental change, a parameter correction strategy based on the least square method is used to directly correct the parameters of the mapping weight matrix , so that the algorithm can adapt to environmental changes more quickly and maintain the ability to track the Pareto optimal front. Therefore, the parameter correction strategy based on the least square method is shown in equation (13):

[0077] wherein, is the approximate Pareto optimal solution after static optimization of the current environment, is the result after mapping with nonlinear factors.

[0078] Exemplarily, in step three, the population auxiliary adjustment strategy includes: Calculate the loss vector and loss value at the current time.

[0079] Count the number of elements in the loss set that are greater than the loss value, and calculate the parameter probability.

[0080] According to the parameter probability, generate a mutant solution and a random solution to join the initial population.

[0081] Save the loss value to the loss set, and after quickly non-dominated sorting of the initial population, use the truncation operation to select a number of individuals as the new initial population.

[0082] In step three, according to the parameter probability, the mutation solution and the random solution are generated and added to the initial population. When the parameter probability is less than a preset probability threshold, the mutation solution and the random solution are generated and added to the initial population.

[0083] The number of mutation solutions is determined by a parameter probability, and the number of random solutions is determined by the parameter probability.

[0084] Specifically, although the nonlinear prediction strategy based on long short-term memory network information accumulation can utilize historical information to predict new solutions in the future environment, in some cases, relying only on the prediction strategy may not obtain accurate results. For example, when the historical information is insufficient (early stage of the optimization process), the model may not be well trained, resulting in a large prediction error. In addition, when the severity of environmental changes is high, the prediction strategy cannot accurately capture the pattern of changes, resulting in a significant error between the predicted solution and the Pareto optimal solution. Therefore, the population auxiliary adjustment strategy is introduced to adjust the population, and the introduction proportion of mutation solutions and random solutions is dynamically adjusted according to the prediction error, which increases the diversity of the population, thereby maintaining the convergence while better exploring the new solution space.

[0085] For the calculation of the parameter probability, first, the average value of the squared error loss of all samples is called the loss function, which is defined as shown in equation (14).

[0086]

[0087] wherein, is a loss vector. is the number of input samples. is the th predicted sample, is the th accurate sample. Therefore, the loss value at the current time step can be calculated by the average value of all variables of the loss vector, as shown in equation (15):

[0088] wherein, is the number of decision variables. is the th variable of the loss vector Initialize the relevant parameters, wherein is the total number of loss sets greater than . Then, the parameter probability is calculated as shown in equation (16):

[0089] wherein, is an integer value in . The greater the value, the smaller the prediction error.

[0090] In this embodiment, when , a mutation solution of is generated and a random solution of is added to the population. By introducing the mutation solution and the random solution, the local search ability and the global search ability of the algorithm are balanced to better adapt to environmental changes. Finally, the loss value is saved to the loss set , and the population is quickly non-dominant sorted and then selected by truncation operation to select individuals as the population of the new time.

[0091] Exemplarily, in step four, the static multi-objective evolutionary algorithm adopts the MOEA / D-DE algorithm.

[0092] Specifically, the MOEA / D-DE algorithm is an existing static multi-objective evolutionary algorithm, which will not be described in detail here. In other possible embodiments, other static multi-objective evolutionary algorithms can also be used.

[0093] The computational complexity of the dynamic multi-objective path optimization method based on long short-term memory network information accumulation of the present application is as follows: In each generation cycle, the computational resources of the dynamic multi-objective path optimization method based on long short-term memory network information accumulation are mainly consumed by the long short-term memory network information accumulation nonlinear prediction strategy and the population auxiliary adjustment strategy. In the long short-term memory network information accumulation nonlinear prediction strategy, information accumulation is associated with the information of the previous time , the current time and the initial population , therefore, the time complexity of information accumulation is , wherein, is the population size, is the decision variable dimension, is the hidden layer dimension. The time complexity of nonlinear mapping is . For parameter modification, the time complexity of modifying is . In the execution process of the population auxiliary adjustment strategy, the time complexity of calculating is , wherein, is the size of . The fast non-dominant sorting and the truncation operation are performed to obtain a population size of population of the species The time complexity of the method is wherein, is the target dimension. In summary, the total time complexity is Generally, Therefore, the total time complexity is which is the same as most existing prediction strategies.

[0094] In one possible embodiment, in order to verify the effectiveness of the proposed dynamic multi-objective path optimization method based on long short-term memory network information accumulation, a series of detailed experiments are carried out. First, the dynamic multi-objective path optimization method based on long short-term memory network information accumulation is compared with five representative algorithms (including: KT-DMOEA algorithm, IT-DMOEA algorithm, SGEA algorithm, MOEA / D-RNN algorithm, SVR-DMOEA algorithm) using MIGD and MHV indicators to prove its effectiveness in dynamic environment. Then the MIGD values obtained by the dynamic multi-objective path optimization method based on long short-term memory network information accumulation and the five representative algorithms are plotted with the change of the environment, in order to fully visualize the performance of each algorithm under different environments. Subsequently, an ablation experiment is carried out to evaluate the contribution of the two key strategies proposed in this application, including the nonlinear prediction strategy of long short-term memory network information accumulation and the population auxiliary adjustment strategy. The proposed method is also compared with the DMIGD performance indicators of the other five representative algorithms to verify the generalization ability of the dynamic multi-objective path optimization method based on long short-term memory network information accumulation. Finally, Friedman test is used to compare different algorithms to determine whether there is a significant difference between them.

[0095] For the fourteen dynamic multi-objective path optimization problems (three 3 test cases are set for each problem, a total of 42 test cases), the experimental comparison results of the MIGD index are: the dynamic multi-objective path optimization method based on long short-term memory network information accumulation achieves the best result in 28 of the 42 test cases, the KT-DMOEA algorithm achieves the best result in 3 of the 42 test cases, the IT-DMOEA algorithm achieves the best result in 2 of the 42 test cases, the SGEA algorithm achieves the best result in 4 of the 42 test cases, the MOEA / D-RNN algorithm achieves the best result in 3 of the 42 test cases, and the SVR-DMOEA algorithm achieves the best result in 2 of the 42 test cases.

[0096] The MHV index better quantifies the convergence and diversity of the solution obtained by the algorithm, and the experimental results of the MHV index are as follows: the dynamic multi-objective path optimization method based on long short-term memory network information accumulation obtains 31 best results in 42 test cases, the KT-DMOEA algorithm obtains 4 best results, the IT-DMOEA algorithm obtains 0 best results, the SGEA algorithm obtains 4 best results, the MOEA / D-RNN algorithm obtains 2 best results, and the SVR-DMOEA algorithm obtains 1 best result.

[0097] In a possible embodiment, the MIGD values (environmental change intensity , environmental change frequency ) obtained by the dynamic multi-objective path optimization method based on long short-term memory network information accumulation and five representative algorithms are plotted with the change of the environment, and the experimental results show that in most cases, the curve generated by the dynamic multi-objective path optimization method based on long short-term memory network information accumulation is lower, so it has superior performance in convergence and diversity. In addition, compared with other algorithms, the curve generated by the dynamic multi-objective path optimization method based on long short-term memory network information accumulation is smoother, which means that the algorithm not only has better performance, but also is more stable.

[0098] In a possible embodiment, an ablation experiment is performed, and the original dynamic multi-objective path optimization method based on long short-term memory network information accumulation is adapted into two variants, the first variant eliminates the nonlinear prediction strategy of long short-term memory network information accumulation, and the second variant eliminates the population auxiliary adjustment strategy. Under the same dynamic setting, the original dynamic multi-objective path optimization method based on long short-term memory network information accumulation is compared with the two variants on ten dynamic multi-objective path optimization problems, including five two-target and five three-target test functions. The experimental results show that in terms of the MIGD index, the dynamic multi-objective path optimization method based on long short-term memory network information accumulation is significantly better than the two variants in most test problems, which shows that the two key strategies play a crucial role in the high performance of the experimental problems.

[0099] In a possible embodiment, the dynamic multi-objective path optimization method based on long short-term memory network information accumulation is compared with the DMIGD performance index of five representative algorithms for fourteen dynamic multi-objective path optimization problems, and the results show that the dynamic multi-objective path optimization method based on long short-term memory network information accumulation fully utilizes the information from the new environment and the historical environment, and can effectively predict the change of the Pareto optimal frontier or the Pareto optimal solution set in 9 of the fourteen dynamic multi-objective path optimization problems, which shows that the method has good generalization performance in dealing with dynamic multi-objective path optimization problems with different dynamic characteristics.

[0100] In a possible embodiment, Friedman test is used to compare the dynamic multi-objective path optimization method based on long short-term memory information accumulation with five representative algorithms. The best ranking score of the dynamic multi-objective path optimization method based on long short-term memory information accumulation is 1.76, which is significantly better than 4.90 of the KT-DMOEA algorithm, 4.33 of the IT-DMOEA algorithm, 2.80 of the SGEA algorithm, 3.07 of the MOEA / D-RNN algorithm and 3.93 of the SVR-DMOEA algorithm.

[0101] In summary, the overall performance of the dynamic multi-objective path optimization method based on long short-term memory information accumulation of the present application is better than that of the existing dynamic multi-objective evolutionary algorithm.

[0102] In other possible embodiments, the dynamic multi-objective path optimization method based on long short-term memory information accumulation of the present application can also be applied to the field of industrial scheduling, system control and other fields with dynamic multi-objective optimization problems.

[0103] The embodiments of the present application improve the environmental adaptability while ensuring the accuracy of dynamic multi-objective path optimization by constructing and combining the nonlinear prediction strategy and the population auxiliary adjustment strategy based on long short-term memory information accumulation.

[0104] When time , the information accumulation set starts to accumulate historical information, and the information accumulation set is updated after each environmental change, when time , the information accumulation set at the current time is determined by the input at the current time and the information accumulation set at the previous time, and the information accumulation is performed by learning the information of all solutions at all time steps, rather than only the information of solutions at the last one or two time steps.

[0105] After each environmental change, the parameter correction strategy based on the least square method is used to correct the parameters of the mapping weight matrix, thereby reducing the error accumulation.

[0106] The proposed population auxiliary adjustment strategy includes: calculating the loss vector and loss value at the current time, counting the number of elements in the loss set greater than the loss value, calculating the parameter probability, generating mutant solutions and random solutions to join the initial population according to the parameter probability, saving the loss value to the loss set, performing fast non-dominated sorting on the initial population and using the truncation operation to select a number of individuals as the new initial population, and improving the inaccuracy of prediction. When the parameter probability is less than the preset probability threshold, mutant solutions and random solutions are generated to join the initial population, the number of mutant solutions is determined by one minus the parameter probability, and the number of random solutions is determined by the parameter probability. This adaptive adjustment of mutant solutions and random solutions increases the diversity of the population, so that the convergence can be maintained while better exploring the new solution space.

[0107] Although preferred embodiments of the application have been described, those skilled in the art will appreciate that other modifications and variations to the preferred embodiments are possible without departing from the spirit and scope of the application. Therefore, it is intended that the appended claims encompass all such modifications and variations as fall within the scope of the application.

[0108] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, it is intended that the present application encompass all such modifications and changes as fall within the scope of the claims and their equivalents.

Claims

1. A dynamic multi-objective path optimization method based on long short-term memory network information accumulation, characterized in that, The method comprises the following steps: Step 1: obtaining a dynamic multi-objective path optimization problem, initializing a population and parameters; Step 2: performing an environment change detection strategy in the initial stage of each generation evolution, if the environment changes, performing the next step, if the environment does not change, jumping to step 4; Step 3: performing a long short-term memory network information accumulation nonlinear prediction strategy, and then performing a population auxiliary adjustment strategy; Step 4: using a static multi-objective evolutionary algorithm to optimize the dynamic multi-objective path optimization problem, returning to step 2 for iteration until the maximum number of iterations is reached, and outputting the final optimization result as the optimal path. 2.The dynamic multi-objective path optimization method based on long short-term memory network information accumulation of claim 1, wherein, In step 2, the environment change detection strategy comprises: A change detection operator is used to randomly extract a certain proportion of individuals in the population as environment detectors, which are re-evaluated in the initial stage of each generation evolution, if the current objective value of the environment detector is different from the objective value of the last generation, it is considered that the environment has changed, otherwise it is considered that the environment has not changed. 3.The dynamic multi-objective path optimization method based on long short-term memory network information accumulation of claim 1, wherein, In step 3, the long short-term memory network information accumulation nonlinear prediction strategy comprises: Calculating an information accumulation set based on the forget gate, input gate and candidate memory cell state of the long short-term memory network; Calculating the hidden state output of the long short-term memory network based on the output gate and the information accumulation set; Mapping the hidden state output back to the decision variable value range through a mapping function to obtain a mapping result set; Using a mapping weight matrix and the mapping result set to predict the population to obtain a predicted population; Selecting a number of individuals from the union of the initial population and the predicted population to form a new initial population. 4.The dynamic multi-objective path optimization method based on long short-term memory network information accumulation of claim 3, wherein, In step three, when time the information accumulation set starts to accumulate historical information, and is updated after each environmental change; when time the information accumulation set at the current time is determined by the input at the current time and the information accumulation set at the previous time. 5.The dynamic multi-objective path optimization method based on long short-term memory network information accumulation of claim 3, wherein, In step 3, after each environment change, a parameter correction strategy based on the least squares method is used to correct the parameters of the mapping weight matrix. 6.The dynamic multi-objective path optimization method based on long short-term memory network information accumulation of claim 1, wherein, In step 3, the population auxiliary adjustment strategy comprises: Calculating the loss vector and loss value at the current time; Counting the number of elements in the loss set that are greater than the loss value, and calculating the parameter probability; According to the parameter probability, generate a mutant solution and a random solution to join the initial population; Saving the loss value to the loss set, and after performing fast non-dominated sorting on the initial population, using the truncation operation to select a number of individuals as the new initial population.

7. The dynamic multi-objective path optimization method based on long short-term memory network information accumulation according to claim 6, characterized in that, In step 3, according to the parameter probability, generating a mutant solution and a random solution to join the initial population comprises: When the parameter probability is less than a predetermined probability threshold, generate a mutant solution and a random solution to join the initial population; The number of mutant solutions is determined by the parameter probability, and the number of random solutions is determined by the parameter probability. 8.The dynamic multi-objective path optimization method based on long short-term memory network information accumulation of claim 1, wherein, In step 4, the static multi-objective evolutionary algorithm uses the MOEA / D-DE algorithm.

Images