Asynchronous parallel hyper-heuristic optimization method

Abstract

The invention relates to the field of intelligent optimization algorithms, and discloses an asynchronous parallel hyper-heuristic optimization method for a large-scale problem. The method specifically comprises the steps of dividing an evolution process into a plurality of stages based on a search area, constructing a stage evolution model, and storing an elite solution set of each stage in each layer; a population interaction queue is introduced, and a slave node sends individuals to a master node and stores the individuals in the interaction queue for information exchange among multiple populations; the population advanced state determines whether to explore or develop after entering the next stage; generating a strategy selection probability matrix according to the strategy expression by using a population quality improvement degree-based underlying strategy expression quantification method; and the slave nodes adaptively adjust the strategy according to the probability matrix. According to the method, a master-slave asynchronous communication architecture deeply coupled with algorithm characteristics is used, an upper-layer hyper-heuristic mode can reasonably allocate computing resources according to underlying strategy performance, and the global exploration capability and the solving efficiency of the algorithm are enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent optimization algorithm, and relates to an asynchronous parallel hyper-heuristic optimization method. BACKGROUND

[0002] Hyper-heuristic algorithm is considered as a promising method to solve high-dimensional optimization problems, and has been widely used in recent years. However, in high-dimensional optimization, it faces the problem of high computational cost, so it is necessary to design a parallel hyper-heuristic algorithm. Parallel computing is not only an acceleration method, but also a search paradigm. The optimization framework of hyper-heuristic algorithm is closely combined with the parallel computing model, thereby significantly improving the ability to solve high-dimensional optimization problems in terms of speed and quality.

[0003] The design of efficient parallel hyper-heuristic algorithm faces many challenges and requirements, mainly including: the synchronization of different low-level strategies in parallel processing will cause a large amount of waiting cost of computing units, so an asynchronous parallel computing framework needs to be designed; the asynchronous parallel computing of low-level strategies will cause each strategy to be in different evolutionary computing stages, thereby affecting the fair evaluation of the upper strategy to the bottom strategy. Using parallel computing in the evolutionary stage may be an effective solution to this problem;

[0004] In the asynchronous parallel hyper-heuristic computation allowing parallel processing in the evolutionary stage, improper resource allocation in the evolutionary stage may cause a large waste of computing resources. The population that performs poorly in the current stage does not need to enter the next stage of evolution. The populations in different computing stages in the same evolutionary stage may repeat the exploration of the same area. An efficient hyper-heuristic parallel design needs to consider the performance of the computing nodes and the different evolutionary stages they are in. SUMMARY

[0005] In view of the above, the purpose of the present application is to provide an improved asynchronous parallel hyper-heuristic optimization method, which uses a master-slave asynchronous communication architecture deeply coupled with the characteristics of the algorithm, and the upper hyper-heuristic can reasonably allocate computing resources according to the performance of the bottom strategy, enhance the global exploration ability, and efficiently solve large-scale problems.

[0006] The first aspect of the present application provides an asynchronous parallel hyper-heuristic optimization method.

[0007] An asynchronous parallel hyper-heuristic optimization method comprises the following steps:

[0008] S1, the master node initializes the hyper-heuristic algorithm parameters, including but not limited to: initial population, bottom strategy performance matrix, bottom strategy selection probability matrix, stage evolution model and population interaction queue;

[0009] S2, the master node sends the population to the slave nodes, and the slave nodes randomly select a strategy to start evolution after receiving the population, until the preset communication condition is met, and the master node communicates with the slave nodes in point-to-point mode;

[0010] S3, the master node receives the message from the slave node, and executes the upper-layer super-heuristic strategy:

[0011] (a) updating the population interaction queue according to the slave node message;

[0012] (b) evaluating the performance of the underlying strategy used by the slave node and the evolution stage of the population thereof;

[0013] (c) updating the underlying strategy probability selection matrix and the elite solution set of the pre-entry stage according to the evaluation result;

[0014] (d) according to the population evolution, a small number of individuals are randomly selected from the population interaction queue and sent to the slave node;

[0015] The master node communicates with the slave node;

[0016] S4, after the slave node receives the message from the master node:

[0017] If the master node determines that the population successfully enters the next evolution stage, the slave node selects a new underlying strategy by using the roulette algorithm according to the underlying strategy selection probability of the new stage, and replaces part of the individuals in the current population with the small number of individuals sent by the master node;

[0018] If the master node determines that the population fails to enter the next evolution stage, half of the individuals in the population are replaced by randomly generated individuals, and part of the individuals in the current population are randomly replaced by the small number of individuals sent by the master node, and then a new underlying strategy is randomly selected for evolution, and the master node is communicated when the communication condition is met again;

[0019] S5, repeat steps S3-S4 until the algorithm termination condition is met.

[0020] The S1 is specifically:

[0021] S1.1: start N+1 parallel nodes, including 1 master node and N slave nodes S={S1,...,S i ,...,S N}; the master node initializes N populations and sends them to each slave node;

[0022] S1.2: the master node initializes the stage evolution model;

[0023] Set the difference d between the number of elite solutions in the elite solution set of two adjacent stages;

[0024] According to the number N slaveDetermine the total number of stages H, where H is a positive integer and satisfies d×(H - 1) ≤ N slave ;

[0025] Each evolutionary stage has an elite solution set Q i , Q i The number of elite solutions that Q i can accommodate is denoted as |Q i |, and |Q i | decreases as the stage increases. |Q H | = (i - 1)×d + 1. In particular, |Q i | = 1; The number of existing elite solutions in Q i is q i ; Initialize q

[0026] S1.3: The master node initializes the population interaction queue Queue. Queue is a circular queue, represented by an array with a capacity of capacity, indicating that capacity individuals can be stored;

[0027] The specific content of S2 is as follows:

[0028] When the population in the slave node evolves a certain number of generations, it checks whether the communication condition is satisfied. When any of the following conditions is met, communication is triggered:

[0029] (a) The population in the slave node meets the advancement condition:

[0030] Let the centroid of the population p be represented as (x1, x2,..., x n ) ∈ R n , then the calculation formula for the minimum hypercube region Ri(x, r) of all individuals in the population is as follows:

[0031]

[0032] Among them, r k = max{|up k -x k |, |lp k -x k |}, and that is, up k and lp k are the upper and lower bounds of the k-th dimension of the decision variable x;

[0033] When the population p is in the i-th (1 ≤ i < H) stage, the formula representing that the population meets the advancement condition is:

[0034]

[0035] Among them, ub and lb are the upper and lower bounds of the decision variable x.

[0036] (b) Set the maximum number of generations limitGen allowed between two communications, if the advanced condition is not met and the population is updated with the optimal solution, the communication probability is the ratio of the current number of generations evolvedGen and the maximum number of generations, i.e.

[0037]

[0038] Randomly generate a floating point number pb∈(0,1), if it satisfies pb≤Pb comm , then initiate communication, the slave node sends the optimal individual and randomly selected n individuals to the master node.

[0039] (c) The current number of generations has reached the upper limit value, i.e. evolvedGen=limitGen.

[0040] The S3 is specifically:

[0041] S3.1: The master node receives the message from the slave node, if the slave process satisfies the above communication condition (b), the master node only executes the upper strategy step (a), adds the individual sent by the slave node to the population interaction queue, if the interaction queue is full, the head of the queue is dequeued n+1 individuals, and then re-queued;

[0042] S3.2: If the slave process satisfies the above communication condition (a) or (c), the master node only executes the upper strategy steps (b), (c) and (d);

[0043] The master node calculates the population p of the slave node using the bottom strategy LLH according to the following formula i The quality improvement value of evolution:

[0044]

[0045] Where p' represents the population after evolution of p, And respectively represent the average fitness value of the individuals in the population p and p', and ε is a very small integer to prevent the denominator from being 0;

[0046] According to the calculated performance value, update the bottom strategy LLH i The selection probability in the evolution stage i:

[0047]

[0048] Where, The initial value is 1) represents the strategy LLH i The total number of times selected in the h i Stage;

[0049] S3.3: the population is in the i-th evolution stage and wants to enter the (i+1)-th stage;

[0050] If q i+1 <|Q i+1 |, the optimal solution of the population is directly included in Q i+1 , q i+1 =q i+1 +1;

[0051] If q i+1 ≥|Q i+1 |, the master node compares the optimal solution of the population with the worst solution in the optimal solution set Q i+1 of the (i+1)-th stage; if the optimal solution of the population is better than the worst solution in Q i+1 , it is determined that the population evolution is successful, otherwise, it is determined to fail.

[0052] S3.4: the master node randomly selects n+1 individuals from the population interaction queue and sends them to the slave node.

[0053] The second aspect of the application provides a computer readable storage medium.

[0054] A computer readable storage medium, which stores a computer program, the program being executed by a processor to implement the steps of the asynchronous parallel hyper-heuristic optimization method.

[0055] The third aspect of the application provides a computer device.

[0056] A computer device, which comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the steps of the asynchronous parallel hyper-heuristic optimization method.

[0057] In view of the above, the beneficial effects of the application are: a population evolution degree evaluation method based on population aggregation state is introduced, the evolution process is divided into multiple stages, so that the asynchronous parallel mechanism with hyper-heuristic characteristics can be realized for different evolution stages, and the unfair evaluation caused by different convergence speeds of underlying strategies is solved. At the same time, the asynchronous parallel mechanism also provides a solution to the expensive computing overhead when solving large-scale problems. In addition, a performance quantification method of underlying strategies based on population quality improvement is introduced, which can fairly compete for computing resources for different underlying strategies. Such a competition mechanism can also control the exploration and development balance of the population, and realize survival of the fittest. Finally, the combination of hyper-heuristic characteristics and multi-core asynchronous parallel mechanism has natural diversity, enhances the global exploration ability, and can efficiently solve large-scale problems. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1is a simple flowchart of the method described in the present application;

[0059] Figure 2 is a framework diagram of the method described in the present application;

[0060] Figure 3 is an acceleration ratio and acceleration efficiency fold line chart of the asynchronous hyper-heuristic method described in the present application compared with the synchronous hyper-heuristic method. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0062] Example 1

[0063] The embodiment of the present application provides an asynchronous parallel hyper-heuristic optimization method, comprising the following steps:

[0064] S1, the master node initializes the hyper-heuristic algorithm parameters, including but not limited to: initial population, bottom strategy performance matrix, bottom strategy selection probability matrix, stage evolution model and population interaction queue;

[0065] Specifically, the present embodiment uses IEEE CEC2013 LSGO Benchmark suite to verify the optimization effect of the method, and initializes the dimension D of the decision variable, the upper bound ub and the lower bound lb according to the function number to be tested; the population size pop_size is set to 40; the maximum evaluation number maxFEs is set to (N+1) x 10 6 , N+1 is the number of parallel nodes.

[0066] In the present embodiment, the probability matrix is initialized as M∈R H×num , The bottom strategies used in the present embodiment include: gorilla troops optimization algorithm (GTO, Gorilla Troops Optimizer), crayfish optimization algorithm (COA, Crayfish Optimization Algorithm), prairie dog optimization algorithm (PDO, Prairie Dog Optimization), advanced differential evolution (ADE, Advanced Differential Evolution), whale optimization algorithm (WOA, Whale Optimization Algorithm), differential evolution (DE, Differential Evolution).

[0067] S1.1: Start N+1 parallel nodes, including 1 master node and N slave nodes S={S1,...,S i ,...,S N}; the master node initializes N populations and sends them to each slave node respectively;

[0068] The embodiment adopts Intel Xeon Gold 6458Q processor (32 cores / 64 threads, base frequency 3.1 GHz); the asynchronous parallel mechanism is realized through MATLAB SPMD (Single Program Multiple Data) model.

[0069] The master node and the slave node communicate through 'labSend / labReceive' command; after receiving the population, the slave node randomly selects a strategy for evolution, when the number of slave nodes is greater than the number of underlying strategies, each underlying strategy can be assigned a slave node, and the remaining slave nodes randomly select a strategy for evolution, to ensure that each underlying strategy has computing resources.

[0070] S1.2: The master node initializes the stage evolution model;

[0071] Set the number difference d of elite solutions contained in the elite solution set of adjacent two stages;

[0072] Specifically, according to the number of slave nodes N slave , the number of stages H is calculated, H is a positive integer and satisfies d×(H-1)<N Slave ;

[0073] Each stage has an elite solution set Q i , the number of elite solutions that Q i can accommodate is |Q i |, |Q i | decreases with the increase of the level, in particular, |Q H |=1; the number of elite solutions in Q i is q i , and q i =0 is initialized.

[0074] In the embodiment, d=2 is set, i.e. divided into 16 evolution stages.

[0075] S1.3: The master node initializes the population interaction queue Queue, Queue is a circular queue, represented by an array with a capacity of capacity, which can store capacity individuals;

[0076] Specifically, in the embodiment, capacity=8×(N+1) is set.

[0077] S2, the master node sends the population to each slave node, and each slave node randomly selects a strategy to start evolution after receiving the population, until a specific communication condition is met, and then the slave node communicates with the master node in point-to-point mode;

[0078] Specifically, the population in each slave node is checked for whether the communication condition is met every certain number of generations, and when any of the following conditions is met, the slave node actively communicates with the master node:

[0079] (a) the population in the slave node meets the advanced condition:

[0080] Let the centroid of the population p be represented as (x1, x2,..., x n )∈R n , and the minimum hypercube region R1(x, r) of all individuals in the population is calculated according to the following formula:

[0081]

[0082] where r k = max{|up k -x k |, |lp k -x k |}, and that is, up k and lp k are the upper and lower bounds of the kth dimension of the decision variable x; when the population p is in the ith (1≤i<H) stage, the formula representing that the population meets the advanced condition is:

[0083]

[0084] where ub and lb are the upper and lower bounds of the decision variable x.

[0085] In this embodiment, 10 individuals are randomly selected from the population by sampling to calculate the minimum hypercube region thereof as the minimum hypercube region of the population p, so as to save the calculation overhead.

[0086] (b) in the case where the population in the slave node updates the optimal solution, the number of generations after the last communication has reached the set upper limit; if this condition is met, the slave node sends the optimal individual and random n individuals of the population to the master node;

[0087] Specifically, in this embodiment, n=3, that is, the number of individuals sent is 0.1×pop_size, and the upper limit of the number of generations that the population can evolve after the last communication is set to 0.8×slave_max_gen / level, where slave_max_gen=maxFEs / pop_size / (N-1), and level is the current evolution stage of the population.

[0088] S3. The master node receives the message from the slave node, and executes the upper hyper-heuristic strategy:

[0089] (a) updates the population interaction queue according to the message from the slave node;

[0090] (b) evaluates the performance of the underlying strategy used by the slave node and the evolutionary stage of the population;

[0091] (c) updates the underlying strategy probability selection matrix and the elite solution set of the pre-entry stage according to the evaluation results;

[0092] (d) randomly selects a number of individuals from the population interaction queue according to the population advancement results and sends them to the slave node;

[0093] The master node communicates with the slave node;

[0094] S3.1: The master node receives the message from the slave node, and if the communication condition (b) is met, the master node only executes the upper hyper-heuristic strategy (a), and puts the individuals sent by the slave node into the population interaction queue. If the interaction queue is full, the front n+1 individuals are dequeued and then queued;

[0095] S3.2: If the communication condition (a) and (c) is met, the master node only executes the upper hyper-heuristic strategy (b), (c) and (d);

[0096] The master node calculates the performance of the underlying strategy LLH used by the population p of the slave node according to the following formula i The quality improvement value of evolution:

[0097]

[0098] Where p' represents the population after evolution of p, and respectively represent the average fitness value of the individuals in the population p and p', and ε is a small integer to prevent the denominator from being 0;

[0099] According to the calculated performance value, the underlying strategy LLH is updated i The selection probability at the evolutionary stage i:

[0100]

[0101] Where, (initial value is 1) represents the selection probability of the strategy LLH i in the h i stage;

[0102] S3.3: The population is at the i-th evolutionary stage and wants to enter the i+1 stage;

[0103] If q i+1 <|Q i+1 |, the optimal solution of the population is directly incorporated into Q i+1 , q i+1 =q i+1 +1;

[0104] If q i+1 ≥|Q i+1 |, the master node compares the optimal solution of the population with the worst solution in the optimal solution set Q i+1 of the i+1 stage; if the optimal solution of the population is better than the worst solution in Q i+1 , the population is successfully advanced, otherwise, the population fails to be advanced.

[0105] S3.4: The master node randomly selects n+1 individuals from the population interaction queue and sends them to the slave node.

[0106] S4: The slave node receives the message from the master node, if the master node evaluates that the population successfully enters the next evolution stage, the slave node selects a new underlying strategy using the roulette algorithm according to the selection probability of the underlying strategy of the new stage, and selects a small number of individuals from the individuals sent by the master node to replace the individuals in the current population;

[0107] If the master node evaluates that the population fails to enter the next evolution stage, half of the individuals in the population are replaced by randomly generated individuals, and part of the individuals are selected from the individuals sent by the master node to replace, and a random underlying strategy is selected to continue evolution;

[0108] S5: Steps S3-S4 are repeated until the termination condition is met.

[0109] In order to verify the asynchronous parallel hyper-heuristic optimization algorithm (APHOA) proposed in the application, comparative experiments are conducted on the parallel versions (P-GTO, P-COA, P-ADE, P-PDO, P-WOA and P-DE) of the six underlying strategies, the synchronous parallel hyper-heuristic optimization algorithm (SPHOA) and the algorithm (APHOA) of the application. At the same time, time comparison experiments are conducted on APHOA and the synchronous hyper-heuristic method (SPHOA) using the same underlying strategy.

[0110] In order to ensure the fairness of the experiment, the maximum evaluation times of each algorithm are equal, the number of parallel cores is equal, and the IEEE CEC2013 LSGO Benchmark suite is used. For each function in the test suite, each algorithm is independently run 20 times, and the minimum value, average value and standard deviation after 20 times of algorithm running are represented by Best, Mean and Std respectively. The statistical results are shown in Table 1, and from Table 1, it can be seen that the comprehensive performance of the application algorithm APHOA is obviously better than each bottom strategy and the synchronous super heuristic method, which verifies the effectiveness and superiority of the application algorithm.

[0111] Table 1 Comparison of results of six bottom strategies, synchronous super heuristic methods using the same bottom strategy set and the super heuristic method proposed in the application

[0112] The parallel comparison test of the application algorithm APHOA and SPHOA under different CPU core numbers is shown in the following table 2. Figure 2 Figure 2 The speedup ratio and acceleration efficiency of APHOA and SPHOA under different parallel core numbers are shown in the following line chart.

[0113] Embodiment two

[0114] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps in the asynchronous parallel super heuristic optimization method in the above-mentioned embodiment one.

[0115] Embodiment three

[0116] The embodiment provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor realizes the steps in the asynchronous parallel super heuristic optimization method in the above-mentioned embodiment one when executing the program.

[0117] In summary, the application algorithm APHOA has the following advantages:

[0118] In order to solve the problem of large-scale problem space and complexity, the super heuristic adopts a master-slave asynchronous communication architecture deeply coupled with the algorithm characteristics, encapsulates the bottom algorithms with different advantages to increase diversity;

[0119] Based on the search area, the evolution process is divided into multiple stages, a stage evolution model is constructed, and whether the population advancement is successful determines the search strategy of the next step of the population, and meanwhile, information exchange can be carried out among multiple populations.

[0120] ​For the fair scheduling of underlying strategies in hyper-heuristic, APHOA designs a performance quantification method of underlying strategies based on the improvement degree of population quality, and calculates the selection probability of each underlying strategy according to the strategy performance;

[0121] Combined with the characteristics of asynchronous parallel and the above strategy selection method, each slave node can adaptively select the underlying strategy, thereby developing different regions.

[0122] Therefore, compared with a single strategy, APHOA has better diversity and exhibits more excellent performance; the asynchronous parallel architecture effectively reduces the time required for traversing the complex solution space through concurrent exploration. At the same time, the architecture naturally fits the master-slave parallel paradigm, providing strong support for the design and implementation of hyper-heuristic algorithms.

[0123] The specific embodiments of the application are described in detail above with reference to the accompanying drawings, but the application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. An asynchronous parallel hyperheuristic optimization method, running on a master node and at least one slave node, characterized in that, It includes the following steps: S1. The master node initializes the parameters of the hyper-heuristic algorithm, including but not limited to: the initial population, the underlying policy performance matrix, the underlying policy selection probability matrix, the stage evolution model, and the population interaction queue; S2. The master node sends the population to the slave nodes. After receiving the population, the slave nodes randomly select a policy to start evolution until the preset communication conditions are met, and then conduct point-to-point communication with the master node; S3. The master node receives the messages from the slave nodes and executes the upper-layer hyper-heuristic strategy: (a) Update the population interaction queue according to the messages from the slave nodes; (b) Evaluate the performance of the underlying policy used by the slave node and the evolution stage of its population; (c) Update the underlying policy probability selection matrix and the elite solution set of the pre-entry stage according to the evaluation results; (d) Randomly select a small number of individuals from the population interaction queue according to the population advancement situation and send them to the slave nodes; The master node communicates with the slave node; S4. After the slave node receives the messages from the master node: If the master node determines that the population has successfully entered the next evolution stage, the slave node selects a new underlying policy using the roulette algorithm according to the underlying policy selection probability of the new stage, and randomly replaces some individuals in the current population with the small number of individuals sent by the master node; If the master node determines that the population fails to enter the next evolution stage, half of the individuals in the population are replaced by randomly generated individuals, and some individuals in the current population are randomly replaced by the small number of individuals sent by the master node. Then, a random underlying policy is selected to continue evolution, and communication with the master node is conducted when the communication conditions are met again; S5. Repeat steps S3 - S4 until the algorithm termination conditions are met.

2. The method as described in claim 1, characterized in that, In the above S1, the initialization of the stage evolution model includes: Set the difference d between the number of elite solutions in the elite solution sets of two adjacent stages; Based on the number of nodes N slave Determine the total number of stages H, where H is a positive integer and satisfies d×(H-1)≤N slave ; Each evolutionary stage has an elite solution set Q. i Q i The number of elite solutions that can be accommodated is denoted as |Q i |,|Q i | Decreases as the stage progresses,| Q i |=(i-1)×d+1, in particular, we have |Q H |=1;Q i The number of existing elite solutions is q. i Initialize q i =0.

3. The method as described in claim 1, characterized in that, In the above S2, the specific communication conditions include any of the following situations: (a) The population in the slave node meets the advancement conditions: Let the centroid of population p be represented as (x1, x2, ..., x...). n )∈R n The formula for calculating the minimum hypercube region Ri(x,r) for all individuals in the population is as follows: Where, r k =max{|up k -x k |,|lp k -x k |}, and up k and lp k Let x be the upper and lower bounds of the k-th dimension of the decision variable; When the population p is in the i-th (1 ≤ i < H) stage, the formula representing that the population meets the advancement conditions is: where, ub and lb are the upper and lower bounds of the decision variable x. (b) Set the maximum number of evolution generations limitGen allowed for the slave node between two communications. If the advancement conditions are not met and the population updates the optimal solution, the communication probability is the ratio of the current evolved generation evolvedGen to the maximum generation, that is A random floating-point number pb∈(0,1) is generated. If pb≤Pb comm If so, then communication is initiated. (c) The current number of evolution generations has reached the upper limit value, that is, evolvedGen = limitGen.

4. The method as described in claim 3, characterized in that, When the communication condition (b) is met, the slave node sends the optimal individual and several randomly selected individuals in the population to the master node; after receiving them, the master node only executes the upper-layer policy step (a) to update the population interaction queue, that is, add these individuals to the queue; if the queue is full, perform the dequeue operation first and then enqueue.

5. The method as described in claim 3, characterized in that, When the communication condition (a) or (c) is met, the slave node does not send individuals to the master node, and the master node only executes the upper-layer policy steps (b), (c) and (d).

6. The method as described in claim 1, characterized in that, In the above S3, the method for evaluating the performance of the underlying policy used by the slave node is: In the evolutionary stage h i In the middle, when population p adopts strategy LLH i When moving to the next stage, the improvement in population quality is defined as: Where p' represents the population after p has evolved. and Let p and p' represent the average fitness values ​​of individuals in populations p and p', respectively, where ε is a very small integer to prevent the denominator from being 0; In the evolutionary stage h i In the middle, the underlying strategy LLH i The performance is defined as: in, (Initial value is 1) indicates the strategy LLH i In h i The total number of times a stage was selected.

7. The method as described in claim 1, characterized in that, In the above S3, the method for updating the underlying policy probability selection matrix is: where, num is the total number of underlying policies.

8. The method as described in claim 1, characterized in that, In S3, the method for assessing the evolutionary stage of the population is as follows: The population is in the i-th evolutionary stage and wants to enter the (i+1)-th stage. If q i+1 <|Q i+1 If | then the best individual in this population is directly included in Q. i+1 q i+1 =q i+1 +1; If q i+1 ≥|Q i+1 The master node combines the population's optimal solution with the optimal solution set Q for stage i+1. i+1 Compare the worst-case solutions in the population; if the optimal solution of this population is better than Q... i+1 If the worst solution is found, the advancement is considered successful; otherwise, it is considered a failure.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the asynchronous parallel hyperheuristic method as described in any one of claims 1-8.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the asynchronous parallel hyperheuristic method as described in any one of claims 1-8.

Images