Researches On Improved Dynamic Differential Evolution Algorithm Based On Reinforcement Learning

Differential evolution algorithm is a very potential branch of evolution algorithms.As a population-based algorithm,its implementation is simple and intuitive,with low space complexity,good global search ability,and is more suitable for large-scale problems.Differential evolution algorithm can solve static problems well,and it also has good performance in solving dynamic optimization problems.However,there are still some improvements in the theory and application of dynamic differential evolution algorithm.In order to improve the local search ability,diversity and convergence speed of dynamic differential evolution algorithm,this paper proposes the corresponding improvement methods,which mainly includes the following two aspects:(1)An improved dynamic differential evolution algorithm is proposed.On the basis of the traditional distance based exclusion scheme,hill-valley function is introduced to track the adjacent peaks to avoid missing the peaks close to other peaks.At the same time,the neighborhood search technology is designed to enhance the ability of local area search,divide the neighborhood space of the best individual of the population,form a set of candidate solutions in multiple ranges,and update the best individual of the population with the optimal solution iteratively.The experimental results show that the algorithm is more effective than unimproved algorithm and comparative algorithms in overall performance,and it is feasible to solve the dynamic optimization problems.(2)Combining reinforcement learning algorithm with dynamic differential evolution algorithm,each individual in the population is regarded as an agent.The ratio of diversity in population and relative fitness value are used to encode state variable.Three typical DE mutation operations are optional actions of the agent,and a reward function is designed to guide the movement of the population.According to the reinforcement learning experience represented by the corresponding value of Q-table,each agent can adaptively select a mutation strategy to generate offspring.The experimental results show that the combination of Q(?)algorithm in reinforcement learning and dynamic differential evolution algorithm can overcome the randomness of differential evolution algorithm and pave the way for the combination of dynamic differential evolution algorithm and other reinforcement learning algorithms.