Research On Evolutionary Algorithms Combining Historical Information

The evolutionary algorithms have been widely used in various optimization problems,but the existing evolutionary algorithms rarely use historical individuals and their spatial distribution information,so they have great randomness.This makes the algorithm inefficient in search and needs to calculate a large number of fitness values,which is not suitable for costly optimization problems such as time-consuming.Although the evolutionary algorithm based on the agent model uses historical individuals to construct the agent model to calculate approximate fitness values,the accuracy of the agent model will gradually decrease as the complexity of the problem increases.To this end,this thesis combines historical spatial information and proposes an evolutionary algorithm framework based on distance measurement.The framework increases the interaction probability of individuals in each region through a hierarchical strategy to accelerate the generation of good individuals,and samples according to the distance measurement results to replace the poor individuals in the current population.This framework fully combines historical spatial information,which effectively improves the convergence speed and solution accuracy of evolutionary algorithms.This thesis further analyzes the advancement and feasibility of the evolutionary algorithm combined with historical information by using genetic algorithm as an example,and proposes a multi-granularity genetic algorithm through hierarchical strategy,improved genetic operation and multi-granularity spatial strategy.This algorithm applies the hierarchical strategy to the current population and performs corresponding genetic operations to accelerate the generation of excellent individuals.Subsequently,the algorithm divides the feasible domain into multiple subspaces through the multi-granularity spatial strategy based on the completely random tree,and increases the search intensity of the sparse space and the subspace where the current optimal solution is located according to the results of the space partition.The experimental results show that the multi-granularity genetic algorithm can improve the convergence speed and the accuracy of the solution,reduce the number of iterations to reach the convergence condition,and thus increase the adaptability of the algorithm to optimization problems such as time-consuming.In order to improve the low efficiency problem caused by the accurate spatial analysis of multi-granularity spatial strategy,this thesis approximates the whole feasible region through clustering,so as to propose an improved genetic algorithm based on k-means.The algorithm uses the spatial partitioning strategy based on clustering to analyze the space,classify historical individuals to obtain the approximate spatial partitioning results,and accelerates the search probability of the sparse space and the subspace where the current optimal solution is located.Although the convergence speed and solution accuracy of this algorithm are slightly lower than the multi-granularity genetic algorithm in some cases,the time needed for space partition is reduced.