Study Of The Adaptive And Distributed Evolutionary Multi-task Optimization Algorithm

Evolutionary algorithm(EA)stands for a cluster of population-based meta-heuristic optimization algorithms that simulates the natural biological evolution,such as genetic algorithm and particle swarm algorithm.Unlike the classical mathematical optimization methods,EA has the characteristics of low problem dependence and global optimization,has been widely used to solve difficult practical optimization problems.The traditional EA is only used to solve a single task and cannot utilize the relatedness of different tasks.In recent years,an emerging multi-task optimization paradigm,Evolutionary multi-task optimization(EMTO)has been proposed to solve multiple optimization tasks simultaneously by EA.The useful knowledge gained from solving one task can help solve another task,thereby improving the optimization performance of each component task.This article mainly studies how to improve the overall optimization efficiency of multi-task optimization through well-designed knowledge transfer strategies,and how to use distributed computing equipment to accelerate EMTO.As far as we know,this thesis is the first work to study the parallel EMTO framework.The main research work includes the following three aspects:(1)We propose an adaptive knowledge transfer algorithm to improve the overall efficiency of multi-task optimization.Compared to the existed algorithms,the proposed algorithm can adaptively adjust the frequency of knowledge transfer according to the efficacy of optimization of task itself and the efficacy of knowledge transfer,and at the same time,when faced with multiple sources of knowledge for selecting,adaptive knowledge sampling is used according to the helpfulness of different candidate tasks.Experiments have proved that in 21 test problems,the adaptive knowledge transfer algorithm proposed in this paper can effectively improve the overall optimization efficiency.Compared with other state-of-the-art EMTO algorithms,the proposed algorithm is superior in both solution accuracy and run time.(2)We first analyze the difficulties in designing the parallel framework for EMTO and propose a concise parallel framework.In this framework,each task is assigned two threads(in a process),the main thread is used for task solving and synchronization with other tasks,and the auxiliary thread is responsible for sending the knowledge of the current task to other knowledge requesters(other tasks).An asynchronous communication strategy is designed to ensure the efficiency of communication between tasks.Experiments have been verified on several multitask optimization problems.The results show that there is no significant difference between the parallel algorithm and the serial algorithm in solving accuracy.Parallel algorithm has a significant acceleration effect for problems with computationally expensive component tasks.(3)We analyze the drawbacks of the parallel framework proposed in(2)for solving the optimization problem of tremendous component tasks(e.g.,thousands of component tasks),and propose a scalable and parallel framework.In this framework,each computing node is composed of multiple CPU cores,one computing node can solve multiple tasks,and multiple computing nodes jointly solve all tasks.When the number of tasks exceeds the computing resources of each computing node,multiple tasks share the computing node's resources serially.In the experiment,32 computing nodes(each node consists of 8 CPU cores)are used to reduce the running time of a multi-task optimization problem with 2000 tasks from 11 hours to 0.4 hours with the similar solving accuracy of serial algorithm.