Research Of High Performance Evolutionary Algorithm Based On Distributed Parallel Computing

With the popularity of the internet and computers,human society has entered the era of big data.Since the increase of data,the scale of the problem has gradually increased and many practical problems have also become large-scale optimization problems.As the scale of the problem increases,the number of decision variables increases too.The “dimensional disaster” makes these methods which used to slove traditional problems hard to work.Evolutionary algorithm is a classic heuristic algorithm,it has the characteristics of easy understanding,fast convergence and strong search ability.It performs well in solving low-dimensional and smallscale problems.When the scale of the problem increases,evolutionary algorithm's efficiency has also been a sharp decline too.Faced with this situation,combining the cooperative coevolution framework with the evolutionary algorithm is a current popular solution.They decompose the large-scale optimization problem into multiple smaller sub-problems.The solution process can be divided into three stages: problem decomposition,sub-problem optimization and integration of optimization results.Due to the complexity of large-scale optimization problems,its overall solution is very time-consuming.At present,there are many researches on how to improve the algorithm aiming on optimizating results,but the research on how to reduce the solution time of each stage and improve the efficiency of algorithms is very limited.This paper takes the large-scale optimization problem as the research object,and select an improved differential grouping algorithm(Differential Grouping 2,DG2)as the decomposition method.In addition,Cooperative Coevolution Quantun-behaved Particle Swarm Optimization(CCQPSO)is used as the optimization algorithm in the problem solving stage.By analyzing the feasibility of parallelization of DG2 and using the natural parallelism of co-evolutionary algorithm and particle swarm optimization algorithm,the paper realized the parallelization of large-scale optimization problem in the decomposition stage and sub-problem optimization stage.The main work of this paper is as follows:1)The parallel DG2(Parallel Differential Grouping 2,P-DG2)based on OpenMP in shared memory model is studied.This parallel algorithm uses the OpenMP framework to improve and parallelize the traditional serial DG2 grouping algorithm.It is more in line with the parallel program concept and realizes parallel acceleration calculation of any multi-core.The experimental results on the IEEE CEC'2013 LSGO test function show that the P-DG2 algorithm reduces the running time effectively without affecting the grouping result and the algorithm achieves good acceleration ratio and efficiency performance.2)The Parallel-Cooperative Coevolution Quantun-behaved Particle Swarm Optimization(P-CCQPSO)based on MPI in the message communication model is studied.The algorithm builds a topological model which adopts the coarse-grained strategy and combines the cooperation co-evolution with the MPI to ensure the convergence effect of the algorithm.At the same time,the algorithm is further improved by introducing secondary grouping strategy and neighborhood update strategy to raise the overall performance of the algorithm.Experiments show that the improved parallel algorithm has better optimization results on large-scale optimization problems.3)The parallelization of the evolutionary algorithm in the domestic supercomputer Sunway·TaihuLight is studied.The domestic CPU adopts the construction of the master-slave core,and the calculation uses the MPI+OpenACC(process parallel + thread parallel).In the framework of this hybrid parallel computing model,the paper studies the performance of traditional evolutionary algorithm and CCQPSO algorithm in solving high-dimensional problems.Experiments show that the performance of CCQPSO is further improved by largescale parallel,and it has great parallel scalability.By combining the evolutionary algorithm and some characteristics of parallel computing,this paper parallelizes the algorithm in the grouping,solving and computing stages of solving large-scale optimization problems and analyses its efficiency and speedup performance.Experiments show that the algorithms after parallelization greatly reduce the execution time of the algorithm while ensuring the correctness and it shows good performance on the 1000-dimensional large-scale optimization problem.On some issues,the parallelization algorithm based on the decomposition strategy can improve the optimization performance of the algorithm and promote acceleration and optimization of results at the same time.In addition,through the massive parallelism of the supercomputer SunWay Taihu-Light,the distributed parallelism of evolutionary algorithms still has excellent performance.Experiments prove that there are still many work worthy of research in solving large-scale optimization problems.