Research On Parallel Algorithm For Multi-objective Invasion Tumor Growth Optimization

Many problems in daily life and scientific research engineering can be transformed into multi-objective optimization problems through modeling.Multi-objective evolutionary algorithm is an effective method to solve these problems while the scale of calculation is large and has a heavy computation cost.Invasive Tumor Growth Optimization Algorithm for Multiobjective(VITGO)proposed by our lab is a new swarm intelligence algorithm based on the growth mechanism of tumor cells.It simulates the transformation and migration of tumor cells in the nutritional environment to solve the optimization problem.Meanwhile,it has the characteristics including high solving accuracy,good stability and extensibility.However,due to the high computational complexity of VITGO,the computational efficiency needs to be improved.The parallel acceleration can be realized through parallelization and distributed computing.Therefore,this paper focuses on the distributed parallel design and implementation of multi-objective invasive tumor optimization algorithm.The main work of this paper is as follows:(1)Multi-objective invasive tumor growth optimization algorithm(VITGO)draws from the vascular growth mechanism of tumor cells and uses blood vessels to guide the growth of tumor cells for achieving multi-objective optimization.The method based on crowding distance is used to replace the method based on Manhattan distance in the original VITGO for endpoint selection and generation.The pareto front solution of the optimized VITGO were better than the original algorithm on most benchmark test functions while the convergence speed compared with the original algorithm has certain ascend.(2)HP-VITGO,a high-performance parallel algorithm based on GPU and multi-core CPU hybrid parallelization is designed and implemented.Each type of tumor search process in optimized VITGO algorithm is parallelized based on coarse-grained master-slave model,and the parallelization is implemented based on multi-core CPU.The construction of nondominated solution set in VITGO algorithm is parallelized based on fine-grained master-slave model,and the parallelization is implemented by GPU.By parallelizing and accelerating the two most time-consuming parts of the algorithm,the time cost of optimized VITGO algorithm on stand-alone system is effectively reduced.(3)SHP-VITGO,a distributed parallel algorithm based on Spark platform is designed and implemented.In order to make better use of GPU cluster to accelerate the algorithm,the algorithm in this paper is parallelized by hierarchical model.The upper layer of the model is the island model implemented based on Spark,and the lower layer is HP-VITGO.The initial population was divided into several subpopulations,which were distributed to each island by Spark RDD and evolved by using the computing resources of each host.Finally,the population was merged.The distributed cluster computing resources were used to further improve the speed up of the algorithm.Moreover,due to the diversity of population brought by distributed computing,the quality of solution has been further improved.