Researches And Applications Of Random Drift Particle Swarm Optimization

With the informatization,networking and intelligence of human society,a large number of optimization problems have emerged,and these problems are generally solved by search method in computers.However,the traditional brute-force search method is time-consuming and labor-intensive,and cannot cover the entire solution space so that remarkable solutions cannot be obtained.Swarm intelligence algorithm is a general term for heuristic search methods,most of which simulate the social cooperation behavior of animals in nature.An individual in a swarm can be regarded as an intelligent body,which decides the next search by learning from its individual and social experience.Particle swarm optimization(PSO)algorithm is a typical swarm intelligence algorithm,which is widely used in scientific researches and engineering practices because of its simple formulation and implementation.However,in practical applications,especially when dealing with multi-modal problems,the PSO algorithm also has some shortcomings.Firstly,the population diversity is poor.Secondly,the PSO algorithm is easy to fall into local optimum.Thirdly,the PSO algorithm converges relatively slowly.To overcome these problems,researchers have proposed many improvement strategies and PSO variants.Common improvement strategies include the hybridization of different algorithms,multi-population evolution,parameter control,and neighborhood topology.Random Drift Particle Swarm Optimization(RDPSO)is a novel variant of the PSO algorithm,and it’s demonstrated that the RDPSO algorithm generally has better performance than PSO algorithm and most other PSO variants.In order to meet the different requirement of practical problems,this paper applies corresponding improvement strategy to the RDPSO algorithm in order to improve its search performance and optimization effect.The research contents in this paper are as follows:(1)This paper proposes a hybrid algorithm of RDPSO and MCMC(Monte Carlo Markov Chain).The hybrid algorithm is divided into two stages.The first stage mainly executes the RDPSO algorithm,and the second stage executes the Markov mutation operation of MCMC.It has been verified on the multimodal function of CEC-2013 that the hybrid algorithm has has better performance than PSO and its four variants.In the applied research,the proposed hybrid algorithm is used to deal with protein-ligand docking problems that require high efficiency.In order to improve the docking efficiency,when gbest is not updated for a long time,the algorithm performs the second stage.The second stage selects the personal best(pbest)positions of the top-ranked particles,and performs the Markov mutation operation to update global best(gbest)position as soon as possible.This operation greatly enriched conformational results.The re-docking and cross-docking performances of hybrid algorithm are tested On the PDBBind-Core core-set and the Sutherland-crossdock-set dataset,respectively.The results show that the hybrid algorithm has very good docking accuracy.(2)This paper proposes a multi-swarm architecture using the master-slave model,which is inspired by the symbiotic relationship in nature.The master and slave subswarms are regarded as two species that are interdependent and cooperate with each other.The slave subswarms accomplish the global exploration for the search space,and the master subswarm exploits for the remarkable solutions around the slave subswarms’ results.The co-cooperation between two species is mainly through the information exchange between the gbest position of the slave subswarms and the pbest position of the master subswarm.In the PSOVina,the PSO algorithm with chaotic embedding has strong global exploration ability,so it is used for the search algorithm of slave subswarms.The standard PSO and RDPSO algorithms are used for the search of the master subswarm,respectively,and propose two multi-swarm algorithms.Tested on the multi-peak functions of CEC-2013,the results show that the performance of multi-swarm algorithm is excellent compared with PSO and its four variants.The performance of re-docking,cross-docking and virtual screening was tested on the three datasets PDBBind-Core,Sutherland-crossdock-set and DUD-E,respectively.It’s demonstrated that the multi-swarm algorithm is better than the MCMC,PSO and GWO algorithms in terms of docking accuracy,and it shows stable docking performance regardless of using sufficient or insufficient CPUs for parallel implementation.(3)This paper proposes a novel parameter setting method.The global parameter of algorithm is defined as a product of the linear decreased value and a novel scaling factor.The novel factor can adaptively scale the parameters to appropriately increase the globality or locality of the algorithmic search.In order to further improve the optimization ability of the algorithm,the pyramid neighborhood structure is also introduced into the algorithm.Tested on the multi-peak functions of CEC-2013,the results show that the scaling factor and pyramid neighborhood can both improve the optimization ability of algorithm.This paper also presents a new network for pneumonia diagnosis and segmentation,which can complete the COVID-19 diagnosis and lesion area segmentation at the same time.However,we found that the diagnostic accuracy of the network still has certain fluctuations.In order to stabilize the diagnostic accuracy,a XGBoost classifier is added into the joint network.But,the XGBoost parameters are relatively complex and it is difficult to find an excellent parameter combination.Therefore,the RDPSO algorithm with scaling factor and pyramid neighborhood is used to optimize the XGBoost parameters.As a result,the proposed algorithm found an excellent parameter combination,and using it for the XGBoost classifier can stablize the diagnostic accuracy of the network.To sum up,this paper presents three improvements for the RDPSO algorithm,i.e.,the hybridization of MCMC and RDPSO,the multi-population strategy of the master-slave model and the adaptive RDPSO algorithm using the pyramid neighborhood.These algorithm improvements meet the needs of practical problems.For the molecular docking problem,the hybridization and multi-population strategies of algorithm have fast convergence speed and excellent population diversity,which can effectively improve the actual docking accuracy.In the XGBoost parameter optimization,the RDPSO algorithm using the pyramid neighborhood and scaling factor has strong optimization capabilities.In this way,an excellent combination of parameters can be obtained for XGBoost classifier,which stablizes the diagnostic accuracy of nerual network.