Improvement Strategies For Particle Swarm Optimization Algorithms With Applications

Swarm intelligence algorithm is a type of heuristic search method.Since the optimization mechanism of swarm intelligence algorithms does not rely too much on the organizational structure information,these algorithms can be widely used in function combination optimization and calculation.Specifically,when solving some optimization problems with non-linear,non-convex or non-differentiable objective functions,the swarm intelligence algorithms can show good search performance.Acutally,the swarm intelligence algorithm is a family which contains many optimization algorithms,and the particle swarm optimization(PSO)algorithm is one of them.Due to its characteristics of easy implementation and fast convergence speed,the PSO algorithm is considered to be a superior candidate approach for solving many practical application problems,and thus it has been widely used in various kinds of scientific research work and industrial production related to artificial intelligence.However,several issues may occur when swarm intelligence algorithms are used to solving complex optimization problems(i.e.,protein-ligand docking),such as premature convergence,insufficient accuracy,and weak efficiency.Therefore,how to design improvement strategies for swarm intelligence algorithms,especially for the PSO algorithm and its variant algorithms,to further enhance the performance and efficiency of such algorithms when solving complex optimization problems,is one of the current research hotspots in this field.The random drift particle swarm optimization(RDPSO)algorithm is a variant of PSO,which has been proven to have better search performance than most variants of the PSO algorithm.Moreover,the RDPSO algorithm also has some advantages such as simple framework,few parameters,and easy implementation.Therefore,this dissertation selects the RDPSO algorithm as our key research object,and designs a series of improvement strategies based on the RDPSO algorithm,which effectively improves the performance and efficiency of RDPSO in solving complex optimization problems.The modified version of the algorithm based on the various strategies we proposed has also been used to solve one complex optimization problem,namely protein-ligand docking,in this dissertation.The experimental results verified that our methods can significantly improve the docking accuracy and docking efficiency of the molecular docking software.In addition,some of the improvement strategies have a certain degree of versatility,which can be extended to other PSO variants.The specific research work in this dissertation is as follows:(1)Firstly,this dissertation analyzed the diversity measurement methods of swarm intelligence algorithms.This dissertation discussed two classical diversity measurement schemes.According to the experimental results,we have determined that the genotype diversity measurement based on average point distance is more suitable than the entropy-based phenotypic diversity measurement for analyzing the convergence of swarm intelligence algorithms.Furthermore,this dissertation also improved the average-point-distance-based diversity measurement and proposed a normalized definition of the swarm diversity.This improved diversity measurement can be applied to the problems that the magnitudes of different dimensions are not the same.(2)Secondy,this dissertation designed a two-stage diversity guided(2PDG)strategy for the RDPSO algorithm.This dissertation analyzed one canonical diversity guided strategy named attract-repulsive(AR)strategy,and found that this strategy sometimes may stay in the repulsive stage for a long time and cannot be applied to some multimodal problems.To solve these problems and to design a novel strategy for the RDPSO algorithm,this dissertation proposed the 2PDG strategy,which includes a declining diversity lower bound,a parameter-controlled repulsion stage,and an accelerated convergence operation in the attraction stage.Experimental results illustrated that the 2PDG strategy can effectively solve the above-mentioned shortcomings of the AR strategy,but it is not as effective as the AR strategy on the multimodal problems where the AR strategy does work.Nevertheless,the RDPSO algorithm based on the 2PDG strategy can obtain better search performance than that based on the AR strategy on most of the multimodal problems.(3)Thirdly,this dissertation proposed a diversity collaboratively guided(DCG)strategy for the RDPSO algorithm.Since the 2PDG strategy mentioned above has some shortcomings such as complicated parameter settings,slow algorithm convergence speed,and unreasonable operations,in this dissertation,we introduced a novel diversity guided strategy which makes two kinds diversity measures work collaboratively.The experimental results and the corresponding analysis firstly proved the necessity and effectiveness of the settings of various stages in the DCG strategy,and secondly verified that the DCG strategy is more robust and has better ability of balancing the global and local search abilities than the 2PDG strategy when solving the most multimodal problems.(4)Fourthly,in this dissertation,we proposed a general diversity collaboratively guided(GDCG)strategy for most of the swarm intelligence algorithms.In the GDCG strategy,the concept of the virtual attractor was introduced to guide the swarm to perform different search behaviors,thereby avoiding the problem of parameter settings at different stages in the DCG strategy,and then makes the framework of the DCG strategy can be successfully applied to other swarm intelligence algorithm.Another advantage of the GDCG strategy is that the position settings of the virtual attractors at each stage are related to the search range of each dimension of the problem,which makes the GDCG strategy suitable for search ranges of different sizes.The experimental results demonstrated that the GDCG strategy can effectively improve the search performance of many swarm intelligence algorithms when solving most multimodal problems.(5)Moreover,this dissertation also proposed a multi-swarm parallel(MSP)strategy for the RDPSO algorithm,to enable the search process of the RDPSO algorithm to be executed in parallel.The framework of the MSP strategy divides the entire swarm evenly into several sub-swarms,and assigns feature components to each sub-swarm.The algorithm based on the MSP strategy realizes the information sharing between sub-swarms by exchanging the feature components every certain number of iterations,so that the entire search process of the algorithm can be executed in parallel.The experimental results illustrated that the RDPSO algorithm based on the MSP strategy can obtain better results than the canonical RDPSO algorithm and the RDPSO based on the classical island model,which verified the effectiveness of the MSP strategy in improving the search performance of the algorithm.(6)Finally,this dissertation makes the algorithms based on the above strategies integrated with a local search method,which effectively improves the accuracy and efficiency of molecular docking.Specifically,several versions of the RDPSO algorithms based on the aforementioned diversity guided strategies and the MSP strategy were integrated with the pseudo Solis&Wets local search method in this dissertation,in order to solve the normal docking and blind docking problems.The experiments in terms of molecular docking in this dissertation have drawn several important conclusions as follows: a)The hybrid algorithm based on the DCG strategy can obtain the best binding-energy results and the best docking accuracy among all the compared docking methods for normal docking problems;b)The hybrid algorithm based on the GDCG strategy can achieve good docking performance for both normal docking and blind docking,and benefit from the good adaptability of the GDCG strategy to different sizes of search ranges,the docking accuracy on blind docking obtained by this algorithm is far superior to other compared docking methods;c)After modifying the implementation of the local search algorithm,the hybrid algorithm based on the MSP strategy can be fully parallelized based on the number of sub-swarms,which can significantly improve the efficiency of normal docking under the premise of ensuring the docking accuracy.In conclusion,this dissertation proposed several novel diversity guided strategies and one novel multi-swarm strategies for swarm intelligence algorithms,which can significantly improve their search performance and efficiency.Furthermore,this dissertation further improved the swarm intelligence algorithms based on the aforementioned strategies to make them be high-precision and high-efficiency solutions for the protein-ligand docking problems.Therefore,it can be summarized that the research work in this dissertation is of value in the academic field and real applications.