Research And Application On Two Kinds Of Evolutionary Algorithms

As two main strategies of human adaptation to environment,Evolution and learning are providing effective means for making artificial intelligence a reality.In modern artificial intelligence technology,evolutionary computation and machine learning have become the two major directions in the field of artificial intelligence.Evolutionary algorithms originated from the simulation of biological evolutionary factors and the phenomenon of natural population,and produced a system that adapts to the environment by disturbance and elimination.Machine learning is based on previous experience,and generalizes a model that is suitable to predict future events by minimizing the risk of error.The object of this paper is the evolutionary algorithms.Because of the large number of black box optimization problems in the real world,evolutionary algorithm is very important in scientific research,engineering practice and production management.In this paper,we have some further studies about the evolutionary algorithms,and then propose a more reasonable evolutionary algorithm model and strategy.Besides,we also discuss the combination of evolutionary algorithm and machine learning.The main content and innovative research include the following aspects:In the aspect of single object optimization,this paper first takes the fireworks algorithm(FWA)as the research carrier,and puts forward some improvements to its existing problems.Note that the FWA selects the individual according to the density of solutions,which has some certain reasons,but it is easy to cause the marginalization of solutions,that the individuals always prefer to run to the edge direction and eventually lead to the poor quality of the solution.This is essentially a preference choice problem.Because only considering the influence of one factor to select the next generation of individuals,it will lead to the occurrence of preference choice.This is the "no free lunch" theorem.Therefore,we propose an effective direct solution.In the update next generation solution,should not only consider the effect of a single factor,but also the multiple factors or indicators.Such as considering the density factor and the function value,we do not consider only the one of them,but instead both of them,which can force the interaction between factors or indicators,and achieve the diversity effect in the force.This paper is called the MIS-FWA(Multi-Indicator Selection Mechanism FWA).In recent years,multi-objective optimization algorithm is attracting more and more attention because of its practicability and universality.The MOEA/D which is based on multi-objective decomposition is an important branch of multi-objective optimization algorithm.The key problem of multi-objective optimization is how to achieve a good balance between convergence and dispersion of the algorithm,specifically,is the selection of aggregate function.At present,there are different advantages and disadvantages in the aggregation function of the mainstream algorithm.In this paper,an improved aggregation function is proposed.The aggregation function is a quadratic function,and the contour line of a population under the function is a quadratic curve,so it is named HYB method.Commonly used aggregate functions,such as PBI and Chebyshev,may encounter some disadvantages on some problems because of their discontinuities.The proposed quadratic aggregate function is a generalization of them,makes it continuous and increases its rationality at the same time.Compared with the PBI method,the HYB method emphasizes the convergence,and can be easier to balance between the convergence and dispersion.We also have put forward some application examples of its algorithms.In the machine learning problems,we have put forward the application of MOEA/D based on our generalization to solve these problems,providing some feasible solutions for the combination.Finally,we carried out simulation experiments on the algorithms mentioned above.The single objective aspect is compared with SPSO,EFWA and EFWA-AOBL and the multi objective aspect is compared with MOEA/D-PBI,MOEA/D-IPBI,and NSGA-III.The experiments show that our proposed algorithm has its validity,in most examples are better than other algorithms,a few examples have similar performance.