Research On Enhancing Diversity Oriented Benchmark Construction And Evolutionary Algorithm Design

As a class of essential optimization methods,evolutionary algorithms have been widely applied in solving many single objective continuous optimization problems due to their advantages of simple implementations and excellent effects.However,the No Free Lunch Theorem reveals that the performance of the optimization algorithm is related to the matching degree between it and the tested problem.Real problems provide most accurate tests yet with expensive costs,thus we generally use artificially designed problems.The artificial problem can reflect the real problem as much as possible by analyzing and constructing the real problem properties on artificial ones,thus the desirable performance of the evolutionary algorithm on the artificial problems is expected to be generalized into the real problem domain.However,facing of the endless complex and diverse real problems,the algorithm design and problem design in the field are suffering from great challenges.On the one hand,most evolutionary algorithms are not designed specifically based on the analysis of known problem characteristics,thus it is difficult to effectively deal with these complex real problem properties.On the other hand,compared with the entire real problem domain,the existing test problems are insufficient in diversity and complexity to substantially detect the comprehensive performance of the algorithm.Inspired by hybrid intelligence,this paper makes some explorations in the construction of test problem and the design of evolutionary algorithm by integrating the ideas and technologies such as neural network,diversity,and local search.The main research contributions are summarized as follows:1.From the perspective of test problems,a novel neuro-diversified problem generation framework(Chaotic Landscape Generator,CLG)is proposed.Due to the random weights of the recurrent neural network and the multiple distortions of the new activation function,the model can generate a variety of test problems to comprehensively evaluate the performance of the optimization algorithms.2.From the perspective of evolutionary algorithms,some specific improvements for particle swarm optimization(PSO)are involved in solving the test problems with special properties.In order to improve the ability of solving multi-modal and ill-conditioned problems,Powell Method and velocity reinforced mechanism are introduced into PSO,which improves the exploration ability of the algorithm by increasing the diversity of search directions and the ability of local search.Experiments show that the targeted improvements of various mechanisms enhance the competitiveness of the original algorithm.