Behavior Optimization Of Simulation Soccer Robots Based On Improved Covariance Adaptive Evolution Strategy

Stable and fast walking is the premise of mobile robots' dynamic confrontation and collaboration.At present,Robo Cup3 D competition uses covariance adaptive evolution strategy(CMA-ES)to optimize the robot skills and achieve good results.But with the increase of the parameter dimension,the algorithm has the problems of slow convergence and falling into local optimality.This paper uses Robo Cup3 D as the experimental platform,which mainly analyzes the optimization characteristics of CMA-ES algorithm and the correlation of robot parameters in the optimization process.In the paper,The main work and research contents of this paper are as follows:First,aiming at the problem of slow convergence of CMA-ES multi-dimensional parameter optimization for robots,an optimization method based on parameter curve similarity clustering is proposed.Pearson correlation coefficients(PCCs)are used to calculate the similarity of various parameters in the optimization process,and a weight threshold is set to obtain a similarity matrix for spectral clustering(SC),and then the classification parameters are optimized.The feasibility of the method is verified by clustering optimization experiments on typical functions and classification prediction experiments of robot parameter curves.Aiming at the problem of adding self-collision penalty to Robo Cup3 D platform and part of the robot's skill action freezing,the optimization task is designed based on the hierarchical overlap optimization framework,and the robot optimization is re-combined with the cluster optimization method.Firstly,the walking training task is decomposed into three subtasks: walking to the target point,brisk walking to the target point and dribbling to the target point through a hierarchical overlap optimization framework,and then clustering optimization of the corresponding walking parameters of the subtasks.The advantages and disadvantages of the obtained parameters are compared on the Simspark platform in terms of walking trajectory,gait direction,and start and stop times.Experiments show that the parameters obtained by clustering optimization are superior to those obtained by the original hierarchical overlap optimization.To deal with the problem that CMA-ES is easy to fall into local optimum,a CMA-ES algorithm based on chaotic mapping and Levi's flight disturbance is proposed.Guide the update of mean value through Tent map and Levi's flight disturbance,so that it has a certain probability to jump out of the local best.The influence of the perturbation operator on CMA-ES is verified and compared on typical functions.Finally,the improved algorithm is used to re-optimize the robot and apply it to the Robo Cup3 D competition,and compare with other teams to verify the optimization effect of the robot.