Research On Learning-guided High-dimensional Optimization

When solving black-box optimization problems,Evolutionary Algorithms can be viewed as a class of general-purpose optimizers.But as the dimension of the problem increases,Evolutionary Algorithms heavily suffer from the so-called "curse of dimen-sionality" problem.Nowadays,high-dimensional optimization is still one of the most important challenges in the field of evolutionary computation.This paper aims to help the algorithms achieve better performance in high-dimensional optimization by learning a prior for the problem.The main contributions of this paper are listed as follows:1.Aiming at the high cost and poor performance of the CMA-ES algorithm in large-scale global optimization,the Correlation Coefficient based Grouping(CCG)strategy and Model Complexity Control(MCC)framework are used to remove re-dundant degrees of freedom in the Gaussian model.Under the simplified model,the cost of modeling and sampling is greatly reduced,and the optimization per-formance is also improved.2.Aiming at the difficulty of choosing the planning horizon in the sequential decision-making problems,the policy network and value network are introduced to learn from historical experience.Under the guidance of neural networks,the agent can obtain better planning results with shorter planning horizon and less search cost.On the CEC 2010 large-scale global optimization benchmark functions,the pro-posed MCC-CCG-CMAES algorithm is significantly better than the compared state-of-the-art counterparts.In the OpenAI Gym movement control environments,the pro-posed p-RHEA algorithm is also significantly better than the compared online planning method and reinforcement learning method.Both of the works studied in this paper are of universal significance,which will promote the application of Evolutionary Algo-rithms in the field of high-dimensional optimization.