The Research On Symbolic Regression Based On Reinforcement Learning

Symbolic Regression(SR)is the task of discovering a symbolic expression that best fits a given dataset from the space of symbolic expressions.Genetic Programming(GP)is the most common algorithm for dealing with the SR problem.However,it is a random search method without direction and memory so that it repeatedly searches for the same or local positions in symbolic expression space,and easily falls into premature convergence.For addressing these problems,the thesis proposes a space search algorithm(named SE-MCTS)based on Monte Carlo Tree Search and Actor-Critic.SE-MCTS utilizes the monte carlo tree to represent the symbolic expression space,and solves the memoryless search problem through traverse this tree.It utilizes Actor-Critic method to evaluate the node in the monte carlo tree and lets the search move to the “high evaluated” nodes so that solving the non-directional search problem.The experimental results show that,compared with GP,SE-MCTS algorithm has more stable search effect and can find better expression.Owing to SE-MCTS need to maintain and traverse the monte carlo tree during the search process,it costs a lot of time to search SE-MCTS costs much time to search.For addressing the problem that cost much time to search in SE-MCTS,the thesis proposes a GP algorithm(called MARL-GP)based on Multi-Agent Reinforcement Learning.MARL-GP embeds Multi-Agent Reinforcement Learning into the GP evolution process,which can keep the characteristics that parallel search and faster search time in GP.And it can utilize Multi-Agent Reinforcement Learning to guide the crossover and mutation operations in GP so that find “high evaluated” expression.The experimental results show that,compared with SE-MCTS,MARL-GP can find an expression with high fitness in a shorter time.