Qiang Lu, Fan Tao, Shuo Zhou & Zhiguang Wang.

Received: 29 May 2020 / Accepted: 11 December 2020.

Abstract

Most traditional genetic programming methods that handle symbolic regression are random algorithms without memory and direction. They repeatedly search for the same or similar positions in the symbolic space, and easily fall into premature convergence. To overcome these shortcomings, we propose a new symbolic expression search algorithm based on the Monte Carlo tree search named SE-MCTS. It creates a tree to represent the symbolic space and remembers its visiting positions. Moreover, it incorporates two neural networks—Actor and Critic into the upper confidence bound to direct its search based on the given dataset features and decide when to use particle swarm optimization to find fitted constants in a mathematical expression. The experiment results show that SE-MCTS can discover more proper mathematical expressions than canonical genetic programming methods.

More info about thie paper

Please see Springer-Link

Code Link

Please see SE-MCTS on our GitHub.