Learning Behavior Trees by Evolution-Inspired Approaches

As a reactive and modular policy control architecture, Behavior Tree (BT) has been used in computer games and robotics for autonomous agents' task switching. However, constructing BTs manually for complex tasks requires expert domain-knowledge and is error-prone. As a solution, researchers have proposed to auto-construct BTs using evolutionary algorithms such as Genetic Programming (GP) and Grammatical Evolution (GE). Nevertheless, their effectiveness in practical situations is in doubt and there are different drawbacks in the application. In this paper, we present a novel BT evolutionary system that integrates both GE and GP as modules and auto-checks the complexity of a given task to select which module to use. In addition, our system collects BTs that are either previously generated or manually designed by the user, which are utilized to further improve the convergence speed and the quality of generated trees for new tasks.