Building Modular, Human-Interpretable AI Systems with Behavior Trees

Abstract

Behavior Trees provide a structure to control the execution and switching of tasks and actions for autonomous agents, such as game characters or robots. Complex behaviors or tasks can be compartmentalized into sub-problems that correspond to actions. These actions become leaves of behavior tree, under the control of high level composite nodes. This modular arrangement makes behavior tree an ideal candidate for constructing solutions for many AI applications, advantages includes readily changeable and human-interpretable. Two main topics covered in this thesis are: 1) IKBT: solving inverse kinematics with behavior trees. IKBT demonstrates how manually designed behavior trees with domain-specific knowledge are capable to solve problems that were usually handled by human experts before. 2) Behavior tree for efficient hierarchical reinforcement learning. Behavior tree provides a modular problem formulation that facilitates searching within feasible actions and re-applying acquired knowledge. Behavior tree augmented reinforcement learning agents are more efficient in learning long horizon, sparse reward problems. Such problems are challenging for reinforcement learning.