Learning of Behavior Trees for Autonomous Agents

In this paper, we study the problem of automatically synthesizing a successful behavior tree (BT) in an a priori unknown dynamic environment. Starting with a given set of actions, a reward function, and sensing in terms of a set of binary conditions, the proposed algorithm incrementally learns a switching structure, in terms of a BT, that is able to handle the situation encountered. Exploiting the fact that BTs generalize AND-OR -trees and also provide very natural chromosome mappings for genetic programming. we combine the long-term performance of genetic programming with a greedy element and use the AND-OR analogy to limit the size of the resulting structure. Finally, earlier results on BTs enable us to provide certain safety guarantees for the resulting system. Using the testing environment Mario AI, we compare our approach to alternative methods for learning BTs and finite state machines. The evaluation shows that the proposed approach generated solutions with better performance, and often fewer nodes than the other two methods.