Behavior Modeling for Autonomous Agents Based on Modified Evolving Behavior Trees

In modern training, entertainment and education applications, behavior trees (BTs) have been the fantastic alternative to FSMs to model and control autonomous agents. However, manually creating BTs for various task scenarios is expensive. Recently the genetic programming method has been devised to learn BTs automatically but produced limited success. One of the main reasons is the scalability problem stemming from random space search. This paper proposes a modified evolving behavior trees approach to model agent behavior as a BT. The main features lay on the model free method through dynamic frequent subtree mining to adjust select probability of crossover point then reduce random search in evolution. Preliminary experiments, carried out on the Mario AI benchmark, show that the proposed method outperforms standard evolving behavior tree by achieving better final behavior performance with less learning episodes. Besides, some useful behavior subtrees can be mined to facilitate knowledge engineering.