Modeling agent behavior through online evolutionary and reinforcement learning

The process of creation and validation of an agent-based simulation model requires the modeler to undergo a number of prototyping, testing, analyzing and re-designing rounds. The aim is to specify and calibrate the proper low-level agent behavior that truly produces the intended macro-level phenomena. We assume that this development can be supported by agent learning techniques, specially by generating inspiration about behaviors as starting points for the modeler. In this contribution we address this learning-driven modeling task and compare two methods that are producing decision trees: reinforcement learning with a post-processing step for generalization and Genetic Programming.