Abstract for: Learning Control Policies in System Dynamics Models
Advances in artificial intelligence and optimal control provide increasingly better algorithms for controlling dynamical systems. These algorithms can be applied for policy design in system dynamics models. In this paper we introduce some basic solution concepts and apply the Q-learning algorithm to a simple dynamic model from system dynamics literature to demonstrate potential value of such cross-fertilization. We also extend a state aggregation and partitioning algorithm that may increase the efficiency of basic reinforcement learning models in application to continuous time and space problems. Simulation analysis demonstrates the value of this approach and offers guidelines for future research.