Reinforcement learning a new approach for the cultural geography model

Abstract

The Cultural Geography (CG) model, under development in TRAC Monterey, is an open-source agent-based social simulation, designed to offer an insight into the response of the civilian population during Irregular Warfare (IW) operations. It implements social and behavioral science theories that govern the behaviors of agents within the simulation using Bayesian belief networks. At this stage, the agents within the CG model do not select their actions at all. Instead, all their actions are hard coded into the model's scenario file. As part of an attempt to improve the model, this effort sought to enhance the functionality within the model by exploring the use of utility functions and, more specifically, the concept of reinforcement learning. This study began with the development of a learning agent prototype. After the initial testing for its functionality, the code that was developed was inserted into the main CG model. Based on specially developed scenarios, and by employing a design of experiments methodology, we created experimental runs. By applying statistical and analysis techniques, we showed that reinforcement learning works properly inside the Social Network environment and produces the desired results. This study can be used as a starting point for the research of the effects of reinforcement learning in social modeling in general.