Shaping the Behavior of Reinforcement Learning Agents
George K. Sidiropoulos, Chairi Kiourt, Vasileios Sevetlidis, George P. Pavlidis · 2021
With the advent of machine learning and agent-based approaches, behavior-shaping in environments composed of several autonomous entities has become a popular and active research field for the development of unique realistic behaviors. Realistic simulations have been particularly studied in the fields of crowd management, swarm behavior analysis and civilization simulation. In this study, we present a new dynamic rewarding approach for shaping the behavior of reinforcement learning agents in mixed (cooperative and competitive) multi-agent environments. The evaluation of the proposed rewarding approach is tested in a developed 3D environment of two groups of ancient Greek warriors fighting inside an octagonal arena, testing different agent behaviors in various scenarios. Interestingly, the results reveal that the trained agents’ behaviors vary based on the situations and the constraints of the environment, resembling realistic behavior variations.