Elaborate Operational Requirements to Address Reward Hacking in Reinforcement Learning Agents

Abstract

Autonomous agents, in recent times have been used to address several problems, but these agents in their course of achieving their task also emit side effects to the environment in which they operate. Paramount of these side effects is reward hacking. In this report, we try to address reward hacking using elaborate operational requirements. The results is evaluated on the unity machine learning platform using multi agents, a goalkeeper and a striker where the elaborate operational requirements helped address these agents from hacking or gaming their results.