The success of deep reinforcement learning methods in single-agent settings has fostered a boom in research interest which as technology progress has now expanded into multi-agent settings. Modern methods are now able to play a number of games at a super-human level. Sports games have long been used as testbeds for AI research, a distinguishing characteristic of sports games from other genres is that dominance of the basic game mechanics is not enough to be considered successful; they are also expected to exhibit realistic human-like behaviours and be fun to play.
This dissertation project investigates the combination of reinforcement and imitation learning methods applied to a novel American football themed environment. The experimental results obtained demonstrate that there is some evidence of emergent behaviours and the combination of reinforcement and imitation learning leads to faster training and stronger game playing agents than reinforcement learning alone. The results also suggest that end-to-end reinforcement learning on its own is not enough to produce the type of explicitly coordinated actions as would be required in a commercial game.
Keywords: C#, Unity ML-Agents, Deep Reinforcement Learning, Multi-Agent, Learning Environments