On the Pitfalls of Learning to Cooperate with Self Play Agents Checkpointed to Capture Humans of Diverse Skill Levels

When engaging in collaborative tasks with unknown team members, humans demonstrate the ability to predict the behavior of their partners and adapt to it. Autonomous agents do not exhibit such adaptability, often struggling to integrate with new partners in multi-agent cooperative scenarios. Past work towards tackling this problem includes sampling from a population of diverse training partners. This consists of self-play agents at various skill levels, generated by checkpointing at various points throughout their training. In this work, we show that such a set of agents isn't representative of human skill levels by evaluating their qualitative and quantitative performance on the Overcooked Domain. Our results demonstrate that self-play agents exhibit distinct learning patterns in contrast to humans and a partially trained self-play agent demonstrates behaviors that diverges significantly from that of a lower-skilled human counterpart.