Better Understanding of Humans for Cooperative AI through Clustering
Cooperative AI and AI alignment research are increasingly important fields of study as machine learning models are becoming more prevalent in society. Applications such as self-driving cars, realistic AI in games, and human-AI teams, all require further advancement in cooperative and alignment research before more widespread applications can be achieved. However, research in these fields has typically lagged behind other machine learning applications due to the difficulty of creating models that are robust to and can adapt to novel human partners. We attempt to address this through the creation of a framework that uses Archetypal Analysis, a unique clustering algorithm that finds extremal ‘archetype’ points in a dataset and expresses each other point as a convex combination of these archetypes. This framework creates understandable archetypes of players which a reinforcement learning agent can use to adapt accordingly to unseen partners. We show that this framework not only results in performance comparable to other cooperative benchmark models but also achieves higher levels of perceived cooperativeness without the need for human involvement during the training process. As such, we demonstrate that the use of clustering techniques to better model different types of human behaviour and strategies can be an effective approach in improving the ability of AI models to adapt to and improve cooperation with novel partners.