Artificial intelligence has seen great success in many applications, including image classification, natural language processing, and more. The key element in these examples is the machine learning from static datasets with clear desired outputs.

However, instead of learning knowledge from static datasets, it is also possible (and indeed necessary and desirable) to learn through feedback from interactions in their environment. This is especially important for machine learning models to be deployed in the real world, since the models themselves will impact each other, turning their individual decision-making into a multi-agent problem. Agent learning in a complex multi-agent world is therefore a fundamental challenge for next-generation AI.

Reinforcement learning is one form of AI that learns by interacting in the world. When interacting with humans, it can become especially effective. Our work shows how an agent can learn to coordinate with humans without knowing about interacting with humans in advance, in the context of the video game, Overcooked.

This is part of a wider research agenda concerned with the cooperation of machines with other machines and with humans. It has also involved machines learning from humans for robotic control tasks, and the use of reinforcement learning to play multiplayer video games such as StarCraft II.

Project Lead

Yali Du
Department of Informatics
Faculty of Natural, Mathematical & Engineering Sciences
King’s College London