Multi-Agent Knowledge Based Reinforcement Learning

Recent work has scaled up game playing Artificial Intelligence (AI) to
world champion performances in Go and generalised applicability to a
wide range of early digital-games. However,
the focus of most work in this area has been on using AI to beat (or
outperform) human players and not on the betterment of humans. It is our
view that AI should augment (not automate) human
abilities. This project focuses on the development of intelligent
assistants for human game players in the context of multi-player games.

The main aim of this research is to apply reinforcement learning to
genrate an intelligent assistant, i.e., a collaborative AI that enables
non-player characters (NPCs) to enhance the experience of human players
attempting a range of tasks in multi-player games. The performance of
the intelligent player assistant will be measured by human player
enjoyment, immersion, engagement and performance. To do so we will
research (1) imitation learning to provide a predictable companion; (2)
reward functions for encouraging contribution towards a shared goal
instead of individual performance; and (3) predictive models of player
behaviour to identify human goals in domains with no explicit extrinsic
reward.