Artificial intelligence in co-operative games with partial observability

This thesis investigates Artificial Intelligence in co-operative games that feature Partial Observability. Most video games feature a combination of both co-operation, as well as Partial Observability. Co-operative games are games that feature a team of at least two agents, that must achieve a shared goal of some kind. Partial Observability is the restriction of how much of an environment that an agent can observe. The research performed in this thesis examines the challenge of creating Artificial Intelligence for co-operative games that feature Partial Observability. The main contributions are that Monte-Carlo Tree Search outperforms Genetic Algorithm based agents in solving co-operative problems without communication, the creation of a co-operative Partial Observability competition promoting Artificial Intelligence research as well as an investigation of the effect of varying Partial Observability to Artificial Intelligence, and finally the creation of a high performing Monte-Carlo Tree Search agent for the game Hanabi that uses agent modelling to rationalise about other players.