PAIndemic: A Planning Agent for Pandemic

Abstract

Cooperative games are an important challenge for AI research. One example of this genre of games is the board game Pandemic, characterized by the challenges it presents, including its considerable search space and hidden information, especially when the cooperators are not able to communicate explicitly. The game pushes players to prioritize between short term and long term objectives, which forces them to not only plan their own individual actions but also to cooperate with the other players in order to win the game. In this paper we present a planning agent which uses state-space planning with a domain-specific heuristic, combined with a Monte Carlo sampling approach to predict possible outcomes in the face of hidden information. We performed several experiments with our agent, including a comparison with a baseline version that does not use planning. Our experiments showed that our agent is able to win about a third of the games played in a realistic game setup.