Using Cultural Coevolution for Learning in General Game Playing

Traditionally, the construction of game playing agents relies on using pre-programmed heuristics and architectures tailored for a specific game. General Game Playing (GGP) provides a challenging alternative to this approach, with the aim being to construct players that are able to play any game, given just the rules. This thesis describes the construction of a General Game Player that is able to learn and build knowledge about the game in a multi-agent setup using cultural coevolution and reinforcement learning. We also describe how this knowledge can be used to complement UCT search, a Monte-Carlo tree search that has already been used successfully in GGP. Experiments are conducted to test the effectiveness of the knowledge by playing several games between our player and a player using random moves, and also a player using standard UCT search. The results show a marked improvement in performance when using the knowledge.