Model Decomposition for Forward Model Approximation

In this paper we propose a model decomposition architecture, which advances on our previous attempts of learning an approximated forward model for unknown games [1]. The developed model architecture is based on design constraints of the General Video Game Artificial Intelligence Competition and the Video Game Definition Language. Our agent first builds up a database of interactions with the game environment for each distinct component of a game. We further train a decision tree model for each of those independent components. For predicting a future state we query each model individually and aggregate the result. The developed model ensemble does not just predict known states with a high accuracy, but also adapts very well to previously unseen levels or situations. Future work will show how well the increased accuracy helps in playing an unknown game using simulation-based search algorithms such as Monte Carlo Tree Search.