Research

This paper introduces an environment for simple 2D maze games, designed as a sandbox for machine learning approaches to reasoning and planning. Within it, we create 10 simple games based on algorithmic tasks (e.g. embodying simple if-then statements). We deploy a range of neural models (fully connected, convolutional network, memory network) on these games, with and without a procedurally generated curriculum. We show that these architectures can be trained with reinforcement to respectable performance on these tasks, but are still far from optimal, despite their simplicity. We also apply these models to games involving combat, including StarCraft, demonstrating their ability to learn non-trivial tactics which enable them to consistently beat the in-game AI.