MazeBase: A Sandbox for Learning from Games
This provides a versatile sandbox for researchers to test machine learning methods on reasoning and planning tasks, though it is incremental in building on existing game-based environments.
The paper tackles the problem of creating a sandbox environment for machine learning approaches to reasoning and planning by introducing MazeBase, a 2D game environment with 10 simple games, and finds that neural models perform suboptimally on these tasks, but models trained on MazeBase can beat the in-game AI in StarCraft.
This paper introduces MazeBase: an environment for simple 2D games, designed as a sandbox for machine learning approaches to reasoning and planning. Within it, we create 10 simple games embodying a range of algorithmic tasks (e.g. if-then statements or set negation). A variety of neural models (fully connected, convolutional network, memory network) are deployed via reinforcement learning on these games, with and without a procedurally generated curriculum. Despite the tasks' simplicity, the performance of the models is far from optimal, suggesting directions for future development. We also demonstrate the versatility of MazeBase by using it to emulate small combat scenarios from StarCraft. Models trained on the MazeBase version can be directly applied to StarCraft, where they consistently beat the in-game AI.