MazeBase: A Sandbox for Learning from Games
Sainbayar Sukhbaatar Department of Computer Science Courant Institute New York University \ANDArthur Szlam, Gabriel Synnaeve, Soumith Chintala、Rob Fergus Facebook AI Research New York
Abstract
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.
中文速览
为了让神经网络学会推理和规划，研究者构建了一个叫做 MazeBase 的二维迷宫游戏沙箱，并在其中设计了 10 种涵盖条件判断、集合运算、环境操控等算法推理能力的简单小游戏。他们用强化学习（reinforcement learning）分别训练了全连接网络、卷积网络和记忆网络（Memory Network）等多种神经模型，并测试了自动课程学习（curriculum learning）对训练效果的影响。实验结果显示，尽管这些任务看起来相当简单，现有模型的表现仍远未达到最优，说明当前神经网络在结构化推理上仍有明显短板。此外，研究者还将 MazeBase 用于模拟《星际争霸》的小规模战斗场景，在沙箱中训练好的模型可以直接迁移到真实的星际争霸环境并稳定击败游戏内置 AI，展示了该平台在跨场景知识迁移方面的价值与潜力。
原文 arXiv:1511.07401;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1511.07401v2