The real-time strategy game StarCraft has proven to be a challengingenvironment for artificial intelligence techniques, and as a result, currentstate-of-the-art solutions consist of numerous hand-crafted modules. In thispaper, we show how macromanagement decisions in StarCraft can be learneddirectly from game replays using deep learning. Neural networks are trained on789,571 state-action pairs extracted from 2,005 replays of highly skilledplayers, achieving top-1 and top-3 error rates of 54.6% and 22.9% in predictingthe next build action. By integrating the trained network into UAlbertaBot, anopen source StarCraft bot, the system can significantly outperform the game'sbuilt-in Terran bot, and play competitively against UAlbertaBot with a fixedrush strategy. To our knowledge, this is the first time macromanagement tasksare learned directly from replays in StarCraft. While the best hand-craftedstrategies are still the state-of-the-art, the deep network approach is able toexpress a wide range of different strategies and thus improving the network'sperformance further with deep reinforcement learning is an immediatelypromising avenue for future research. Ultimately this approach could lead tostrong StarCraft bots that are less reliant on hard-coded strategies.
展开▼
