What it does

This repo wires DeepMind’s pysc2 environment to OpenAI’s baselines algorithms so you can train agents to play simplified StarCraft II mini-games. The flagship example: a Deep Q-Network that learns to collect mineral shards on a minimap. You get two algorithms (DQN with prioritized replay and dueling networks, plus A2C/A3C) and a handful of tunable hyperparameters.

The interesting bit

The value here is plumbing, not novelty. In 2017, getting three Google-sized dependencies—DeepMind’s environment, OpenAI’s RL library, and Blizzard’s game client—to actually talk to each other was non-trivial. This repo is essentially a working docker-compose.yml for your GPU.

Key highlights

• Supports DQN (with dueling + prioritized replay toggles) and A2C out of the box
• Hyperparameters exposed as CLI flags: learning rate, exploration fraction, agent count, n-step returns
• Includes enjoy_mineral_shards.py for watching your agent stumble around after training
• One working GIF of the mineral-shard task, which is more documentation than some repos provide

Caveats

• TensorFlow 1.3 dependency dates this to roughly the Obama administration; expect dependency hell on modern Python
• Only one mini-game (CollectMineralShards) is actually demonstrated
• README numbering jumps from 4 to 5 back to 4-1, suggesting maintenance has been sparse

Verdict

Worth a look if you’re trying to understand how pysc2 and baselines once fit together, or need a historical baseline for SC2 RL research. Skip it if you want maintained, multi-task StarCraft agents—this is a fossil, not a foundation.