Engineer Trains AI Agent on a Single PC to Beat Warcraft III at Insane Level

🤖 AI on a Single PC Beats Warcraft III at Insane

Engineer Robin Kahlow (Warlock.AI) spent three weeks training an agent using self-play RL on a PC with an RTX 3090: on the Echo Isles map, it wins 89% of games against the built-in AI (patch 1.29). The Windows client ran under Wine, with ~50 instances driven from Python via w3shim.dll, which replaces game clocks: ~730 game hours per real-time hour. The network is a 5-layer Transformer with minGRU (~3.2 million weights), with inference taking ~13 ms per step — the agent plays in real time.

🌍 Techniques from AlphaStar (behavioral cloning, self-play league, UPGO) worked without an API or data center: the harness turns the game “black box” into an RL environment, and the recipe is transferable to other closed simulations.

👤 Code and logs are open at github.com/RobinKa/warcraftsim — you can run and further train the agent yourself. The author confirmed via measurements that the Insane AI gathers resources twice as fast as a human.

Source 1: https://warlock.ai/blog/warcraftsim-1/ Source 2: https://github.com/RobinKa/warcraftsim