Balancing Player Experience Without Players At Supercell: Case mo.co
summary
Supercell builds a headless AI simulator (C++ client exposed as a Python wheel) and trains thousands of bots with proximal policy optimization, running about 1,000 parallel simulators on one GPU to balance gear before real players exist. Reinforcement learning finds exploitable combos (one setup kills the boss in one second), so designers iterate: limit bot action rate to ~0.5s, use genetic search for gear allocations, and adjust reward shaping to stop reward-hacking like the invisibility bypass.
AI-written summary · about
chapters (13)
- 0:00Supercell and Moo intro
- 1:05Supercell's mission
- 2:10Balancing without players
- 3:34Reinforcement learning
- 7:28AI simulator for Moo
- 9:36Training AI bots
- 11:02Proximal policy optimization
- 12:47Retraining models
- 13:49Designing rewards
- 16:40Encouraging exploration
- 17:01RL in Moo summary
- 18:06Q&A: Bots vs humans
- 19:32Q&A: Rewarding survival
related talks
53mBalancing gameplay of aRPG using player behavioral patterns
01h 01mIntroduction to Practical AI
20mEngaging Bots for Multiplayer Games
24mIEEE CoG 2019 - Industry Day - Christoffer Holmgård (modl.ai)
21mMeet your new and improved teammate! Human-directed AI agents and gaming
59mGame AI 4N00bz
02mDesigning a Video Game AI
52mThe Journey of AI in Games: A Technical and Philosophical overview