Clustering players' strategies for generation of non-player characters
summary
Clustering turns millions of recorded formations into hundreds-to-thousands of playable templates by splitting data by formation size, applying a two-step clustering (cluster by champion class using normalized effective health/damage, then cluster by positions using Manhattan distance), and pairing champions with a fast greedy matching instead of exponential optimal matching. This reduces per-comparison time from days to seconds, yields diverse human-like NPC templates, and produces templates that can be tuned by pick, win-rate filtering, and level-scaling to fit difficulty progression.
AI-written summary · about
chapters (12)
- 0:00NPC strategy clustering
- 1:19Problem definition
- 2:01Player strategies explained
- 4:27Multiplayer matchmaking
- 5:31Adapting matchmaking
- 7:38Smart opponent templates
- 10:10Clustering formations
- 13:46Practical clustering
- 19:27Position-based clustering
- 21:32Template generation
- 23:16Key lessons
- 25:04Q&A
related talks
53mBalancing gameplay of aRPG using player behavioral patterns
06mUnderstanding Player Patterns by Combining Knowledge-Based Data Abstraction with Interactive ...
50mToo Clever isn't Fun
03mPlayer-Centric Procedural Content Generation: Enhancing Runtime Customization by Integrating...
21mBalancing Player Experience Without Players At Supercell: Case mo.co
57mSolutions That Evolve: Building Self-Improving Systems with Genetic Algorithms
24mIEEE CoG 2019 - Industry Day - Christoffer Holmgård (modl.ai)
52mLessons from the Character Creation Process of Destruction AllStars