← All modules/Module 11 · AI/ML
RU

AI / ML in games

From pathfinding and tree search to LLM NPCs. The thread running through it: classical AI is for control, modern AI is for content, and maturity is knowing when ML is NOT the answer. Nine topics below, four with interactive labs (🏠).

0 / 0 read
1

Classical control and search

60 FPS, determinism, debuggability. What actually drives NPCs in shipped games, plus tree search for games you can compute your way through.
FSMs and Behavior Trees~15 min
Finite state machines and behavior trees — the backbone of NPC control.write-up
Pathfinding: A* and NavMesh~15 min
Graph search for paths, heuristics, navigation meshes.🏠 lab
Minimax, αβ, MCTS~18 min
Game tree search: αβ pruning, MCTS/UCB1, hybrids with a learned evaluation (AlphaZero, NNUE).🏠 lab
2

Procedural generation

Content from rules. The core principle: generation needs constraints.
Wave Function Collapse~15 min
Constraint satisfaction as a generator of tile worlds.🏠 lab
PCG: noise and dungeons~16 min
Perlin/Simplex noise, L-systems, BSP, grammars. Constrained randomness > ML for content.🏠 lab
3

The modern frontier

Generative AI: dialogue, worlds. Expensive, slow, powerful — and moving fast.
LLM NPCs: anatomy of the system ★~30 min
Persona, memory, function calling, latency, economics, platforms.🏠 lab
World models~15 min
Genie 3, Muse/WHAM — a neural net as the game engine. Research frontier.write-up
4

When ML is not the answer · synthesis

The most valuable judgement call: where learning earns its keep and where the classical approach wins. A map of the module's connections.
RL agents: why they don't ship~16 min
Superhuman in demos (OpenAI Five/AlphaStar), almost never shipped: determinism, cost, tuning. Where ML in games is genuinely needed.write-up
Classical vs ML: how to choose~12 min
Control versus generalization. A decision tree plus the module's patterns.write-up