AI Takeover
summary
A Pac‑Man clone is reimplemented and successive components are replaced with machine learning: a random-forest classifier (trained on ~50k synthetic moves, reported 99.42% on a 5k test set but still fails about 10% of the time), GANs trained on scraped images to generate 32×32 sprites (GANs produce sprites after ~5 hours on an i7), genetic algorithms used to evolve level layouts (yielding playable but biased corridor designs), and a Deep Q‑Learning agent to play the game. The results demonstrate working ML substitutions but reveal issues of non‑determinism, poor generalization, data bias, and significant compute requirements.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:10Talk goals Pac-Man
- 2:35Game architecture
- 3:59Replace engine with AI
- 13:25Model limits
- 14:07Neural networks intro
- 22:38GANs for graphics
- 27:10Genetic level generation
- 32:04AI bias objectives
- 33:50Deep Q-Learning bot
- 36:38Debugging UI
- 38:47AI recap limits
- 40:10AI as tool
- 44:03AI future dev
- 45:26Thanks conclusion
related talks
01h 31mReinforcement Learning- Pac-Man
01h 02mTeach your Pacman to play with ML and Reactive Streams
01h 01mReinforcement Learning: Pac-Man
56mCan I do AI?
35mHow I Built an AI to Beat the 1990s Arcade Game Double Dragon
39mBuilding self evolving video game AI in Python
48mMachine Takeover - Improving Your Game With Machine Learning & AI
01h 01mIntroduction to Practical AI