Gamified Reinforcement Learning: Mastering AI through Video Games
summary
Builds a simple Flappy Bird gym in Python using pygame + numpy and trains a population with a genetic algorithm (no backprop). Agents use a 4-input MLP with 32 hidden neurons (inputs: bird y, pipe gap y, pipe x distance, velocity), run in parallel as a population of ~100, and evolve via keep/crossover/mutation (roughly 25% keep, 25% children, 25% mutants, 25% new). Simulation runs, agents improve and sometimes pass pipes, but premature convergence and tuning remain necessary.
AI-written summary · about
chapters (16)
- 0:00Gamified reinforcement learning intro
- 6:18Reinforcement learning theory
- 16:32Real-world example of reinforcement learning with a pet
- 21:08Agents in game environments
- 23:35Neural networks for AI
- 27:26RL gym design
- 37:22Code walkthrough setup
- 40:55Entity component system
- 54:21Multiple agents in game
- 1:01:04Neural nets and genetic algorithms
- 1:23:37Live simulation demo
- 1:30:19Q-learning comparison
- 1:35:32Metaheuristic algorithms
- 1:40:14Saving trained models
- 1:45:01Future work
- 1:45:50Q&A
related talks
39mBuilding self evolving video game AI in Python
32mMachine Learning for videogames in Unity3D
02mDesigning a Video Game AI
06mCreating and Teaching Games in Python to Teach Artificial Intelligence Concepts
31mDesigning game AI with reinforcement learning
26mGames Are Smarter Than Us
56mCan I do AI?
37mTensorflow and deep reinforcement learning, without a PhD