Learnfun and Playfun: A Nintendo automation system
summary
A program learns what “winning” means by observing human play and extracting lexicographic orderings from the NES’s 2048 bytes of RAM, producing ~250 orderings from ~4,000 frames; it then searches by saving states and doing time travel in an emulator, evaluating 40 futures of 300 inputs each and simulating steps at ~500 μs apiece. The system automates dozens of games, sometimes reaching impressive wins and sometimes producing hilariously bad behavior.
AI-written summary · about
chapters (4)
related talks
30mControlling the NES with Javascript
37mEmulators, Immutability, and Time Travel
30mHacking your Nintendo with JavaScript.
31mUsing Python to build an AI to play and win SNES StreetFighter II
01h 37mShall We Play A Game
33mAlpha2048 - how I taught my computer to play 1024
26mSuper MarioScript: Tools and Tips for NES development
46mAI Takeover