Reinforcement learning as a testing methodology
summary
Claim: reinforcement learning finds subtle, human-discoverable heisenbugs that fuzzing and property tests miss; uses Rurel to train agents against a pure-physics reward. Built six domain-specific integration tests that reproduce speedrun techniques, enforces a 40-hour training cutoff, and uses baseline max speed as a safety threshold. Electromouse is an early-stage library that orchestrates these experiments; it does not perfectly automate verification but gives actionable guidance, reproduces known exploits, and speeds debugging and porting decisions.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:32Quake 1 speedrunning
- 8:27Strafe jumping exploit
- 13:03Bugs vs features in ports
- 16:52Game engine port testing
- 22:24What is reinforcement learning
- 27:43Designing RL tests
- 30:14Electromouse framework
- 34:45Why use RL testing
- 36:51Summary
related talks
42mSurprise yourself with Property Based Testing
28mBuild bigger in less time: code testing beyond the basics
21mA Tour of Empirical Verification in Rust
55mComprehensive Testing of Embedded Software
19mA Hybrid Approach to Semi-automated Rust Verification
52mLearning Rust the wrong way.
52mLearning Rust the wrong way
34mGood or Bad Vibes? Coding with AI in Rust