RL Lab / failure analysis
Failure Analysis
A record of experiments that missed a target, unresolved anomalies, numerical issues, and what was changed or still needs to be tested.
Failure is evidence when recorded clearly
Failure notes distinguish observations from diagnosis. An unexplained result remains unresolved until a reproducing case or new evidence supports a cause.
- Tabular Q-Learning on CartPole-v1: The first 500-episode run produced a measurable but weak policy. Mean evaluation reward was 52.0 ± 36.7 with a 0.0% success rate at the project threshold. Lesson: The experiment remains open. Reproducibility, discretisation, exploration, and training stability need to be investigated before treating the method as solved.