arxiv.org/abs/2611.08144

Detecting Silent Faults at the Environment-Trainer Boundary in Reinforcement Learning

In reinforcement-learning experiments, injected faults at the environment-trainer boundary often produced no error, and learning-curve checks missed most runs unless every value was corrupted. A step ledger kept on both sides detected and located every boundary fault in the reported tests, adding a median 3% training time on Pendulum.


No comments yet

Be the first to share your thoughts.