- Machine learning model for NetHack suddenly performs 40% worse due to a unique bug related to the full moon
- Researchers tried various solutions like code reversion and software stack restoration without success
- The bug was eventually traced to the game’s response to a full moon, affecting player luck and enemy behavior
- The model’s lack of training data on full moon variables led to suboptimal decision-making
- Despite the unconventional nature of the bug, it posed a significant challenge and provided valuable insights into machine learning limitations
Related Video
Published on: June 18, 2020
Description: Seminar by Tim Rocktäschel at the UCL Centre for AI. Recorded on the 17th June 2020. Abstract: Progress in Reinforcement ...
The NetHack Learning Environment
Related Wikipedia Articles
Topics: No responseResponse
Response may refer to: Call and response (music), musical structure Reaction (disambiguation) Request–response Output or response, the result of telecommunications input Response (liturgy), a line answering a versicle Response (music) or antiphon, a response to a psalm or other part of a religious service Response, a phase in emergency management...
Read more: Response
