NetHack is considered great for machine learning due to its difficulty, choices, and being a 'single-agent' game suitable for rapid generation and play on modern computers.
Despite initial success in training a neural network to play NetHack and improve itself using reinforcement learning, the model's performance suddenly dropped by 40%, perplexing the researchers as machine learning typically progresses gradually in one direction.
[
add
]
[
|
|
...
]