<crunchbubba> reads aos comment from last night about the nethack challenge at the AI/ML conference
<crunchbubba> there are too many variables for ML to get very far. i suppose if you let it play a billion training games you might get somewhere. maybe
<aosdict> I still contend that an AI that tries things randomly is statistically never going to stumble upon doing the invocation sequence correctly
<crunchbubba> right. but the metric was something like "found the down staircase 50% of the time"
<aosdict> they were trying to maximize score, weren't they? and so the bots just farmed monsters because that's how you get score
<crunchbubba> there were also hand-coded bots
<amateurhour> there were different categories, the ones that didn't use neural networks outperformed the ones that did, iirc
<crunchbubba> You can design and train your agent however you please — with or without machine learning, using any external information you’d like, and with any training method and computational budget.
<amateurhour> I considered writing a parody paper where I trained a coworker to play it
<amateurhour> DAVE has been trained on visual input for approximately 30 years