Re:Build your own MENACE (matchbox-powered machine learning ...
by gwern on Sunday February 27, 2022 @04:38PM
(#30492792)
I think it would be most similar to tabular Q-learning. It has some weaknesses due to its simplicity, like not decaying the learning rate (you always add/remove from a fixed number of matches; maybe if you used an unbounded number of matches it'd be equivalent?) or having any exploration built-in, so it doesn't enjoy the guarantees tabular Q-learning does like converging to the optimal game play - MENACE would presumably jitter around the optimal strategy if it approached it at all.
[ Parent
| on HN ]
There are no comments on this comment yet.