Re:Learning ‘Montezuma’s Revenge’ from a single demonstration
by jhbadger on Wednesday July 04, 2018 @04:13PM
(#17459239)
I played it on the Apple ][ -- it seemed to be ported to many platforms. But it is a fairly obscure game -- it makes me wonder why they picked it rather than a more common one.
[ Parent
| on HN ]
Re:Learning ‘Montezuma’s Revenge’ from a single demonstration
by jonrei on Wednesday July 04, 2018 @04:58PM
(#17459417)
It was picked because it is difficult to train a reinforcement learning model to play it well. In most other games you can create a reward function based on the score or something similar, and then the AI can explore possible actions that gives the best score. In those cases AI players are already doing quite well.
In this case finding the key requires long term planning to get an actual reward and the AI has previously got stuck before that.