3 ms·
this is hilarious and also a great idea. i dont see any reason why you can't play a few million games against itself and other engines and see where it takes yo
by wwarner 7y ago
this is hilarious and also a great idea. i dont see any reason why you can't play a few million games against itself and other engines and see where it takes you. less efficient than alpha zero probably, but how much so?
- sillysaurusx 7y agoIt should be similarly efficient. AlphaZero used 1,000 TPUv1's to generate self-play games, and a much smaller number of TPUs to train the model on the previous self-play results. Whenever it generated a model that was >= 55% better, that became the new model. The same algorithm could be applied here.
- jeffshek 7y agoIt would not be close to similarly efficient. They have completely different loss functions.
- sillysaurusx 7y agoYou're right, "efficient" should be substituted with "possible". We're certainly not claiming that this is a smart way to do it, just that you can. Still, I think that there's a chance it could work well. Each move could be prefixed with the final outcome of the game, which is the technique either alphazero or muzero uses.