| |
BGonline.org Forums
Very impressive result from DeepMind team - Xavier's take?
Posted By: Maik Stiebler In Response To: Very impressive result from DeepMind team - Xavier's take? (eXtreme Gammon)
Date: Tuesday, 12 December 2017, at 9:38 a.m.
Do any of the player here think that a backgame is strategy to seek against any top level player?
It's not so much about seeking as about being less averse to letting it happen.
I imagine that XG has seen just enough snakes to learn that they are utterly hopeless. So it's become very good at avoiding them. From our human perspective, it's very clear that XG underestimates the value of snakes. Shouldn't we conclude that correct strategy is avoiding snakes somewhat less than XG says we should (I'm aware that's probably too vague a description to become a viable human strategy optimization, but isn't the conclusion logical?)?
On a related note, state-of-the-art NNs are often "Convolutional NNs", which Wikipedia says "are also known as shift invariant or space invariant artificial neural networks (SIANN), based on their shared-weights architecture and translation invariance characteristics." I am not an expert. I don't know to what extent that is still what Deepmind is doing. At least for earlier versions of their spectacular "AlphaGo" I know they explicitly mentioned using convolutional NNs. At any rate, my understanding is that what convolutional NNs do on the higher levels of the evaluation process (that is, close to the input nodes) is extracting local features of positions no matter where they appear. In that sense, to them a prime is a prime, and the concept of rolling it home should be kind of "intuitive" to them, much more so than to a brain which is shaped like XG's. But I'm speculating.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.