On Jan 5, 2017, at 10:49 PM, Robert Jasiek <jas...@snafu.de> wrote: > > On 06.01.2017 03:36, David Ongaro wrote: >> Two amateur players where analyzing a Game and a professional player >> happened to come by. >> So they asked him how he would assess the position. After a quick look he >> said “White is > > leading by two points”. The two players where wondering: “You can count > > that quickly?” > > Usually, accurate positional judgement (not only territory but all aspects) > takes between a few seconds and 3 minutes, depending on the position and > provided one is familiar with the theory.
Believe it or not, you also rely on “feelings” otherwise you wouldn’t be able to survive. Some see DNNs as some kind of “cache” which has knowledge of the world in compressed form. Because it's compressed it can’t always reproduce learned facts with absolute accuracy but on the other hand it has the much more desired feature to even yield reasonable results for states it never saw before. Mathematically (the approach you seem yourself constrain into) there doesn’t seem to be a good reason why this should work. But if you take the physical structure of the world into account things change. In fact there is a recent pretty interesting paper (not only for you, but surely also for other readers in this list) about this topic: https://arxiv.org/abs/1608.08225 <https://arxiv.org/abs/1608.08225>. I interpret the paper like this: the number of states we have to be prepared for with our neural networks (either electronic or biological) may be huge, but compared to all mathematically possible states it's almost nothing. That is due to the fact that our observable universe is an emergent result of relatively simple physical laws. That is also the reason why deep networks (i.e. with many layers) work so well, even though mathematically a one layer network is enough. If the emergent behaviours of our universe can be understand in layers of abstractions, we can scale our network linearly by the number of layers matching the number of abstractions. That’s a huge win over the exponential growth required when we need a mathematical correct solution for all possible states. The “physical laws” for Go are also relatively simple and the complexity of Go is an emergent result of these. That is also the reason why the DNNs are trained with real Go positions not just with random positions, which make up the majority of all possible Go positions. Does that mean the DNNs won’t perform well when evaluating random positions, or even just the "arcane positions” you discussed with Jim? Absolutely! But it doesn’t have to. That’s not its flaw but its genius. David O.
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
