Bill Spight wrote:
That depends upon how the program is written. IIUC, AlphaGo Zero and Leela Zero start out with no theoretical go knowledge and end up with none. It may be possible that other strong programs will come along that utilize go theory. Don't top level chess engines incorporate chess theory in the form of opening books and tablebases?
Mike Novack wrote:
The neural net programs aren't written to be able to do anything but to implement a neural net. A neural net has the property of being able to learn (of being taught) to implement a function. In OUR case (the human brain) we probably start out with some connections, biasing us to form "theories" of why things happen << presumably this biasing useful in our evolutionary process >>
IF you have ever had "experimental psychology" you will have learned that not only is this tendency not uniquely human, but that it is "superstition" developed in the learning process << the rat might have accidentally "learned" to do a back flip before pressing the reward lever -- the back flip of course having nothing to do with the mechanism giving the reward >>
Understand? We humans are more comfortable with "theories", thinking that we understand why.
IF you have ever had "experimental psychology" you will have learned that not only is this tendency not uniquely human, but that it is "superstition" developed in the learning process << the rat might have accidentally "learned" to do a back flip before pressing the reward lever -- the back flip of course having nothing to do with the mechanism giving the reward >>
Understand? We humans are more comfortable with "theories", thinking that we understand why.
A lot of go theories are heuristics. I suppose by superstition you mean a heuristic that is unsound. For instance, when I was learning go textbooks said that it was incorrect to extend further from a 3-4 point than an enclosure. That is wrong. Curiously, this was at a time when the Chinese Fuseki was becoming popular.
This was a heuristic that held sway for a time. If you go back a couple of centuries players often extended further, just as we do today.Quote:
The "theories" of WHY "the next move is a game of go is such that there is no better move" would represent shortcuts to the evaluation of that function, shortcuts that we humans might find useful even when slightly incorrect. This makes sense since the practical evaluation of the function is an approximation of the exact function. These neural net AIs are also learning to evaluate the function, also still inexact approximations. But they are learning to do that directly, skipping the theory business.
Go theories need not be incorrect at all. For instance, one of the first theories that many human beginners are taught is the basic theory of ladders. The theory is simple and because the reading of simple ladders does not tax human working memory, even human beginners can play ladders that span the board. IIUC, modern go neural nets have not learned that theory to the same depth. I suppose for two reasons: first, they cannot represent the intensive theory like humans do, and second, that they have not encountered enough instances where such long ladders are relevant to have learned them.
, wanting to play at "a".
fails.