My point is that statistical methods were never considered in GO until just recently. A bunch of old-timers in computer GO were really skeptical and resisted well beyond the point where it was clear that you compete unless you had Monte Carlo Tree Search.bob wrote:Been aware of it for some time. And have also been aware of the rather serious shortcoming it has as well, since "deep threats" are completely misunderstood because of the very fast games used in the MC analysis.Don wrote:The "real deal" is very difficult to come by. Are you aware of the recent amazing progress in computer go? They are using a combination of tree search and statistical methods (based on Monte Carlo methods.) They call it MCTS which stands for Monte Carlo Tree Search. Essentially they find out what the "real deal" really is by playing a lot of (somewhat random) games. It is all about statistical sampling now.bob wrote: I do not believe that statistical sampling will ever beat "the real deal" in terms of what to reduce or not reduce. It could be improved, but more direct application of chess knowledge makes more sense. Otherwise we could do evaluation in the same way and not have so many different terms for various chess features.
It's an idea. And for go, because of the branching factor, it appears to be better than normal alpha/beta search. We are not "there" in computer chess however. AB works more than good enough.
I don't think that you can say that AB works more than good enough. Good enough for what? I want my program to be 500 ELO stronger than it is, and AB is not cutting it - so I cannot say that it works more than good enough.
I'm not suggesting that we can do it the same way in Chess. What I am suggesting is that I think there are some ideas out there based on statistical methods that we have not yet imagined. I don't know what form they might take.
And it is the "real deal" in the sense that no non Monte Carlo Go program these days has any chance to compete.
I think Marco probably has it right in this regard. I don't like it any more than you do, but I think it's likely the case that the best programs in a few years will be much more about probability and statistics and sampling. I don't what exact form this will take, if I did I would be doing it now!
Of course in theory it probably seems cleaner and simpler to just write a function to "give you the right answer", the "real deal" as you say and this is probably what you are imagining as being possible. But this is MUCH harder than statistical sampling.
(1) much harder to do;
(2) much more accurate, too.
When I first started playing with Monte Carlo and Go I spend a lot of time trying to write an evaluation function that had some kind of reasonable sense of the goodness of a go position. GO is really hard - there is no known simple and fast evaluation function that works very well. What is trivial for a human to see at a glance seems impossible to program heuristically.
Then I started playing random games to see what move was wining the most games. I quickly discovered that playing 1000 games and keeping statistics on which points were alive at the end of the games was the evaluation function I always wanted! It was amazingly good at having a "feel" for which points were safe and which were dicey and which groups were alive. It wasn't perfect, but it beat anything I could manually construct by a long shot. But then people started utilizing the statistics in interesting ways and improving the "Monte Carlo" playouts by using various methods (many of them also statistical in nature) the programs really took off.
In MCTS, the entire search tree is directed solely by statistical methods. Whether you try a move or not is based on complex formulas about prior results. It's actually quite incredible and it has caused programs to do what many predicted could never happen in the foreseeable future, programs reaching the Dan level.
Bob, do you remember when people were saying that about Chess? They would never reach expert, they would never reach master, etc? In chess it happened gradually, but in GO it happened very suddenly due to statistical methods.
If such a program could rip thru Fine's book with near-perfect accuracy, I would be impressed. But it is unlikely in the extreme, because the nature of chess is deeper and narrower, compared to go. Different tree shapes require different search approaches, mainly because for some trees, alpha/beta just can't go deep enough.
I don't want to give the wrong impression, Go program are still no where close to the top human players, but the progress in just 2 or 3 years was totally insane.