Comment Re:Parallelism required? (Score 1) 246
> Winning a tactical battle is something computers are okay at now. MPP is a
> useful optimization for analyzing a tactical situation. However, winning a
> particular battle may lead to you losing the game (and frequently it
> does), so the key problem is choosing the battles to fight. If there are
> three fights on the board, how do you choose among them? Right now we
> cannot answer that question, so we don't know if MPP will help or not. I
> was honestly just conjecturing on the topic.
Actually, in Go, computers aren't even very good at tactics. Very basic tactics yes, but at a high, or even middling amateur level, issues that have yet to be successfully quanitified, become as important as raw territory in evaluating tactical success. Capturing 5 stones while letting your opponent get an outside wall is nearly always a tactical loss.
The other problem is that Go computers don't have great life and death engines, so the ultimate tactical issue of "If I let this group get cut off, will it live or die" is not something a computer is able to answer without reading the tree out to its most bitter end, or having a very specific pattern to match.
There's a sense in which computers are *weaker* at tactics in go than in strategy. If I play a very strategic, "peaceful" game with a computer, I will often still win, even giving 9 stones, but the better programs will appear to play at a 9-10kyu level. OTOH, I or players of my level can play all of these programs at 9 stones and win by well over 100 points, if we don't play "peacefully" but start lots of complicated tactical fights.
The real key to the weakness is the intersection of tactics and strategy -- the evaluation of *what* constitutes a tactical "win". A pro player has an excellent but hardly perfect sense of this. As a low amateur dan, I have a decent but very fuzzy sense of this. Weak players have little to no sense about it beyond the most obvious question of who has more completed territory. This is the position the computer is in. This is in contrast to chess where 95%+ of what can be considered tactical success can be well defined mathematically. So when the computer goes searching through the move tree in a chess game, it will recognize any clearly advantageous positions within its search depth. In go, existing evaluation functions are likely to grossly mischaracterize the value of various positions in the same way that rank beginners do.
So maybe you're right that this is essentially a strategic weakness, but oddly it shows up most strongly in what go players would consider "tactical fighting.
Michael
> useful optimization for analyzing a tactical situation. However, winning a
> particular battle may lead to you losing the game (and frequently it
> does), so the key problem is choosing the battles to fight. If there are
> three fights on the board, how do you choose among them? Right now we
> cannot answer that question, so we don't know if MPP will help or not. I
> was honestly just conjecturing on the topic.
Actually, in Go, computers aren't even very good at tactics. Very basic tactics yes, but at a high, or even middling amateur level, issues that have yet to be successfully quanitified, become as important as raw territory in evaluating tactical success. Capturing 5 stones while letting your opponent get an outside wall is nearly always a tactical loss.
The other problem is that Go computers don't have great life and death engines, so the ultimate tactical issue of "If I let this group get cut off, will it live or die" is not something a computer is able to answer without reading the tree out to its most bitter end, or having a very specific pattern to match.
There's a sense in which computers are *weaker* at tactics in go than in strategy. If I play a very strategic, "peaceful" game with a computer, I will often still win, even giving 9 stones, but the better programs will appear to play at a 9-10kyu level. OTOH, I or players of my level can play all of these programs at 9 stones and win by well over 100 points, if we don't play "peacefully" but start lots of complicated tactical fights.
The real key to the weakness is the intersection of tactics and strategy -- the evaluation of *what* constitutes a tactical "win". A pro player has an excellent but hardly perfect sense of this. As a low amateur dan, I have a decent but very fuzzy sense of this. Weak players have little to no sense about it beyond the most obvious question of who has more completed territory. This is the position the computer is in. This is in contrast to chess where 95%+ of what can be considered tactical success can be well defined mathematically. So when the computer goes searching through the move tree in a chess game, it will recognize any clearly advantageous positions within its search depth. In go, existing evaluation functions are likely to grossly mischaracterize the value of various positions in the same way that rank beginners do.
So maybe you're right that this is essentially a strategic weakness, but oddly it shows up most strongly in what go players would consider "tactical fighting.
Michael
