Your characterization is pretty spot on, I think. However, given that at the end game, its relatively easy for a computer to play perfectly, this seems like a desirable enhancement to supplement the bots with. Admittedly, as you say, this is somewhat at odds with the statistical approach of MC bots, and it would probably be a hard integration exercise, in the sense that once you start distorting the MC approach with other techniques, the whole thing tends to fall apart. But the pure MC approach is pretty much maxed out already, so supplementing it with extra abilities, e.g. strategic play, or perfect endgame, is necessary to move it up a level.

This doesn't seem correct to me, but I know very little about go bot programming so I could be wrong. Surely the MC approach is simply the method of ranking moves, accompanied by some evaluation function. The current bots perhaps do not play a good endgame for whatever reason, but this is a problem with the evaluation function (maybe because it's designed for all-round play, or whatever), not the MC technique itself.

Were a bot designed solely to judge an endgame objectively, I don't see why the MC method shouldn't be extremely powerful. I assumed that bots didn't play such a good endgame or play strategically just because this wasn't in their design yet.

Monte Carlo code is an interesting idea, but IMHO very poor for the future of skilled playing Go engines. I await to see if I'm right or wrong I guess.

In simplest terms, MC bots take a given position, it plays out huge numbers of sequences to the end for a number of moves, and then picks the one that "wins the most times". Over enough iterations, without a great deal of skill or logic in the moves, this works very effectively, but you'll often find an MC bot play a poor endgame, turning a 7.5 point lead into a 1.5 point win - however, get it close to the wire, like 1.5 or 0.5 points, and they play very strongly indeed.

Most monte carlo bots don't use very sofisticated evaluation functions. It's a tradeoff on speed vs accuracy, being able to analyse more positions instead. It seems to work pretty well.

Probably each bot has a tweaked evaluation function or MC algorithm, but I doubt every bot maker is keen to reveal its innards.

Edit:Also what topazg said.

_________________
a1h1 [1d]: You just need to curse the gods and defend.
Good Go = Shape.
Associação Portuguesa de Go

Perhaps it's the evaluation function that is the real limitation. If it was easy to evaluate a position it would also be easy to tell what the biggest move is. Position and next move go hand in hand. Then you wouldn't need to do any random playouts and don't need MC in the first place.

_________________
FusekiLibrary, an opening library.
SGF converter tools: Wbaduk NGF to SGF | 440 go problems | Fuseki made easy | Tesuji made easy | Elementary training & Dan level testing | Dan Tutor Shortcut To Dan

Yes, that's true, if all players would bother to learn a few overplays that would only ever work against a specific bot, then it would have a rating of 5k.

But in practice, it will still get a rating of 1d on KGS. Why? It's simple: The bot always makes the same mistakes, so you can rely on them happening. You can play a special strategy that exploits the mistakes. That wouldn't work against human players, because they would adjust their style after a few games, or even learn from their mistakes. So most people don't bother with finding out these special weaknesses of the bots, because it doesn't help their game against humans or even different bots.

That's the reason why the rating of a bot doesn't really tell you anything. It's KGS rating represents its strength when you play it like a human, using normal strategy. Once you adjust your game to its weaknesses, you'll see its "true" rating. And that will happen at least subconciously when you play bots too frequently. What have you gained then? Basically nothing, except picking up bad habits along the way...

This is also the reason why I don't believe in the recent "successes" of bots against pros in handicap games, because they were mostly one-off matches, with the pro having no opportunity to analyze the specific weaknesses of the bot.

Ummm. You are disproving your own point here. An MC program plays immensely strong endgame, because it's always looking for the best way to win, even if that's with a small margin. And in my experience, once it is a tiny bit ahead in the endgame, an MC bot is not going to let you catch up. Just try playing a few 9*9 games against any strong program and you know what I mean - the sheer number of 0.5 and 1.5 point losses becomes almost frustrating .

Also, whether it's by 50.5 points or 0.5 points, a win is a win. Therefore, questions like "could a 10.5 pt loss be reduced to a 0.5 pt loss" or "could the 10.5 point win be stretched to a 15.5 point win" are irrelevant to strong endgame play. What matters is taking home the win; counting the value of moves is merely an imperfect tool to help the imperfect human brain work towards that goal.

_________________
My name is Gijs, from Utrecht, NL.

When in doubt, play the most aggressive move

Where's pasky when you need him?

Interesting, I know it's subjective, but I wouldn't consider a strong endgame "minimise margin for going wrong" which is what MC bots do, I'd consider it "making perfect moves", which would maximise the margin of victory at the risk of complications and dangerous lines. In that sense, MC don't play a strong endgame unless the game is very close. If simply "winning" is the criteria, then that's different.

I disagree.

If you assume perfect play by your opponent, then yes, it's irrelevant. But in practice, if I'm 10.5 pt behind and I see a chance to reduce it to a 0.5 pt loss, it may still be a better strategy to try that. Let me elaborate:

Let's say on the board I'm behind, with a few good endgame points left. Now I have two strategies:
a) Play a solid endgame and reduce it to 0.5 pts behind with a high probability. I'm good at endgame, so let's say 90%. I'm not playing risky, in the worst case I'll lose by 5 pts.
b) Play a risky endgame (overplay) and make it a 5 pt win. It's risky, let's say the probability to pull it off is 30%, but if it fails it will be a 30 pt loss.

Current MC bots use strategy b). It would just see that a) has a very low probability to win, and b) has a 30% probability to win.
But I argue that there is another factor, namely the probability of a mistake by the opponent:
Let's say the probability that the opponent makes another 1 pt mistake is 40%, for a 5 pt mistake it's 5% and a 30 pt mistake just won't happen.
Suddenly strategy a) looks a lot more appealing: In 90% of the cases, I just need a 1 pt mistake from the opponent (40% chance), at worst case (10%) I need a 5 pt mistake (still 5%). This still gives a total win probability of 36.5%, assuming a just small mistake by the opponent.
Strategy b) has just 30% chance, and it's also assuming a mistake by the opponent (namely falling for the overplay).

Of course all these numbers are just made up on the spot to demonstrate a point.

Sure, you could argue that my strategy is not theoretically best play but exploiting mistakes by the other player. Against a hypothetical "perfect opponent", my strategy would be worse. That's true, but that argument is flawed: The assumption of mistakes by the opponent is already there in the MC method itself by assuming a win probability instead of a fixed final score.

Sure, Monte Carlo bots get defensive and play moves that are suboptimal in terms of points when they are ahead, but why is this considered "unskilled"? Isn't a win by half a point as good as a win by a landslide? In a sense, the computer is just not being greedy by taking more points than it needs to win.

In a position where one player is visibly ahead, there will often be several moves that qualify as perfect play -- that is, the opponent has no sequence of responses that cannot be refuted (in terms of who wins at the end of the game). A sufficiently powerful Monte Carlo bot will choose one of these more or less at random, not caring whether or not it'll win by 7.5 points or 1.5 points at the end of the game. A practical, finite Monte Carlo bot will tend to prefer the more defensive moves, and obviously, it won't always choose perfect play. But your position seems to be one where you could even criticize actual perfect play as "unskilled" or as a "poor endgame", which seems off.

The way I see it, Monte Carlo bots can absolutely become "skilled". They will, however, always play in a style that is not natural to any human, and so their play won't in general resemble that of a skilled human player. The situation we're seeing is one where the strongest computers are not strong because they're imitating strong humans (with opening books etc.), but because they (or their programmers) are actually developing a new style that works very well with the way computers think. That's just the way humans make sense of the game, by thinking of it in terms that work well with the way humans think -- for instance, "good shape" or "influence".

I actually find that an extremely interesting development in the ongoing mind sports human vs. machine rivalry. To the best of my knowledge that's not the way it went with chess?

Exactly. So we just need to convince it to play this strongly, even when it knows it is behind. If the deficit is small enough - I would say anything less than 15 points - the bot will have a good chance for an endgame comeback win against middling amateurs. It just doesn't know that.

_________________
Main Entry: zing·er Pronunciation: \ˈziŋ-ər\
1 : something causing or meant to cause interest, surprise, or shock
2 : a pointed witty remark or retort

Agreed. Go is a game of endurance. It isn't the good moves that win you the game. It is the lack of bad moves.

The only problem with Monte Carlo bots is that they generally over exaggerate the level of carefulness when they are ahead. And over exaggerate risk taking when behind. This generally doesn't cause any problems in the end game because MC bots read pretty damn good. But it does cause problems earlier in the game, and also makes them somewhat unstable in handicap games.

Now, with all this said. This is why MC bots actually makes for pretty good computer opponents. Their weakness is less in actual game play and more in maintaining risk/reward throughout the game. So I don't think you will pick up any real bad habits from them.

Agreed. And, I'd like to see it do it even if it was 20 points ahead. I'm sure most to all professional Go players would frown on being casual in the end just because you don't have to be that precise to win any more.

It is actually pretty much the way it went with chess (maybe to a less extreme degree).

For decades people tried to make chess programs that mimicked the thought processes of human players. This is one reason that creating a strong chess-playing program was considered really important in the field of AI.

In the 1970s, people basically abandoned the idea of making a chess program that thought like a human and switched to exploring the tree of variations as fast as possible. Of course they still try to discard obviously bad branches and look at interesting branches more closely, as a human does, but the whole concept of making a plan and implementing it, which had been seen as the interesting part of chess AI, was thrown out the window. Now it's mostly a question of having an static evaluation function that is both fast and high-quality, and of exploring the tree in the most efficient manner possible.

I'm with topazg here. I'm talking from a human's perspective when I say that MC bots is unsuitable for endgame analysis. Humans are generally interested in the best order of moves assuming perfect play. For a human it would make sense to go back and review a 10.5pt loss and improve it to a 0.5pt loss. He would be happy to have found a better endgame, maybe even leant something. Same reasoning when you are ahead, trying to maximize your win.

But MC bots only care about winning or losing. This is what makes them unsuitable for the task of endgame analysis in the human sense. Their algorithm just doesn't work that way.

_________________
FusekiLibrary, an opening library.
SGF converter tools: Wbaduk NGF to SGF | 440 go problems | Fuseki made easy | Tesuji made easy | Elementary training & Dan level testing | Dan Tutor Shortcut To Dan

For one thing, Making a plan still just relies on static evaluative heuristics put together. Moves that naturally improve a position will automatically work towards a dynamic and flexible plan with regards to the board as it stands.

Rybka has started to reverse the process again

I used to do UCI engine leagues as a hobby, and I was always amazed at the node differences. Fritz and Shredder would run between 1 and 2 million positions a second getting to about 13/14 ply per move at the time controls I used, whereas Rybka did about 30 to 40 thousand - a 500th the amount - and outperformed them.

The difficulty is that actually coding how people think is not an easy task. It helps that the Rybka author is both a programmer and a titled International Master

What I'm trying to say is that I think MC bots are inherently unable to play a strong endgame (in the human sense) when one side has a significant lead. I hope someone who has actually programmed an MC bot can say if this is true or not.

_________________
FusekiLibrary, an opening library.
SGF converter tools: Wbaduk NGF to SGF | 440 go problems | Fuseki made easy | Tesuji made easy | Elementary training & Dan level testing | Dan Tutor Shortcut To Dan

This is what I don't understand. Naively, the only problem here is that the bot doesn't distinguish between winning lines. Wouldn't just allowing them to distinguish this to aim for the best endgame possible change this completely?

I'm well aware that I'm probably missing something, but I can't see what it is...

In short, they abandoned "thinking" in favor of brute force calculation, and it worked, thanks to Moore's law. Can this happen to Go?

_________________
Main Entry: zing·er Pronunciation: \ˈziŋ-ər\
1 : something causing or meant to cause interest, surprise, or shock
2 : a pointed witty remark or retort