Bots that undo

[RBerenguel]

I perfectly understand how an MC evaluator works. But the reason the AI gives may not have anything to do about why the game was won in most of the MC variations. Imagine an AI that does not understand ladder breakers, and plays a move because "threatens to attack a weak group" as a reason, this move turns to be a ladder breaker for some other area. Then, most MC branches pick this ladder and mark as a win. The reason was not "threatens to attack a weak group", so, as a learning tool it is not specially good.

Imagine that we have before us a go playing program that claims to be able to give go reasons for the move made. Why do you want to decide what it can or cannot do without looking? Why are you saying that if there was more than one go reason for a move it will display just one reason? Is this based upon your observation of how this program behaves?

How about if we go back to just before the first MCTS based program. The program in question was using an AI to select a set of plausible moves based on go reasons and then it made a choice of "best move" from among these. It was making that choice by an AI deciding which reasons counted more in the particular situation. You could ask it to show you why. And it was able to play go at just a stone or so weaker than you are now. You think that could have been done without the AI knowing about ladders or that a move could have more than one go reason behind it?

You are totally missing the point, so, whatever.

[daal]

Teachers give reasons based on their understanding of a position. In some cases these are verbalized heuristics, but often a teacher will show a few sequences and point out the difference in results. A 6k AI might offer some heuristics, but an MC will want to show you zig-thousand sequences.

[Mike Novack]

I am sorry if I was being too obscure by simply saying look and asking have you looked.

What I mean is that the question "can a program give multiple reasons why a move was made?" (or is this impossible) can be settled immediately if there is at least one program which can display multiple reasons. If we have in front of us at least one program which can do that then it clearly is possible.

So I will present this as fact, not endorsement for a product. If you turn on the reason option in MFOG 12 it will show a varying number of reasons. Sometimes only one (for example, it says that it is following a known joseki) but sometimes several (increase liberties, separate groups, get out, ....).

If you don't believe this, then look for yourself.

[quantumf]

The point is that the heuristic (explanation) that MFOGO chooses could be utterly wrong, and thus mislead the learner. i.e. MFOGO chooses the right move for the wrong reason. The learner will think, "In this position, extending two space from a wall is the right move", when in fact the correct reason was "playing a ladder breaker here is the correct move"

[ez4u]

The point is that the heuristic (explanation) that MFOGO chooses could be utterly wrong, and thus mislead the learner. i.e. MFOGO chooses the right move for the wrong reason. The learner will think, "In this position, extending two space from a wall is the right move", when in fact the correct reason was "playing a ladder breaker here is the correct move"

And this is different than a human teacher?

[Boidhre]

And this is different than a human teacher?

This is the part that's been bugging me. Most of the objections seem to apply equally well to human teachers explaining why an individual move was the right one.

[quantumf]

The point is that the heuristic (explanation) that MFOGO chooses could be utterly wrong, and thus mislead the learner. i.e. MFOGO chooses the right move for the wrong reason. The learner will think, "In this position, extending two space from a wall is the right move", when in fact the correct reason was "playing a ladder breaker here is the correct move"

And this is different than a human teacher?

Yes, because the human doesn't have an MCTS engine built in.

Humans could pick a move that's right for the wrong reasons, but the human doesn't know it. The combo-bot picks the right move for the right reason, but gives the wrong reason.

[Mike Novack]

The point is that the heuristic (explanation) that MFOGO chooses could be utterly wrong, and thus mislead the learner. i.e. MFOGO chooses the right move for the wrong reason. The learner will think, "In this position, extending two space from a wall is the right move", when in fact the correct reason was "playing a ladder breaker here is the correct move"

Please (re)consider what you have just said.

In effect, you are claiming that in a case where there is more than one reason why a move is the best move (indeed, precisely because more than one reason, because the opponent can't counter both with one move) the bot may be displaying the wrong reason.

That is a continuation of the "belief" that if there are multiple go reasons supporting a move the bot could not possibly be displaying more than one. Why do you continue to maintain this belief in the face of the clear evidence that it is possible for the bot to display multiple reasons? I will repeat, look at the evidence (MFOG 12 does display multiple reasons).

There is, of course, the important possibility that there is some "reason" that the AI of the MFOG program does not recognize. In fact, this is almost surely the case based on the evidence that there are MCTS programs that play a stone or so stronger on comparable hardware resources (they must be making some moves that MFOG did not consider) . Which simply means that the AI could be improved by figuring out what this reason was and adding a recognizer for it. But as far as I can tell, Fotland hasn't been working on improving the PC version of the program but instead putting the effort into improving the version that runs on weaker hardware. Since many folks seem to be wanting to play against the program on their ipads or whatever this is probably the right choice based on economics. While many folks decry the price of a program like this, I don't believe the potential market is large enough to pay for the hourly rate of somebody like Fotland (the market for 1) on a PC and 2) playing a stone or two stronger than now ---- minus for folks for whom 1 dan on a standard PC or 2-3 dan on a powerful workstation the present version is good enough as it stands)

[HermanHiddema]

@Mike: Perhaps you need to step back and reread? Nobody is claiming that programs cannot display multiple reasons. They are claiming that it is possible that a program will find wrong reasons and/or miss right reasons for moves, and will thereby mislead the student. That is absolutely a possibility. Especially if the move reason evaluation is not based on the same code that gives the program its playing strength. Which, given that MCTS code has no reason beyond "this move wins more on average over many simulations" and given that programs not based on MCTS barely reached SDK level, is not an unrealistic assumption at all.

[Mike Novack]

No, I am taking into account that the AI might not recognize some go reason. I've said so. And BTW, before MCTS the strongest AI's were playing at about 5kyu and 6kyu (go++ and MFOG 11 respectively).

But you need to think of this problem not at the level of code but of algorithm, and to consider how we, as humans, decide what move to make vs how MCTS is doing it.We need to understand the difference between "go reasons" at the lower level and the decision "best from the set of moves all of which have some good go reasons". The latter is not the same level of "reason". The MCTS algorithm is making that choice by a different mechanism than we do but conceptually both the MCTS algorithm and we are deciding that "this move is better than that because it makes us more likely to win the game". For us that is by somehow evaluating that in this position, the set of reasons behind this move rather than that one is more important, more urgent, leaves behind a choice of follow up moves, etc. Those are higher level go reasons than "the move is a ladder breaker".

And I think we need to consider why the 5kyu human player is a 5kyu player rather than a 1dan player. Is that because the 5 kyu player did not even see the go reasons behind some move? Or is it more likely made an error in judgement about which was more urgent, more valuable, left more possibilities behind, not in keeping with the nature of the game (small territory vs framework) etc. OK, the 5kyu human player might in fact miss some move for which (at the 5kyu level) the reason should have been known. But note that the 5kyu AI will not do that. It wasn't playing at 5kyu because it missed reasons but because the evaluator incorrectly judged which most critical in this position.

Let's consider how a human teacher might be trying to get a human pupil to improve (from 5 kyu to dan). Isn't that going to be more breaking bad habits like "following" or "allowing being forced" or "mistaking the direction of the game" than specific "go reasons" at the lower level which the student in fact already knows? (not that the student didn't know all the reasons, but misjudged the context as to which in this situation more important).

[quantumf]

From the point of view of explanations, then, the AI will only be able to give 5k reasons. The MCTS aspect is essentially meaningless (from a tutoring point of view).

The program will give occasionally wrong reasons for the right move, and that pains me, but I can accept that the 5k explanations are still useful for someone aspiring to 5k.

[Uberdude]

What are you lot wittering on about? I can learn from good moves without any accompanying explanation.

[skydyr]

What are you lot wittering on about? I can learn from good moves without any accompanying explanation.

I agree with this completely. The only issue, as I see it, is that it's hard as a weaker player to tell when an AI is playing a move that's actually good as opposed to a crap move that doesn't change the game result, or a move the player just doesn't know how to deal with but is actually bad if you know how to punish it. This is a problem with humans too, but bots certainly have a reputation for off-the-wall tenukis and other moves that don't make much sense.

[quantumf]

What are you lot wittering on about? I can learn from good moves without any accompanying explanation.

Are you mocking us? The argument has admittedly gone on way longer than it needed to, but your statement is bizarre. The logical extension from this is that strategy books are pointless, go lessons are pointless, trips to Korean go schools are pointless, etc.

I know many have pointed out that our "Western" obsession with explanations, heuristics, algorithms and so on are arguably a poor for our game, but there is a (as far as I can tell) an almost universal thirst for explanations as to "why" a certain move is good or bad, and the AI's do their best to meet this demand (successfully in chess, much less so in go). I can't believe you've never asked a stronger player "why".

[xed_over]

What are you lot wittering on about? I can learn from good moves without any accompanying explanation.

I can't believe you've never asked a stronger player "why".

I don't believe that's what he said.

I believe he is saying that he doesn't have to ask, that it is possible to learn on your own. Its discoverable.

Personally, I think each and every move played has, or should have a purpose or reason for being played (even if the person, or bot playing it, plays it for the wrong reason, or doesn't know the reason themselves), and its our job to try and find out what that reason might be. If we know why a move or sequence of moves was played, it can help us improve our own games. No one can know for sure what was in the mind of another player (or bot), but if the moves themselves can have meaning, then its not entirely undiscoverable their intent and purpose.