Expected Value Profit or Absolute Profit?

[Kirby]

...
I didn't say I had no way of evaluating the value of a move - that's not easy, but as you say it's no harder than counting the board. It's assigning probabilities to different outcomes when you're not sure that is hard.

For example, let's say you have a very standard choice - reduce or invade. It's easy enough to estimate the score if you reduce, or if you invade and live, or if you invade and die. What's hard is to figure out the probability of invade and die versus invade and live.

I'm referring to a situation where you know the correct result, but there's a good chance that your opponent does not.

So in your example, you know the "value" after you invade - say you get +50 points if you invade and your opponent screws up and lets you live. And you also know that if your opponent plays correctly, you get -20 points.

Given your opponent's rank, you may be able to estimate, "at X-kyu, there's probably like an 80% chance that he doesn't know how to kill me".

It's a rough estimate, but you could perhaps say that it's worth 0.8*50 - 0.2*20... However, as Joaz pointed out, issues such as how late in the game it is could affect the value of such an estimation.

[Kirby]

I should also say that I may be pressing this idea a bit too much, but that's because I actually feel this way during games. It's not as precise as an expected value calculation, but I often am faced with a situation where I feel, "I know is isn't the correct play, but if I play it there's a good chance my chances of winning will improve".

So sometimes I play the incorrect move.

[gowan]

It all depends on what your goal is in playing. If your goal is just winning then there may be times when you should play a move you know is bad. If your goal is to play the best possible move in the abstract sense then obviously you should never deliberately play a bad move.

[jts]

Right. I am saying that your count of the board is probably pretty good, whereas "Oh he's 3k, 80% of 3k don't know this trick" is a wild-ass guess. If you are going to calculate an expected value on the basis of a wild-ass guess, you should be willing to go the extra step and make an educated guess about the count.

[xed_over]

For example, I am in a situation, and I know a particular move will not work. But if my opponent screwed up and the move did work, ...

sounds like a ko threat to me

[Kirby]

Well, my counts are also wild-assed guesses

[Loons]

I think several things.

The most common and useful application of this is hamete. You could do a kgs database survey of what rank failure lines turn up in games and what your record is against that rank otherwise.

In this thread we're talking about demonstrably refutable moves. Not active moves, which would be difficult but possible seeming.

We actually do this in handicap games already with classic handicap tricks/unreasonableness that we think will work.

Playing our best, we are striving to always play the strongest move on the board, not one with an apparent failure instead. I don't mean anything silly by the word strongest.

[SmoothOper]

...

Playing a move you know is bad in the hope that your opponent doesn't read or know the refutation is bad.

...

Sounds like a standard answer, but why is this the case? Why is it bad? That is to say, if you feel that you increase your chances of winning the game by doing so, why not go for it?

An example that comes to mind is computer bots for poker. It's not uncommon for them to try to learn and exploit human irrationality to get a greater profit when gambling. They don't play the "optimal" choices given their cards based on the straight up probability of getting certain hands. Rather, they try to maximize gains by learning human playing styles.

Is this type of strategy bad?

Odds are only a simple form of statistic, and they tend to give statistics bad names in poker. An interesting poker statistic, is what is the probability of winning consecutive pots, it does't take genius to figure out that if you could calculate that, you could use that information to help predict a bluff, of course battle field statistics are a different ball game, and calculating more sophisticated odds, could be tricky.

[Kirby]

It all depends on what your goal is in playing. If your goal is just winning then there may be times when you should play a move you know is bad. If your goal is to play the best possible move in the abstract sense then obviously you should never deliberately play a bad move.

Indeed, it seems to come down to this. I have to admit, I have never understood the camp not having the goal of winning. If you are not trying to win, what are you doing?

Sometimes people say they want to play beautiful games. If that's the case, why not play in random, yet pretty looking spots?

I suppose motivation for playing the best possible move might be to enhance your personal understanding and search for truth in the game.

Perhaps that's a good goal to have.

Thanks for sharing.

[Loons]

Sorry, are we talking about playing the most actively to have a chance to catch up when behind (which would be totally standard for everyone) or are we talking about making (over)plays that we have read definitely fail, when behind?

I play for beauty and the most beautiful move is the most functional one I can find, whether I'm ahead or trying to catch up.

[Kirby]

I am talking about playing stuff that should not work, and probably wouldn't work if you played it against yourself. But for whatever reason, you think that it will increase your chances of winning the current game.

[Toge]

I think this question has to do with morality. Should one try his best if the people around them can't tell the difference? Lacking that expert knowledge and having good faith sets the victim in vulnerable position. Media is ridden with stories how some famous person or organization has acted in immoral ways trusting that the possility of being exposed and punished is small. Resulting reputation loss is harsh, but as we gamblers call it "probably +EV".

I find this talk about probability odd on Go forums, though. Go is a game of perfect information, which should make swindles difficult. I play moves that look like they are leading to favorable outcome. Good reading contains built-in principle of charity - in fact the player who thinks that his opponent is going to choose poor line isn't a strong Go player in the first place! Just like you're not a good conversationalist by accepting strawman version of your partner's argument, as the principle of charity suggests. Many of the mistakes I recognize after game are the type where I thought my opponent would play different move. It's easy to understand oriental notion of no-self here. Each move stands on its own qualities and it's not necessary that there even is an opponent to trick.

Very nice thread you started Kirby.

[Loons]

I think about this differently.

If I'm behind but feel the game isn't lost, what that must mean is that statistically I am not the winner. I don't resign (unless I'm really demoralised). I try the most appealing (ie highest probability success I guess) strategy that can get me back in front, or be a step on the path there ("and I'll try really hard in endgame"). Breakfast calls this "choosing to play more actively" (where choosing to play needlessly actively while you are ahead is a mistake).

If I can't see anything to try and feel like I'm down too much to recoup in endgame, I resign (or if it's a capturing race I'm down one liberty on, I play it out and then resign).

I don't play defensively like I'm winning when I'm not winning.

[Kirby]

...in fact the player who thinks that his opponent is going to choose poor line isn't a strong Go player in the first place! ...

This is a very fascinating statement. I suppose minimax algorithms have their place for a reason. If you made a program to search board positions, ignoring the fact that the opponent could play the strongest line of play, it'd be a pretty lousy program indeed.

Thinking along the lines of computer programs, it would be cool to make a computer program capable of doing two things:
1.) Classifying moves in local scenarios based on similarity.
2.) After playing the same person multiple times, start to learn a probability distribution of moves that that particular person might make against a given local position.

It would probably be difficult to accurately do either of these, but if it were to work out well, the computer could, given a board position, match it with positions from the training data, and come up with a probability distribution of possible places the opponent may play if the computer were to play at a given intersection.

If the program could evaluate the value of a board position, it could then select the move that gave the best value based on the probability distribution of opponent responses vs. the resulting board value.

I think minimax is cooler than this, however, since it seems closer to a program that is "seeking the truth".