Life In 19x19

Kirby - yes, pretty much. For example, if in its current favorite move it reads deeper and discovers that it actually blunders a group to die without compensation that the neural net initially didn't perceive as killable, then the evaluation of that move will go way down. (Mechanically, this is because after searching deeper now the search tree for that move contains many positions where the opponent has played the killing move making the dead shape more obvious, which now the neural net correctly recognizes as very bad for the player whose group died).

Basically its doing something very similar to what you might do with more time to think - reading deeper and reevaluating based on that reading.

Kirby wrote:

lightvector wrote:Since again there seems to be confusion about this, relating mostly to oversimplifications of the relevant ideas, here's an attempt to describe succinctly and accurately without any of those usual easy-to-misinterpret oversimplifications.

65% win rate means that following what the bot considers to be likely good play by both sides over the next few moves¹, on average the resulting positions are ones that the neural net believes are "similarly good" to positions it's seen in the training data where the player-to-move in that data won about 65% of the time². The training data usually consists of a slightly old-and-weaker version of that bot playing itself many times using a certain fixed number of playouts³, and with much heavier randomization than normal⁴.

¹Of course, occasionally the bot's reading may entirely be overlooking a good move by one or both sides.
²But still limited by the neural net's ability to understand and compare those positions. Larger nets will on average have better understanding. But they can still massively blunder/misjudge from time to time.
³In particular, this means that the 65% is NOT an estimate of how likely this version would be to win with the potentially very different number of playouts that you are running it with.
⁴More randomization provides the neural net with richer and more varied training data to learn from, but also means that the bot in the training data is much more likely to blunder than normal, which of course also affects the win % just like the other things mentioned in (³).

Thanks for the succinct explanation - and to others who have similarly added to the discussion.

Given this explanation, could you elaborate on what is happening when the win rate is changing as the number of playouts increases? E.g. if I let LZ sit there, the percentages start to change. Obviously, the training data hasn't changed, so something about the playouts happening right now are affecting the win rate, right? Is it just that it's finding different board positions a few moves ahead that match up to newly found positions that are similar to different training data positions (thereby adjusting the probability)?

Thanks.

Emphasis mine.

lightvector wrote:Basically its doing something very similar to what you might do with more time to think - reading deeper and reevaluating based on that reading.

Elsewhere I discuss some differences between Elf's evaluations of the same game with settings of 100K playouts and 200K playouts. As a working hypothesis I assume that the winrates with the 200K setting are more accurate than those with the 100K setting. OC, both settings are far greater than those in the training data, else we would still be waiting for those games to finish. By that token what is being estimated is not clear. Arguably the two different settings are estimating different things. Another unknown factor has to do with the fact that as the level of play increases, the estimated winrates approach 0 or 1, so we should expect the winrates with 200K to be more extreme than the winrates with 100K. How much more we -- at least, I -- don't know.

Nice - thanks for the explanations.

hyperpape wrote:

Tryss wrote:I think our neural net are less sensitive to this, but these kind of things can happens

I don't know if you meant go programs by "our neural nets", but in a very real sense, our human neural nets are similarly vulnerable, as a recent paper apparently demonstrated: https://spectrum.ieee.org/the-human-os/ ... ial-images.

I was talking about go programs, but thanks for this very interesting article!

Back to the initial question (because the discussion may have drifted off)

We may be missing the point. The matter might not be "what is the BEST move possible" but "what is the move in practical across the board play between two humans that makes the outcome CLEAR?"

In other words, human A has analyzed the position in terms of basing saving the game on a particular resource and human B makes a move (the ear reddening move) that makes it clear that this resource does not exist. THAT would explain the ear reddening.

Deeper analysis, by either humans using more time than over the board allows or by our current very strong go programs might reveal even better moves in the sense that they assure victory irregardless of that resource. But that might not be immediately clear (no ear reddening)

Mike Novack wrote:Back to the initial question (because the discussion may have drifted off)

We may be missing the point. The matter might not be "what is the BEST move possible" but "what is the move in practical across the board play between two humans that makes the outcome CLEAR?"

In other words, human A has analyzed the position in terms of basing saving the game on a particular resource and human B makes a move (the ear reddening move) that makes it clear that this resource does not exist. THAT would explain the ear reddening.

Deeper analysis, by either humans using more time than over the board allows or by our current very strong go programs might reveal even better moves in the sense that they assure victory irregardless of that resource. But that might not be immediately clear (no ear reddening)

In my case my toes flush read when caught by surprise. I keep my shoes on.

Mike Novack wrote:Back to the initial question (because the discussion may have drifted off)

We may be missing the point. The matter might not be "what is the BEST move possible" but "what is the move in practical across the board play between two humans that makes the outcome CLEAR?"

In other words, human A has analyzed the position in terms of basing saving the game on a particular resource and human B makes a move (the ear reddening move) that makes it clear that this resource does not exist. THAT would explain the ear reddening.

Deeper analysis, by either humans using more time than over the board allows or by our current very strong go programs might reveal even better moves in the sense that they assure victory irregardless of that resource. But that might not be immediately clear (no ear reddening)

So what we need is not a neural network trained to win games, but rather, a neural network trained to demoralize humans!

Hi Kirby,

rather, a neural network trained to demoralize humans!

Unfortunately, this is also past tense... for some pros.

Kirby wrote:
So what we need is not a neural network trained to win games, but rather, a neural network trained to demoralize humans!

Well ........ first of all, what we have currently are neural nets trained to "find the best next move" and "analyze chances for winning for each side". That is NOT quite the same thing as being "trained to win games". If behind, the objectively best move, but one which keeps things clear and simple might not be as good as a move not quite as good in objective terms but which made the game very complicated with lots of tempting ways for the opponent to go wrong. I ahead, the reverse might be true. In terms of the objective of winning the game.

We MAY be able to train the nets to do this (there may be difference between "simple" and "complex" in terms of the number of alternatives close in evaluation)

I don't think we have been training these neural nets to examine a sequence of moves made by the opponent to predict "what is the opponent aiming at?" << and thus, figuring out how to block that >> That appears to me to be a very different problem.

Mike Novack wrote:

Kirby wrote: Well ........ first of all, what we have currently are neural nets trained to "find the best next move" and "analyze chances for winning for each side". That is NOT quite the same thing as being "trained to win games". If behind, the objectively best move, but one which keeps things clear and simple might not be as good as a move not quite as good in objective terms but which made the game very complicated with lots of tempting ways for the opponent to go wrong. I ahead, the reverse might be true. In terms of the objective of winning the game.

If the win probabilities are based on positions similar to the training data, as I understand, the win percentage would represent the probability that the bot (or maybe an earlier version of it) would win against itself. Making things complicated or simple shouldn't have an impact, unless it results in a difference in calculated win rate for the bot.

Playing against a human might be different. The bot wasn't trained to win against humans, so the move that gives the best win rate for bot vs. bot might be a different move than the move that has the best odds of beating a (given) human.

EdLee wrote:Hi Kirby,

rather, a neural network trained to demoralize humans!

Unfortunately, this is also past tense... for some pros.

Yeah, tell me about it. I've been there myself (demoralized). But I think I got over it, now.