Is reading a suboptimal subset exercise?

2 posts / 0 new
Bloemen
Is reading a suboptimal subset exercise?

AlphaGo has quite an interesting way to decide which move to play, let me roughly summarize the most important characteristics. For every position it encounters AlphaGo selects a set of candidate moves, likely to produce the best result. These candidate moves are input for quite some random move sequences. The move giving the best chances of winning the game is selected.

The moves selected by AlphaGo may theoretically not be the best, as to be sure of that, an exhaustive analysis of all the possible moves has to be performed from start to finish, giving a set of perfect games. This kind of analysis is impossible to perform, due to the forbiddingly huge amount of possibilities. But AlphaGo may come quite close to perfect moves, as it is feasible to suppose that rather often the move giving the highest chance of winning the game may also give the best result, is the best move. At least AlphaGo performs at a consistent high level, avoiding glaring mistakes.

Because of the way AlphaGo decides on the move to play, some of the moves are unexpected to human beings, some moves seem to lack human logic. Certainly the AlphaGo - AlphaGo games are sometimes very difficult to understand. For instance, moves that may be considered sente by human beings, are not seen as sente by AlphaGo. Furthermore, it sometimes is difficult to figure out what a move played by AlphaGo is trying to accomplish. There is no objective to a move AlphaGo knows of, as the only parameter that is useful to it is the chance of a move to win the game.

One of the most powerful possibilities for human beings to analyze a position is reading: mentally working out move sequences in a tactical situation. There are two difficulties here: selecting the right moves in all kinds of relevant variations, and evaluating the results of these move sequences. Being correct or wrong, the nature of all the moves analyzed while reading is that they are sente, the move sequences can be considered forced. The evaluation of a move sequence sometimes is easy, like stones that are captured or saved, groups that live or die. Sometimes the evaluation is quite difficult, like judging whether territory is worth more or less than a thick wall, the value of the aji of stones left alone, or the value of cutting points. The nature of evaluating reading move sequences is that the conclusions are mostly local. When human beings analyze a position, typically these kinds of sente move sequences are laid down on the board, and the results of them are discussed.

When human beings try to understand the games played by AlphaGo, a problem arises. Quite often reading sequences are laid down on the board, we already saw that happening by those hero commentators trying to explain the Ke Jie - AlphaGo games. But the question is: does AlphaGo read out move sequences? It doesn't, although sometimes a human reading sequence may appear at random, could be deduced from some of the many variations tried. The move that according to AlphaGo has the highest chance of winning the game may be everywhere on the board. As reading is about a sequence of sente moves for a local tactical situation, one could see the reading exercise as a narrow subset of possibilities that AlphaGo might encounter. And, looking at the whole board, the result of a local reading exercise may not be instantly relevant. Reading may help human beings to understand the status of a local situation, but when looking at AlphaGo's games it is quite useless at helping to predict where to play next, to understand and to explain what really is going on. So, I wonder how the AlphaGo's games are going to be explained, whether other methods than reading are going to be used in a meaningful way.

Because of the moves that AlphaGo doesn't consider sente, and the restricted usefulness of reading when looking at AlphaGo's games, it seems to me that the human concept of sente is overrated. Ke Jie indicated that many basic concepts of go have to be re-evaluated, is sente one of them?

Kind regards,
Paul

Haiga
No it's not

You're bringing up a lot of topics...

First of all, Alphago reads and it reads a lot, something like 40 moves sequences according to the programmers (I don't have the reference right here, but i can search it back if you need to). It's not "random move sequences"...

Then there's a misconception : reading and judging are different things. Reading is a must, what can you do if not ? Top pros can read hundred of sequences of 30 or 40 moves, how can you get the result of a complicated fight or the perfect last 40 moves yose if you dont read ?

There are 2 issues though, blind spots and evaluation, and that's kinda what alphago is also putting in perspective. It certainly have blind sports as it cannot read everything, but at least it doesn't have the same bias as human have while rejecting moves/pruning the tree search of sequences. Coming up with "new moves" or not ignoring some is what we would call creativity, something that some pros are known for, like Lee Sedol.

Judging/evaluating is the most difficult part, because it deals with instincts and whole board perception. As games in today's world are quite fast (2h+30s byoyomi are the usual in pro tournaments), people cannot read everything, which means they have to rely on intuition, but intuition itself is helped by reading and games played and reviewed.

But there are ways for humans to improve, like deeper analysis/evaluation of games/sequences/tactics/strategies, news tsumegos with long reading sequences full of blind spots in early/late middlegame situations, and so on. After his match with Alphago, Lee Sedol said we should emphasize a lot on reading.

There's a reason why the strong players are strong players, they know a lot about the game and they can read a lot, but they don't know everything and can't read everything....

At last, the topic of sente is not overrated at all : sente is a concept which would translate as the value of a move in another area of the board is superior to the value of a move added to the local sequence, it's not something people just came up from the top of their head... The value of sente is the value of the next bigger place on the board, but what alphago is good at is taking sente, because its evaluation of the position is better due to its deep reading. When a human player thinks a move is sente when the other player doesn't, someone is making a mistake in evaluating the position, aka reading a position and assessing a point value to it, so either one player takes a loss in points, or the other gets punished for ignoring. And when alphago takes gote, it usually do this with a clear gain on the board.