18

While watching Hikaru's Streams/YouTube (GM Chess Twitch Streamer), he's no doubt extremely impressive when it comes to recognizing patterns that win him material and forced mates in like 7+ moves. Things like this require extensive knowledge of certain key positions and a great load of pattern recognition to go on top of that. During or after the sequence, he goes over the sequence fairly briefly to explain his thought process to his viewers, and it's the type of stuff you watch and are really just amazed by. These are usually the flashy calculations that are showcased on highlight channels and things like that.

My concern is for the opposite type of moves, being the less subtle ones. Often, Hikaru will give no thought to a move and say something like "let's just go here" and it often turns out to be the best move there is (Once played a bullet game with 96 percent accuracy). On the surface, these moves don't look like they have any thought to them. Obviously, it doesn't straight up blunder anything, but it doesn't really give him an advantage either. This begs the question: when Hikaru plays these types of moves, does he innately know it's the best move based on his intuition, or is there some kind of hidden calculation/judgment that leads him to make the move? What reason goes into him playing that move?

This also leads to the more general question: do grandmasters have to have a reason for every move they play, or are some moves just "okay" and meant to pass the move to the opponent? Obviously, the opening is an exception to this but even so, there are underlying ideas of center control, development, etc. This might be wrong in some way, so if anyone could elaborate on this that would be great.

2
  • 7
    Is this only in the context of blitz/bullet, and not long time controls? – D M Jan 31 at 15:29
  • As my example comes from watching Hikaru's streams and they are mostly blitz/bullet speedruns, yes. – Timothy Oh Feb 1 at 3:35
22

I will try to answer this in a different manner, the way I understand this topic.

Do we think on every signal, turn, fork when we drive?

Do we think every time we eat food or walk on the street?

The answer is yes, we do, but that thought process has moved to our reflexes to the extent that our brain does not let us know that it is doing a task (Thankfully!) unless there is a critical alert in our brain that puts us to think about the situation.

Just as a fact, the human brain takes 35000 decisions per day and not every decision we are aware, but is still taken.

Just like this, the Grandmasters have reached a state where the rudimentary positions are into their reflexes and they effortlessly play their game, to the extent that they invent positions and theories on the board itself. Reaching that state is "mastery"

7
  • 30
    This answer would be better without that "fact" which is clearly a number the internet invented. – Stian Yttervik Feb 1 at 9:15
  • 8
    @StianYttervik: Indeed. 83% of statistics on the Internet are entirely made up. – Eric Duminil Feb 1 at 15:07
  • 4
  • 3
    @Eric Duminil - and 97% of that 83% is made up by 5% of people. – chasly - supports Monica Feb 1 at 23:10
  • Analogy with professional race-car drivers would be more convincing and yes they do think before every move and in the case of rallies they do it with two brains. – PatrickT Feb 2 at 1:26
5

I can also answer this from AlphaGo's perspective because we know fairly precisely how it works. We can then reason by analogy for the human perspective.

From a bird's eye view AlphaGo has 2 components, a neural network that looks at a snapshot of the board, and a Monte Carlo search step that uses the neural network output to search faster. Both the neural net and search produces (and optimizes for) an estimate of the probability of winning for every move. The key feature is that you can play chess with either of the algorithm components, or a combination. The Monte Carlo search on it's own has dreadful performance, but Deepmind computed the Elo rating of the neural network on it's own and it's pretty good at chess.

From the AlphaGo zero paper:

Figure 6b shows the performance of each program on an Elo scale. The raw neural network, without using any lookahead, achieved an Elo rating of 3,055. AlphaGo Zero achieved a rating of 5,185, compared to 4,858 for AlphaGo Master, 3,739 for AlphaGo Lee and 3,144 for AlphaGo Fan.

Now for the analogy to human thinking. The Monte Carlo search does a lot of heavy lifting for AlphaGo and searches through at least thousands of combinations. Human search (thinking) is a lot slower than that, but the search strategy might be smarter. Still I think we can consciously look ahead at most 100 options, much less than a computer could. This leads me to believe that the subconscious (neural net) analysis makes up a larger percent of the total performance in humans. If we consider the Deepmind performance difference as a low bound, we'd be anywhere between 60% and 90% as performant subconsciously.

3
  • 1
    This is very interesting! Basically, it provides a more concrete way of thinking of "intuition", which is the neural net's evaluation at any given position. The given argument, however, also suggests that in reality, grandmasters can't "calculate" faster than anyone else (as Monte Carlo is an unchanging search algorithm), they just have the intuition to know what to calculate. This was a very helpful read. – Timothy Oh Feb 1 at 21:17
  • The Monte Carlo search on it's own has dreadful performance Without an eval function, I don't see how search would work. One has to have some kind of eval function, no matter how rudimentary, to decide which branches to search or to distinguish between leaf nodes. – Allure Feb 2 at 0:54
  • 1
    The eval function for the search is winning or losing a game after a complete rollout. If you play random moves you'll eventually finish some games accidentally. The variant of Monte Carlo they use picks moves weighted by the neural net probability of winning with that move, so combined with the neural net it mostly explores promising paths. For their latest variant of Alphazero they start training with a random policy and no demonstration games, but it bootstraps itself pretty fast through self play. – csiz Feb 2 at 1:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.