2

For position you never played, book probability is perfectly fine for me. The question starts when you already played position x-times. I'm looking for function that will update book probabilities with your style. Obviously if you play 1.d4 in 97 games out of 100 games, probability of playing 1.e4 runs close to 0%. On the other hand, if you played some late position just once, it's nowhere close to 100% you will repeat the move. I could probably make some workaround myself, but this looks like problem data scientists must have already solved, as it is applicable to virtually any discipline where you want to predict someone's behavior.

Edit: I think there should be one parameter <0;1> measuring how seriously we take book probabilities and how seriously we take players games. With one extreme being ignoring book and one being ignoring player's games.

  • If you're using database games of your opponent, then he is probably using yours as well. So if you always play the same thing against 1.e4 and score very badly with it, the chances that he'll choose that go up a lot. It's more game theory than statistics. – RemcoGerlich Jul 21 '17 at 7:24
  • @RemcoGerlich There are some limitations of results I will get, but I think they will be very usefull anyway. I understand your point, but I find players not exploitive enough. Ruy Lopez players won't exploit you in poisoned pawn Najdorf, even if you lost it every single time. That's not happening. At least my approach will be sound enough to evaluate our own behavior, so I keep waiting for an answer. I can compare our behavior to opponent's and find probable exploits, but first I need something to start with and it is behavior biased by observation of games. – hoacin Jul 21 '17 at 7:54
  • 4
    But then, is it not simply a matter of taking a database of the player's games, and see how often they played with move? E.g. if they played 1.d4 97 times out of 100, why isn't the probablility of 1.d4 just 97%? – RemcoGerlich Jul 21 '17 at 7:58
  • @RemcoGerlich Why would those millions of games be useless once in position we played at least once? Ok for 1.d4 you won't make mistake with that, but in 6th move of Najdorf, imagine one game played, would you really like 100% probability for move from the game? I see no reason why 0 games would use those millions of games to predict and 1+ games would just throw them away. Good prediction should use them. Word prediction will be also much more effective if you used texts of millions of people and only then update recommendations with your writing style. Our actions are not that unique. – hoacin Jul 21 '17 at 8:52
  • 1
    @hoacin MMM. Kind of the point I was making. Off-topic issues seems very important to your stated objective. Over and out. – Philip Roe Jul 23 '17 at 20:32
2

Intuitively we should look at the player's games for early moves and book moves for later. This is because as we get further in a game, the number of possibilities increases, therefore the sample size of player games decreases. In this scheme, we are assuming that the player in question behaves as the average player does in the long run given the positions they usually seek from the opening.

You could use a rule of using the player's games if there are more than 5 to use, and using other players games otherwise.

| improve this answer | |
  • 1
    I will do something similar to what you write, just instead of rule of 5 I will use some interpolation of book and player probabilities to results be more fluent. Maybe it's not that bad workaround after all. – hoacin Aug 8 '17 at 9:35

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.