12

It occurred to me that presumably every chess engine's score is based on the best available continuation move, and thus every other move must be worse (or at least no better).

i.e. if there were a move that the engine could immediately assess as being higher-scored than the computer's chosen move, then that would in fact be the engine's chosen move, and thus would be reflected in the engine's score of this position.

If all that is true, then I conclude that (in terms of position scores) Human Games consist of each side making moves which always either maintain or lower their score. i.e you only increase your score on your opponents turn, when they make sub-optimal moves.

i.e. from a computer's p.o.v. a human chess game is contest of two players making moves to see who can lose positional points the least.

Is this roughly accurate?

In particular, this:

you only increase your score on your opponents turn, when they make sub-optimal moves.

4
  • You probably need to take the effect of contempt factors into account, and possibly null-move heuristics failing for certain types of positions. Engines using endgame tables often find discontinuities in evaluation when the 'true' score of the endgame table is compared with evaluated scores of the engine, making them avoid promotion or capture as the 'true' score is lower than the estimated score.
    – user30536
    Nov 16, 2022 at 16:57
  • Your third paragraph sounds like it's a matter of interpretation. Should I consider all the salary I will earn in my lifetime as earned, and then track any salary that I miss out out due to e.g. unpaid PTO? Or should I assume I get nothing and track what I earn along the way? The way you are describing "making moves which always either maintain or lower their score. i.e you only increase your score on your opponents turn, when they make sub-optimal moves", it sounds like you're applying the former idea whereby all potential is already considered as gained, and it can only be lost, not gained
    – Flater
    Nov 17, 2022 at 5:54
  • @Flater indeed, the two evaluations you mention could be called two different "philosophies", and mathematically they're mostly equivalent (as long as one is consistent in their evaluation). But the fact is that chess engines use a specific, well-defined evaluation, so for the OP's question it's not a matter of interpretation but a matter of fact.
    – Stef
    Nov 18, 2022 at 9:26
  • A similar perspective is expressed in the centipawn loss, which treats the top computer move as quasi-perfect (which is not true as the answers point out, but perhaps its close enough given human standards)
    – Scriddie
    Nov 18, 2022 at 9:51

2 Answers 2

24

In that sense an engine's evaluation is not than different from an evaluation made by a human player who is much stronger than you.

If engines were "perfect", then yes, every move you make can only maintain or worsen the evaluation before the move (note also that in this world where engines are perfect the only possible evaluations are "White wins", "Draw" and "Black wins").

However, real-world engines aren't perfect. They can only analyze a position up to a certain depth. So it's totally possible that the engine evaluates a certain position as +0.5, then make a move and after the engine evaluates the new position, the score goes up.

The same can happen when you make the engine's best move though. In the same example of a position being valued as +0.5, if you follow the top engine moves until the end of the game, it won't stay as +0.5 forever (it's either gonna be a win or a draw, or, in case the engine got it totally wrong, perhaps even a loss).

Add to that the fact that engine evaluations are not set in stone even for a given position. The specific version engine you're using, the hardware it's running on, the time you've let the engine analyze and so on will have an impact on the number you get.

3
  • 5
    +1 to this. OP is describing the decision-making process in a solved game e.g. where every move can be evaluated against the optimal strategy from every possible position.
    – Valorum
    Nov 17, 2022 at 17:31
  • To summarise, can we say the evolution of the positional score is a combination of two things? 1) The score going down when the players make suboptimal moves; 2) The score becoming more accurate as the game gets closer to its end, making it easier for the engine to evaluate the position.
    – Stef
    Nov 18, 2022 at 9:28
  • 1
    @Stef 1 is absolutely true. 2 is true most of the time even though it's possible for it to go back and forth until the engine eventually gets the right answer
    – David
    Nov 18, 2022 at 9:50
2

The answer is "no" and an example (+ detailed explanation) is given here. The brief explanation is that the depth the engine is looking at matters, and making one more move changes the depth the engine searches the main line.

You can find more examples looking at engine games. In this game, Stockfish evaluated the position after 8. Bxf6 as +1.76, then after the obvious recapture 8...Bxf6 its evaluation dropped to +1.40.

1
  • It should also be emphasized that Chess.com's engine has a high depth relative to the average strong human player, but it is shallow relative to e.g. the engines playing in TCEC. So you should expect Chess.com to give worse results than you would get from ChessBase or running your own private instance of Stockfish.
    – Kevin
    Nov 18, 2022 at 19:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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