In engines vs engine game, both think "optimistically": that they are in better shape than what their opponent thinks the evaluation score is

Question

I observed in many Tcec games of season 10 (and possibly repeats trough seasons) that white evaluation score of a position is always greater (not in terms of absolute value, but in terms of the ">" math relation) than what their oponent thinks.

Now, the engines don't both evaluate the same position, they evaluate only their own moves, so to compare, I'm comparing with the previous and following move, in average.

You can see for yourself that in the Tcec link at the beginning, down there is the "Evaluation" graph. And the white "curve" is over the black in most game most of the time.

What causes this phenomena?

Is that after making a move the position is improved for our side and the evaluation changes accordingly?

I don't think, evaluation shouldn't depend so much on side to move. Is that a engine always do the things he likes, which the opponent maybe likes but maybe not and thus differing in the evaluation?

Answer 1

It doesn't really matter what the absolute scores are, they really only serve to compare moves internally so only their relative values matter.

With one exception: positions that are evaluated at a draw get 0.0 by all these engines.

So this "optimism" factor is tuned to create a difference between a position that is merely equal and one that is a draw; if it's set to +0.1 for an equal position, then a position must be worse than -0.1 before taking a draw becomes preferable.

The usual name for this is the "contempt" setting. The current TCEC Rules say that the engine's default setting is used, so it can be set by the programmers before the tournament, and updated versions of the program can be entered for later stages of the tournament.

I think that in stage 1, an engine that has too many draws won't qualify for stage 2, so I'd expect some slightly higher than normal contempt, and this effect should be clearly lower in later rounds. But I'm a complete layman and this may be wrong.

Answer 2

Is that a engine always do the things he likes, wich the oponent maybe likes but maybe not and thus differing in the evaluation?

Yes, if you replace "like" with "judge as the only correct move".

First, "Optimism" goes in both ways: If White thinks that it stands better than Black thinks it does, then Black also thinks that it stands better than White thinks it does.

They may be a half-move away from each other in their evaluation, but modern engines are blunder-proof enough that their evaluation score usually doesn't just wildly jump around after each new move, so you can usually safely compare the average between White's evaluation of move N-1 and the one of move N+1 with Black's evaluation of move N.

The case that both sides are "optimistic" occurs much more often than the case that they are both "pessimistic" because usually, both sides employ the Minimax-Algorithm in one way or the other. In theory, if one side plays perfectly, the evaluation score will never actually improve for it on its own move: "Perfectly" entails that they will not be "surprised" to find a better move. If one exists, it will already be incorporated in the current evaluation and thus not change it. The only way that this evaluation score changes if the opponent makes a mistake.

Modern engines don't play perfectly (yet), but the idea is the same: If one engine plays a move, there are only two ways the other engine will judge it: It either agrees that this was the correct move (the difference between their evaluation scores stays the same), or it will disagree because it thinks there was a better move that it used as the score for the position until now (the difference between the scores will increase in the "optimistic" direction). The third case that it disagrees "the other way round" (and the difference increases in the "pessimistic" direction) never happens, as that would mean that it until now intentionally used a score of a move that it judged non-optimal.

Engines never "like" the opponent's moves - they only judge whether or not they were mistakes. And as soon as both engines disagree over that, both will begin to think that the other has taken a path that is worse for them, causing what you call "optimism".

What does indeed happen is reevaluation based on the now deeper search horizon, and that blurries the effect a bit again (maybe the engine to move saw something which the other couldn't see yet on its previous move, so the other actually does have to decrease its own score). But as said, most modern engines are able to avoid large jumps in their evaluation (because they realize that they missed something) for the most part, so the "optimism effect" is more influential.

Answer 3

This is normal because White has the first move advantage. Most standard openings would give very slight advantage for White due to space advantage and better pieces.

In fact, Stockfish would give you a slight advantage for White in the initial position. Try it yourself!

Please take a look:

https://en.wikipedia.org/wiki/First-move_advantage_in_chess

White has better winning statistics - both human chess and computer chess.

Answer 4

That is simply not true. In a lot of positions, black is 'more optimistic' than white. Those are just 2 separate evaluating entities with their particular evaluation terms and understanding of the position. I guess you simply happened to browse games, where white was more optimistic.