### Question on Texel's Tuning Method

Posted:

**Wed Jul 08, 2020 4:32 pm**I have a theoretical question on Texel's tuning method. If I understand it correctly, the function takes a set of positions, and minimizes the mean squared error of the positions. The error in a position is the difference between the (adjusted) evaluation and the result of the game.

Clearly this is an effective technique, but I can't help but think that most of the positions being analyzed don't have a clear winner. In a pathological case, the game's winner may be slightly losing in the middle game. This would be asking the algorithm to tune a winning position to evaluate as a losing position, and vice versa.

Similarly, wouldn't these unclear positions dominate the contribution to mean squares error? That is, highly accurate late-game evaluations will be ignored in favor of slightly inflated early- and mid-game evaluations.

My thought was, wouldn't it make more sense to instead calculate a position's error by calculating the difference between the current position and the position, say, 5 moves into the future (perhaps an average of several future moves)? You'd need a technique to prevent it from tuning all parameters to 0, but you only need to prevent the final positions of the games from substantially lowering their scores (there are a few techniques that come to mind).

The effect of this should be that the mean squared error is now dominated by positions on the cusp of a large eval swing, and that positions throughout the duration of the game all have similar effects on contributions to mean squared error.

Is my thinking sound, or am I missing something?

Clearly this is an effective technique, but I can't help but think that most of the positions being analyzed don't have a clear winner. In a pathological case, the game's winner may be slightly losing in the middle game. This would be asking the algorithm to tune a winning position to evaluate as a losing position, and vice versa.

Similarly, wouldn't these unclear positions dominate the contribution to mean squares error? That is, highly accurate late-game evaluations will be ignored in favor of slightly inflated early- and mid-game evaluations.

My thought was, wouldn't it make more sense to instead calculate a position's error by calculating the difference between the current position and the position, say, 5 moves into the future (perhaps an average of several future moves)? You'd need a technique to prevent it from tuning all parameters to 0, but you only need to prevent the final positions of the games from substantially lowering their scores (there are a few techniques that come to mind).

The effect of this should be that the mean squared error is now dominated by positions on the cusp of a large eval swing, and that positions throughout the duration of the game all have similar effects on contributions to mean squared error.

Is my thinking sound, or am I missing something?