Engine that allows tuning its evaluations?

[Marek Soszynski]

I'm looking for a (strong) UCI engine that allows the tuning of its numerical evaluations. A good example would be a multiplier setting. While the default setting is "1", when the multiplier is set to "1.5" then a position normally evaluated at, say, "+0.76" centipawns would be evaluated and displayed and processed as "+1.14". In other words, advantages (and disadvantages) would be exaggerated when the multiplier is positive.

[krazyken]

Glaurung has about a billion UCI options for tuning eval. But I don't think there is any engine that has a multiplier like that right now.

[Edmund]

What is the purpose of this multiplier for the total evaluation?

If you want to tune an engine, you should try adjusting individual parts of the evaluation function, such as increasing the relevance of kingsafety.

Adjusting the total evaluation won't have much effect on the tree searched. The only places where I could imagine any differences would be, where absolute numbers are used for cutoffs. Eg Delta Cutoff in Quiescence search. There one uses the assumption, that making a certain move won't increase the value of the position by more than a static amount.

[Dann Corbit]

If everything is scaled by the same linear factor the net effect will be zero.

In other words if a pawn is worth 1000 then a knight will be 3200 or so.

[Stan Arts]

What is the purpose of this multiplier for the total evaluation?

Not sure if the OP intended this, but if you exclude material from the multiplier, you can roughly determine how speculative your program will play. And in my engine that's how my speculative option works. (Although a few evaluationparameters are excluded) A program that allows tuning of evaluation should ofcourse also adjust it's safety margins in search accordingly. (for example those delta pruning margins.)

For example internally my multiplier currently runs on 2, because in the beginning I was afraid to give big bonusses. My speculative option allows this to be set to resp. 4 and 6, and on 6 it will feed pawns and pieces at random for a little development, passed pawn or kingattack. Just how chess should be. (and about 400 Elo weaker. )

Stan

Ps. On a sidenote considering delta pruning, last weeks I have experimented with some sort of intelligent delta pruning in quiescent search, because I don't do SEE pruning. Fe. capturing a knight won't be able to cause a huge positional evaluation jump, while capturing a passed pawn or queen likely can. Along with also always giving recaptures on the previous square a high margin, (that really reduces risk of completely missing positional cases) this turned out to be quite safe and works well for me.

[Marek Soszynski]

>If everything is scaled by the same linear factor the net effect will be zero.

>What is the purpose of this multiplier for the total evaluation?

It is strange that in many engines it is possible to alter all kinds of settings - except this one.

There are at least three situations where a multiplier setting could be useful.

1) It is always possible that an engine has a "wrong" relative evaluation. Older versions of Shredder are an example of an engine exaggerating advantages compared to the evaluations of other engines.

2) Altering the multiplier could affect the length of games. A high (positive) multiplier would cause the engine to resign earlier, and could cause a different attitude to draw offers and agreements.

3) There are some special circumstances where the absolute value of evaluations could be important. For instance, in Aquarium IDeA which converts an engine's evaluations from centipawns to percentages and back again.

[Anil]

You may be able to do it only as a visual reference in Arena GUI:
Right Click on Permanent Statistics and then click "Scale the statistics".
Now, slide the slider to the left (to 1).

The engine evaluation scores would now form taller towers even for shorter values.

Don't know if this is what you were looking for.

[Dr.Wael Deeb]

You may be able to do it only as a visual reference in Arena GUI:
Right Click on Permanent Statistics and then click "Scale the statistics".
Now, slide the slider to the left (to 1).

The engine evaluation scores would now form taller towers even for shorter values.

Don't know if this is what you were looking for.

No,he means real tunning of the engine's evaluation itself....

[Matthias Gemuh]

You may be able to do it only as a visual reference in Arena GUI:
Right Click on Permanent Statistics and then click "Scale the statistics".
Now, slide the slider to the left (to 1).

The engine evaluation scores would now form taller towers even for shorter values.

Don't know if this is what you were looking for.

No,he means real tunning of the engine's evaluation itself....

He wants to tune the final score _after_ search.
He is aware that this will not affect the engine's search: search depth, best move, search time, nodes, speed, etc, will not change and he knows it.

Matthias.