pawn hash and eval tuning

[jwes]

How do people handle pawn hash when tuning? It just occurred to me if I have a pawn hash, then 99% of the time I won't have pawn terms to tune.

[Desperado]

Hi,

you can simply always compute the data when in tuning mode, like...

Code: Select all

pawn_t* PawnTable::getSlot(Position* pos)
{
    pawn_t* slot = &table[getKey(pos) & _HASHMASK];

    if( NOT_TUNING )
    if(BitUtil::hi32(getKey(pos)) == slot->lock) {
        return slot;
    }

    evaluate(pos,slot);
    return slot;
}

At least if the tuning algorithm is not sensitive to speed of the engine and you are talking of parameter tuning.

[tpetzke]

This is not a feature of having a pawn hash or not. The pawn hash only saves some computation. If you remove the pawn hash and compute everything from scratch you have the same outcome from tuning (only a bit slower).

[Desperado]

This is not a feature of having a pawn hash or not. The pawn hash only saves some computation. If you remove the pawn hash and compute everything from scratch you have the same outcome from tuning (only a bit slower).

I can't follow now. If parameters are changed for pawn evaluation you need to re-calculate the score of course.
For example if you have set 100 cp for a passed pawn but change the value to 105 cp you need to compute the outcome again. Otherwise you would retrieve the score which was computed with 100 cp.

[bob]

How do people handle pawn hash when tuning? It just occurred to me if I have a pawn hash, then 99% of the time I won't have pawn terms to tune.

Not quite. 99% of the time you won't have to COMPUTE the pawn scores, but you use them at every terminal node... So the tuning is the same whether you hash or not.

[bob]

This is not a feature of having a pawn hash or not. The pawn hash only saves some computation. If you remove the pawn hash and compute everything from scratch you have the same outcome from tuning (only a bit slower).

I can't follow now. If parameters are changed for pawn evaluation you need to re-calculate the score of course.
For example if you have set 100 cp for a passed pawn but change the value to 105 cp you need to compute the outcome again. Otherwise you would retrieve the score which was computed with 100 cp.

Pawn hash doesn't survive program termination. You play a game, or test a position, then it would be cleared, at least when using games for testing. If you are talking about some sort of automated annealing or whatever, then you would likely want to disable it.

[jdart]

For this reason I disable caching when tuning.

See https://github.com/jdart1/arasan-chess/ ... /scoring.h.

The evaluation function takes a boolean parameter:

Code: Select all

// evaluate "board" from the perspective of the side to move.
int evalu8( const Board &board, bool useCache = true );

[Desperado]

This is not a feature of having a pawn hash or not. The pawn hash only saves some computation. If you remove the pawn hash and compute everything from scratch you have the same outcome from tuning (only a bit slower).

I can't follow now. If parameters are changed for pawn evaluation you need to re-calculate the score of course.
For example if you have set 100 cp for a passed pawn but change the value to 105 cp you need to compute the outcome again. Otherwise you would retrieve the score which was computed with 100 cp.

Pawn hash doesn't survive program termination. You play a game, or test a position, then it would be cleared, at least when using games for testing. If you are talking about some sort of automated annealing or whatever, then you would likely want to disable it.

Well, the tuning method needs to be specified for the original post.

[mar]

I can't follow now. If parameters are changed for pawn evaluation you need to re-calculate the score of course.

Well, you run n threads on million(s) of positions for tuning, one run of course has the same parameters. Naturally you flush the caches between runs, I don't see any problem here.
Eval cache speeds up Texel tuning a lot, at least in my case it did.

[bob]

This is not a feature of having a pawn hash or not. The pawn hash only saves some computation. If you remove the pawn hash and compute everything from scratch you have the same outcome from tuning (only a bit slower).

I can't follow now. If parameters are changed for pawn evaluation you need to re-calculate the score of course.
For example if you have set 100 cp for a passed pawn but change the value to 105 cp you need to compute the outcome again. Otherwise you would retrieve the score which was computed with 100 cp.

Pawn hash doesn't survive program termination. You play a game, or test a position, then it would be cleared, at least when using games for testing. If you are talking about some sort of automated annealing or whatever, then you would likely want to disable it.

Well, the tuning method needs to be specified for the original post.

This is cache 101. If something invalidates the data, it can't be used again. If you allow the score weights to change, you have to take action. I would think that's pretty obvious...

Same problem with position learning. If ANYTHING changes, old position learning should be invalidated.