"stat score bonus" in stockfish

[xr_a_y]

Regarding this code

Code: Select all

    // Step 6. Static evaluation of the position
    if (inCheck) {
        ss->staticEval = eval = VALUE_NONE;
        improving = false;
        goto moves_loop;  // Skip early pruning when in check
    }
    else if (ttHit) {
        // Never assume anything about values stored in TT
        ss->staticEval = eval = tte->eval();
        if (eval == VALUE_NONE)
            ss->staticEval = eval = evaluate(pos);
        if (eval == VALUE_DRAW)
            eval = value_draw(thisThread);
        // Can ttValue be used as a better position evaluation?
        if (    ttValue != VALUE_NONE && (tte->bound() & (ttValue > eval ? BOUND_LOWER:BOUND_UPPER)))
            eval = ttValue;
    }
    else{
        if ((ss-1)->currentMove != MOVE_NULL) {
            int bonus = -(ss-1)->statScore / 512;
            ss->staticEval = eval = evaluate(pos) + bonus;
        }
        else
            ss->staticEval = eval = -(ss-1)->staticEval + 2 * Eval::Tempo;
        tte->save(posKey, VALUE_NONE, ttPv, BOUND_NONE, DEPTH_NONE, MOVE_NONE, eval);
    }
  

I am looking for an explanation of the very little "bonus" added only if evaluate is called and ttHit is false.
Why here ? why only here (and not also when ttHit with VALUE_NONE ?

[Joost Buijs]

I think the explanation is: 'because it tested 0.3 Elo better'.

[xr_a_y]

I think the explanation is: 'because it tested 0.3 Elo better'.

But there is always a real idea before the test try

[Deberger]

I think the explanation is: 'because it tested 0.3 Elo better'.

But there is always a real idea before the test try

The (real) idea is to give a bonus if the previous move of the opponent was historically bad,
and a penalty if the previous move of the opponent was historically good.

The Idea, Tweaked, Simplified

At this point, I wouldn't be surprised if there was more Elo gain by removing it altogether.

[xr_a_y]

Thanks for the details

[abulmo2]

The (real) idea is to give a bonus if the previous move of the opponent was historically bad,
and a penalty if the previous move of the opponent was historically good.

Is not the bonus added recursively? It looks to me that the result is to inflate deep score. Maybe it adds confidence to score returns from deeper searches.

[Eelco de Groot]

I think there is an aging factor for the stat scores, so that they do get refreshed. But more importantly, the bonus is only added in nodes with no ttHit. This means probably that the node is often near the leaves, and a new node. Once the hash probe finds a search result, you have both a search result and a static evaluation, and the static evaluation is modified with that if possible for use with nullmove etc.