Do exact node cut offs from TT produce wrong results?

[kdjones]

Do exact node cut offs from TT produce wrong results?

I have tried writing PVS search both ways and get correct results only when an exact score returned by a TT lookup is not used to give a cut off, but only gives best move. Occasionally the cut off method produces the wrong result.

Does this mean I have a bug in the code?

An example: Fine position 70

with cut off: best move (ply 26) Ka1-b2 ( incorrect - white loses King opposition )

without cut off: best move (ply 26) Ka1-b1 ( correct - white wins a pawn )

[syzygy]

If you get a winning score for Ka1-b2, you have a bug.

Taking exact node cut offs is certainly not a bug in itself. Why would it...

[Steve Maughan]

You'll only get an error if you have path dependant evaluation terms,

Steve

[AndrewGrant]

Are you missing a condition like the following

Code: Select all

if (entry->depth >= depth && entry->type == EXACT) doTheCutOff

[kdjones]

Thanks. I believe I have equivalent code. Here is my hash table retrieval code:

Code: Select all

bool ProbeHashTable( U32 *move, int *score, int alpha, int beta, int depth)
{
	U32 index = static_cast < U32 > ( hashKey ) % hashEntries;
    hashTable.probes ++;

    if ( hashTable.entry[ index ].signature == static_cast < U32 > ( hashKey >> 32 ) ) {
        *move = hashTable.entry[ index ].move;   // for move ordering

        if ( hashTable.entry[ index ].depth >= depth ) {
            hashTable.hits ++;
            *score = hashTable.entry[ index ].score;

            if ( abs( *score ) > mateScore ) {
                if ( *score < 0 )
                    *score = depth - Infinity;
                else
                    *score = Infinity - depth;
            }
            switch ( hashTable.entry[ index ].flags ) {
                case HFnone:
                    return false;
                case HFalpha:
                    if ( *score <= alpha ) {
                        *score = alpha;
                        return true;
                    }
                    return false;
                case HFbeta:
                    if ( *score >= beta ) {
                        *score = beta;
                        return true;
                    }
                    return false;
                case HFexact:
                    return true;   // take cutoff
            }
        }
    }
    return false;
}

[kdjones]

Thanks. I don't believe I do.

[kdjones]

Yes, I get a winning score at 26 ply.

cutoff: 26. Ka1-b2 score 186 ( previous ply score 85 )

no cut: 26. Ka1-b1 score 185 ( previous ply score 85 )

The problem seems to be in the collection of the PV from the TT. The PV then is used to order the moves.