H. G. Muller wrote on Sat, Oct 13, 2018 06:49 AM UTC:
It is clear that the change BD -> BnD should weaken the Clobberers against any opponent, which would be undesirable against opponents that already have the upper hand. But usually such opponents would also be stronger than FIDE, and would also have to be weakened. My earlier testing with Fairy-Max suggested that the performance of armies was reasonably 'transitive', in the sense that when A > B, and B > C, then A > C by an amount approximately equal to the sum of the first two. The only anomaly was that the Nutters under-performed against the Clobberers. I conjectured that this could again be a strategic issue, namely that the more forward-directed strategy and slow backwardness of the Nutters backfires when the army has pairs of pieces (or single pieces) that can easily checkmate a King.
It is a pity that test takes so long (a common problem in computer chess...). I suppose that the new ChessV is stronger than Fairy-Max? Have you ever measured by how much? What TC are you using for these tests? Have you tried how far you can push that, without significantly affecting the result? Large depth is only needed to bring the eventual tactical punishment of strategically bad moves within the horizon, so a more advanced evaluation (e.g. for Pawn structure and King Safety) should allow faster games without play becoming so unrealistic that it is no longer a representative sampling of the pieces their tactical abilities. People are nowadays tuning their engine's evaluation at ~0.25 sec/move (e.g. 10 sec + 0.1 sec/move). It would surely save a lot of time if that would work for piece-strength measurements too.
The danger is that playing at a lower level reduces all 'excess scores', even though the ratio of these scores keep constant (so that you get the same value in terms of centi-Pawn when you divide them by the Pawn-odds score). For a twice-lower Pawn-odds score you would need 4 times as many games to get the same resolution in centi-Pawns. So I suppose there will be an optimum there. Too high a quality of play is also not good. btw; you want the typical evaluation lost per move compared to prefect play to be so large that over the duration of a game it typically accumulates to a range wider than the draw interval (say [-150cP, +150cP]), so that small departures from equality in the initial imbalance already significantly sample the won/lost range.
It is clear that the change BD -> BnD should weaken the Clobberers against any opponent, which would be undesirable against opponents that already have the upper hand. But usually such opponents would also be stronger than FIDE, and would also have to be weakened. My earlier testing with Fairy-Max suggested that the performance of armies was reasonably 'transitive', in the sense that when A > B, and B > C, then A > C by an amount approximately equal to the sum of the first two. The only anomaly was that the Nutters under-performed against the Clobberers. I conjectured that this could again be a strategic issue, namely that the more forward-directed strategy and slow backwardness of the Nutters backfires when the army has pairs of pieces (or single pieces) that can easily checkmate a King.
It is a pity that test takes so long (a common problem in computer chess...). I suppose that the new ChessV is stronger than Fairy-Max? Have you ever measured by how much? What TC are you using for these tests? Have you tried how far you can push that, without significantly affecting the result? Large depth is only needed to bring the eventual tactical punishment of strategically bad moves within the horizon, so a more advanced evaluation (e.g. for Pawn structure and King Safety) should allow faster games without play becoming so unrealistic that it is no longer a representative sampling of the pieces their tactical abilities. People are nowadays tuning their engine's evaluation at ~0.25 sec/move (e.g. 10 sec + 0.1 sec/move). It would surely save a lot of time if that would work for piece-strength measurements too.
The danger is that playing at a lower level reduces all 'excess scores', even though the ratio of these scores keep constant (so that you get the same value in terms of centi-Pawn when you divide them by the Pawn-odds score). For a twice-lower Pawn-odds score you would need 4 times as many games to get the same resolution in centi-Pawns. So I suppose there will be an optimum there. Too high a quality of play is also not good. btw; you want the typical evaluation lost per move compared to prefect play to be so large that over the duration of a game it typically accumulates to a range wider than the draw interval (say [-150cP, +150cP]), so that small departures from equality in the initial imbalance already significantly sample the won/lost range.