Hikaru completes trilogy with Leela

[lkaufman]

Having played a bullet chess match with LeelaQueenOdds (drawn) and a blitz chess match with LeelaKnightOdds (lost 14 to 2), Hikaru completed the trilogy today by playing a blitz (3'2") match with LeelaRookOdds. Despite taking White in eight of the fourteen games ("rook and move" odds), he lost by 9.5 to 4.5 (seven wins for Leela, two for Hikaru, five draws). Although this was noticeably better than his result at knight odds, Hikaru expected the larger handicap to make more of a difference, and was surprised to lose the match, especially by such a score. One factor was that the bot hardware was upgraded from 4090 to 5090 between the matches, but this was probably just a modest help. Although he claimed to be "on tilt" after many losses, both of his wins came near the end, so that cannot be responsible for the result. I don't believe anyone else has made an even or positive score at rook odds at 3'2" or faster against the current net, so it remains an open question as to whether anyone can surpass Hikaru's score under these conditions. The matches do seem to show that although rook odds is larger than knight odds, it is not so much larger as many might have thought.

[M ANSARI]

Yes, that was very interesting match and you could see Hikaru's frustration at the BOT not playing for the strongest moves when it comes to trading down pieces. The BOT is tuned to play the "trickiest" move rather than the "strongest" move in the position and will avoid trading down pieces unless it is absolutely necessary. This works really well against humans and probably not so well against other engines. The NN of Leela takes advantage of humans fallibility of consistently falling for a tactical trick eventually ... with Queen odds you have a lot of material advantage, so even losing a piece is not fatal ... but with only Rook odds ... you lose a piece and all of a sudden you are facing a 4000 ELO monster with only a small advantage ... no contest.

Here is the link for anyone interested

[towforce]

Yes, that was very interesting match and you could see Hikaru's frustration at the BOT not playing for the strongest moves when it comes to trading down pieces. The BOT is tuned to play the "trickiest" move rather than the "strongest" move in the position and will avoid trading down pieces unless it is absolutely necessary. This works really well against humans and probably not so well against other engines. The NN of Leela takes advantage of humans fallibility of consistently falling for a tactical trick eventually ... with Queen odds you have a lot of material advantage, so even losing a piece is not fatal ... but with only Rook odds ... you lose a piece and all of a sudden you are facing a 4000 ELO monster with only a small advantage ... no contest.

This!

Here is the link for anyone interested

Thank you. I'll take a look at that when time permits.

[lkaufman]

Yes, that was very interesting match and you could see Hikaru's frustration at the BOT not playing for the strongest moves when it comes to trading down pieces. The BOT is tuned to play the "trickiest" move rather than the "strongest" move in the position and will avoid trading down pieces unless it is absolutely necessary. This works really well against humans and probably not so well against other engines. The NN of Leela takes advantage of humans fallibility of consistently falling for a tactical trick eventually ... with Queen odds you have a lot of material advantage, so even losing a piece is not fatal ... but with only Rook odds ... you lose a piece and all of a sudden you are facing a 4000 ELO monster with only a small advantage ... no contest.

Here is the link for anyone interested

I disagree with "and probably not so well against other engines". Even against other engines, the best strategy is to avoid even exchanges if at all plausible. It's better to be down by a 9 to 8 ratio than by a 5 to 4 ratio, for example, regardless of who your opponent is. Our testing shows that at least against bots that are trained on human games, these odds bots are way better than any normal engine in odds play. Probably it is less so against traditional engines, but still clearly true. Against the human-simulating bots, there doesn't seem to be much difference between giving odds to a 2500 human or to a bot that is even (in standard chess) with that 2500 human.

[towforce]

I disagree with "and probably not so well against other engines". Even against other engines, the best strategy is to avoid even exchanges if at all plausible. It's better to be down by a 9 to 8 ratio than by a 5 to 4 ratio, for example, regardless of who your opponent is. Our testing shows that at least against bots that are trained on human games, these odds bots are way better than any normal engine in odds play. Probably it is less so against traditional engines, but still clearly true. Against the human-simulating bots, there doesn't seem to be much difference between giving odds to a 2500 human or to a bot that is even (in standard chess) with that 2500 human.

This is turning out to be an unexpectedly insightful thread!

If the key factor is not human susceptibility to "trickiness", then we have a new way to win games that might otherwise end in a draw.

If (and only if) you believe that your engine is stronger than your opponent's engine, then rework it as follows:

1. Evaluate all moves in the usual way

2. Pick the top moves which are roughly equal in eval

3. Evaluate these moves for complexity

3a. Keeps as many pieces on the board as possible

3b. Has as much engagement between opposite coloured pieces as possible

3c. Alternatively, train a separate NN to evaluate a position's complexity (this is 2025 - none of this hand-coding nonsense! )

4. Pick the move with the most complexity

5. The weaker engine will become confused and bewildered, it won't be able to see so far ahead in its game tree, it won't be able to make sense of the large number of interacting chess patterns, and hence will be much more likely to make a bad move