Tournament Time Odds

[Brian Richardson]

Just an update, FWIW.

Many people have recently been drawn to computer chess in connection with the Leela chess project, which is now effectively Lc0.
Of these, most have little experience with various GUIs and engine testing. Moreover, this has become an area with even more diverse environments: in addition to many versions of Windows and Linux with .exe or app engines, now there are more .py (Python engines), cloud systems, and even phones. Distributed efforts like OpenBench (https://github.com/AndyGrant/OpenBench) also look promising. And, a good GUI for matches may not be a good choice for tuning (CLOP, etc) or online (ICS) play. There is even yet another effort to start from scratch (Octagon).

Arena is typically the first consideration, although Fritz and various other GUIs are used. Cutechess-cli seems to be the default match harness. Naturally, I still try to suggest Winboard/Xboard.

Specifically, one tester commented on the Lc0 Discord chat/blog (link from http://testserver.lczero.org/) that they tried to use time odds for a tournament with Winboard, but although the clocks adjusted as expected, once the tournament started, the starting clocks were equal. This was confusing, so they went to cutechess-cli.

Of course, I have pointed out some of the strengths of Winboard: being able to restart tournaments, and changing "concurrency" on the fly with the number of instances running (although this is somewhat less important when GPU-bound). The main weakness for me is lack of SPRT. A minor weakness is lack of random FEN/PGN starting position support (only sequential, I think), although it is easy enough to randomize the starting positions separately, which probably supports the multiple instance approach. Finally, I suspect the somewhat different Winboard and Xboard "look and feel" is due to technical historical differences (although, to be fair, I have not tried the most recent versions).

Well, enough early Sunday morning pre-coffee kick-in ramblings.
Just thought an update might be appropriate.
Huge thanks again for all of your work.
Regards,
Brian

[Brian Richardson]

BTW: The time odds are quite important as people try to replicate the AlphaZero Stockfish results where cpu engine NPS are in the millions and GPU engine NPS are in the thousands.

There is even something dubbed the "Leela ratio":
https://lczero.libertymedia.io//2018/07/04/leela-ratio/

[H.G.Muller]

It should not be difficult to use time odds in a tournament. But you have to specify the time odds for each engine, by installing it with the special WinBoard option -firstTimeOdds N (or add it to the engine line through the Edit Engine List menu after the fact). Then every time the engine is loaded the appropriate time odds will occur.

The people that complain it doesn't work probably just specify some time odds for first and second engine in the Time Control dialog. But that doesn't work. I cannot even imagine what 'working' would mean here. You never know in advance during a tourney which engine is going to be first and which second, for a given pairing.