Internet engine matches

[Krzysztof Grzelak]

It is better to play matches on just 1 computer without pondering, even if you only use half the number of available cores per engine. Windows 11 slows down incredibly with disk I/O; if one of the engines starts loading its endgame databases during pondering, the other engine suffers tremendously from it.

It can also happen that both engines try to load their databases at the same time, which is not exactly ideal either. With 6P databases you won't suffer from this as much, but with 7P and 8P databases it can have a major impact.

As a 'specialist' in this field, you should know that.

Thank you for the information you’ve provided.

[Joost Buijs]

Hi Joost,

Great, all draws. Ares done well. I plan to run a tournament these day and i will be glad to include this latest version.
Congratulations.

Friendly, Sidiki

Hi Sidiki,

I don't think it is very useful to send new versions each time, the difference in strength is so small that you won't notice it in a tournament anyway. The problem with Draughts is that games between strong programs on modern hardware will always end in a draw. Without changing the rules of the game these tournaments don't seem very useful anymore.

I will send you Ares version 2.0 when it's ready.

Joost

[Sidiki]

Hi Joost,
It's true, i think that with the player or programme with most pawn at the end of the game must/could be the winner.
Or if someone create another search and eval.

I will be glad to test and a tournament with this version.

Friendly, Sidiki

[Joost Buijs]

I ran another 3 single-core matches with Ares against Kingsrow using a time control of 30 seconds plus a 0.5-second increment for 90 moves, both with no opening book and a 6-piece endgame database (6P EGDB). The result was that 1 match ended in a complete draw, and in the other 2 matches, both programs scored 1 point, making them completely equal.

I do not expect there is much room left for improvement, except for creating a GUI to make the program more user-friendly.

Perhaps reinforcement learning could still offer something; Google has shown in the past that for games like Chess and Go, reinforcement learning outperforms supervised learning (which is, of course, training on a pre-played set of matches).

According to Google: Transitioning from supervised learning to self-play reinforcement learning allows an engine to discover unique strategies unconstrained by human biases or flawed training datasets.

I've added the .pdn files of the 3 matches.

[BertTuyt]

Yesterday I tested the new damage2026 GUI, and played (with the damage 18.6 engine) 10 games against KR.
Settings 1 core, 6p db, book active, and 5 minutes for 80 moves.
To my surprise, damage won the 10th game, all others were a draw.
But again these are freak accidents.
Think only changing the rules will bring back some tension.

Bert

[Joost Buijs]

Yesterday I tested the new damage2026 GUI, and played (with the damage 18.6 engine) 10 games against KR.
Settings 1 core, 6p db, book active, and 5 minutes for 80 moves.
To my surprise, damage won the 10th game, all others were a draw.
But again these are freak accidents.
Think only changing the rules will bring back some tension.

Bert, I agree.

To say the least, all these draws are getting a little bit boring. Maybe I will try something with reinforcement learning, and a policy network, but I don't want to add that to the old version of the engine anymore.

The new version is still not able to play games, every time I want to work on it I feel a rather large resistance, and I ask myself what on earth I am doing.

Joost

[BertTuyt]

Joost, the problem we face, is that neural networks, policy networks, and reinforcement learning are the way forward and the future.
Current and future processors (think ZEN 7 will include ACE, with AMX), will even support this more, as there is a need for ultra-fast NN inference with limited/no latency on the main processor.
Only disadvantage, you will still end with 157 draws, and the engine will only work on your (latest) computer.
So basically no one can use your engine.......

Bert

[MichelG]

One way out is playing faster games, 1 minute is long already, but Bert's 5 minutes is like an eternity.

Matches at 1 to 5 seconds per game are much better in my opinion, and you don't need to change the rules.

Only thing is, that this can be difficult to enforce from a hub standpoint, because of communication and GUI overhead.

Michel

[BertTuyt]

Michel, although you are right.......
But I do not like the idea that the only way to see differences is to go to milliseconds games.
Or that we all acquire vintage computers from the 1990-1995 era, like a 386, and limit ourselves to Borland Compilers.
And that we conclude that at 6-ply we really see interesting games, with many decisive games.

Bert

[BertTuyt]

Michel, another comparison.
It is like restricting the Formula-1 cars to 10 liter fuel only.....

Bert

[MichelG]

Dragon supports fractional times, down to 0.5 minutes (see my suggestion in 2014). Damexchange allows down to 0.6 seconds/game.

Are there any other programs that support fractional minutes? Ill remove the 30 seconds limit on next version.

Michel

[MichelG]

Or that we all acquire vintage computers from the 1990-1995 era, like a 386, and limit ourselves to Borland Compilers.
And that we conclude that at 6-ply we really see interesting games, with many decisive games.

Run a competition on an c64 emulator

Michel

[BertTuyt]

Michel, good idea, and another requirement only programmed in BASIC, no machine language allowed.

Bert

[MichelG]

It is like restricting the Formula-1 cars to 10 liter fuel only.....

Yeah, but now it it like a 100 meter running match and then rounding to the nearest second to determen who wins.

[Joost Buijs]

Joost, the problem we face, is that neural networks, policy networks, and reinforcement learning are the way forward and the future.
Current and future processors (think ZEN 7 will include ACE, with AMX), will even support this more, as there is a need for ultra-fast NN inference with limited/no latency on the main processor.
Only disadvantage, you will still end with 157 draws, and the engine will only work on your (latest) computer.
So basically no one can use your engine.......

Bert, within a few years from now, Intel and AMD will both have AVX-10.2 which includes ACE, Intel will also put it back in their consumer processors. Removing AVX-512 from their consumer processors was a very big mistake from Intel. Also the big-little approach is something that people don't want in their desktop PC's. It's good for a smartphone, and maybe for a laptop, but most people don't want a desktop computer that can only run at one leg.

I agree, with faster computers the games will be more draw than ever.

Joost