Swordfish 14.2, Blue Marlin 14.2

[drc]

Another forum worthy of a look-see is:
http://www.engine-chess.com/
It looks capable, but unfortunately it is http instead of https.

[Chess_Dragon]

sarona, thank you very much for your comment!!
Also, I used your Robot-Dr.Smith imagae as my avatar here, as you can see. (I had forgotten all about that!?)

_ _ _ _ _ _

drc, very nice looking site!

It is just days old. The Admin. has an very impressive amount posted to start the various categories.

I do not think there are any posts there yet using game or position displays so I can not tell if they work there or not.

[sarona]

You are most welcome.

That avatar brings back fond memories of your analytical posts at a now defunct forum (not Rybka). I was just a kid then. You and a few others there got me hooked on computer chess engine analysis. They were always my favorite threads.

Sometimes I go back there - using archive.org - to read through the old Ivanhoe etc. testing threads. . .

[dorsz]

The next release of engines is very soon. It is mostly due to I was not fully glad with Blue Marlin 14.2 performance. I did many experiments and I found settings I can tell I am really satisfied now. New Blue Marlin will be stronger and it will have in my opinion much more interesting style.
After the release, the next one is not planned in the near future, because I will focus on other things more.

[magicianofriga]

Hi,
Thank you for the excellent work! Swordfish is quite strong in both versions, but Blue Marlin seems to miss some easy tactics sometimes. Is this why you are tuning the parameters? When do you plan to release them?
I would love to apply these changes to engines that have persistent hash like SugaR or ShashChess. What are the exact changes you are making and how do you test them? I would like to contribute to these 'Bluefish' based engines even after you stop.
Thank you again!

[drc]

Both engines suffer from the Stockfish "Freeze on a big fail high" where it will not out put anything, even when the flag falls, after a massive fail high.
This is only a problem with extremely difficult checkmate positions.
For your average tactical position or checkmate of medium difficulty, they are excellent.

Ironically, the best engine for really hard checkmate problems is Houdini with 6 man Nalimov files only. If you give it the Syzygy files, it also hits the wall.

[dorsz]

Swordfish is quite strong in both versions, but Blue Marlin seems to miss some easy tactics sometimes. Is this why you are tuning the parameters?

No. My main aim for Blue Marlin is to be wild. Obiously it is better if it is wild and strong, but unfortunately I have to compromise both things. The stronger it is the less attractive style it has. Attractiveness is a priority here. In Swordfish I take the average.

When do you plan to release them?

Tomorrow.

What are the exact changes you are making and how do you test them?

Original Bluefish by Eelco de Groot had 3 simple changes: lowered pieces values, increased king safety and additional bonus for material imbalance. Bluefish by Michael Byrne had exactly the same changes + some additional uci options and later 2 little changes in search regarding pruning.

If I compare Swordfish 14.1 and 14.2 to Bluefish I use slightly lower piece values. From v14.2 I reset material imbalance to default values. It works better. In v14.3 the kingsaftety will be slightly adjusted.

The same changes will be made to Blue Marlin 14.3 but with much different values. Without material imbalance bonus like it was in v14.1 and v14.2 I had to compensate its sacrificial and attacking style with other values. I also tested 6 diffrent versions of king safety, but I will leave it as it has been before. It was just good enough.

I tried some other changes: values for pruning like they were used in Bluefish by M. Byrne, pruning and LMR values from Crystal, but I will not use any of them because the results in test suites are maybe 1% more on average higher but they are always at the bottom in the tournaments.

At first I test quickly all changes on chosen test positions. I choose the best versions for longer tests. I also do some additional sacrificial positions test that engine must solve quickly just to know if the engine keeps the style I want. Than I play tournaments between them to find the strongest engine from these that had similar results in previous tests and at the end I test old vs new.

Both engines suffer from the Stockfish "Freeze on a big fail high" where it will not out put anything, even when the flag falls, after a massive fail high.
This is only a problem with extremely difficult checkmate positions.
For your average tactical position or checkmate of medium difficulty, they are excellent.

Ironically, the best engine for really hard checkmate problems is Houdini with 6 man Nalimov files only. If you give it the Syzygy files, it also hits the wall.

It is because I did not make any changes there. The changes should not affect the way engines are playing, but like modyfing prunning they can make them weaker.

[Chess_Dragon]

dorsz, Interesting and complete reply, thanks!

How are Swordish and Blue Martin in solving problems? Is that a priority?

[magicianofriga]

Swordfish is quite strong in both versions, but Blue Marlin seems to miss some easy tactics sometimes. Is this why you are tuning the parameters?

No. My main aim for Blue Marlin is to be wild. Obiously it is better if it is wild and strong, but unfortunately I have to compromise both things. The stronger it is the less attractive style it has. Attractiveness is a priority here. In Swordfish I take the average.

When do you plan to release them?

Tomorrow.

What are the exact changes you are making and how do you test them?

Original Bluefish by Eelco de Groot had 3 simple changes: lowered pieces values, increased king safety and additional bonus for material imbalance. Bluefish by Michael Byrne had exactly the same changes + some additional uci options and later 2 little changes in search regarding pruning.

If I compare Swordfish 14.1 and 14.2 to Bluefish I use slightly lower piece values. From v14.2 I reset material imbalance to default values. It works better. In v14.3 the kingsaftety will be slightly adjusted.

The same changes will be made to Blue Marlin 14.3 but with much different values. Without material imbalance bonus like it was in v14.1 and v14.2 I had to compensate its sacrificial and attacking style with other values. I also tested 6 diffrent versions of king safety, but I will leave it as it has been before. It was just good enough.

I tried some other changes: values for pruning like they were used in Bluefish by M. Byrne, pruning and LMR values from Crystal, but I will not use any of them because the results in test suites are maybe 1% more on average higher but they are always at the bottom in the tournaments.

At first I test quickly all changes on chosen test positions. I choose the best versions for longer tests. I also do some additional sacrificial positions test that engine must solve quickly just to know if the engine keeps the style I want. Than I play tournaments between them to find the strongest engine from these that had similar results in previous tests and at the end I test old vs new.

Both engines suffer from the Stockfish "Freeze on a big fail high" where it will not out put anything, even when the flag falls, after a massive fail high.
This is only a problem with extremely difficult checkmate positions.
For your average tactical position or checkmate of medium difficulty, they are excellent.

Ironically, the best engine for really hard checkmate problems is Houdini with 6 man Nalimov files only. If you give it the Syzygy files, it also hits the wall.

It is because I did not make any changes there. The changes should not affect the way engines are playing, but like modyfing prunning they can make them weaker.

This is quite interesting to note. Does this make both the engines good correspondence chess weapons? I am mainly looking at exciting options for LTC. Dkappe's Dark Horse net is superb, but the net arch has not been updated so I can't use it with recent versions of Stockfish. Do you have any plans of adding persistent hash? This might really help in engine performance without compromising on style. Dark Horse with SugaR AI is as strong as regular Stockfish in games and absolutely brutal when it comes to sacrificial ideas.

[dorsz]

dorsz, Interesting and complete reply, thanks!

How are Swordish and Blue Martin in solving problems? Is that a priority?

They are better than Stockfish if we are talking about sacrifices and king attack. Maybe it is better to say they are just faster in that type of positions. They can't solve positions that Stockfish can't solve like deep zugzwang, difficult mates and weird composed positions. It is a task for Crystal or Sting.

Tactics is not a priority. Priority is to have fun from the way they are playing, to have engine that plays actively, to meet more sacrifices and more attack, but I am trying to make it still reasonable. It is hard to define, tactics is not a priority, but I seek for the best values in tactical positions because it is highly related.

I also use positions to avoid unjustified sacrifices. For example this position. It is from your tournament. Swordfish 14.1 on my slow laptop from beginning likes Rxh2 and it even see an advantage for black. I had to repair it.
2kr3r/pp1n1p2/2p1pPb1/q5p1/PbBP2P1/1QN5/1P4NP/R1B2RK1 b - - 0 1