Subject: little fun with TSCP Fri Feb 12, 2021 9:54 am
One evening I have modified venerable Tom's Simple Chess Program, so that its evaluation became just the piece values and piece-square tables from Rofchade 1.0, scaled according to game phase. Tom Kerrigan has just uploaded it at http://www.tckerrigan.com/Chess/TSCP/Community/
You can view it as a demonstration of the power of tuning - Ronald Friederich did an extremely good job with Rofchade's tables and new TSCP scores like +200 Elo against the original.
Beside that, it is really odd opponent to play against.
Admin, Mclane and matejst like this post
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Sat Feb 13, 2021 8:14 pm
Very interesting, I am going to spend some time on this.
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Sat Feb 13, 2021 8:35 pm
BTW, are you going to use the tables in Rodent?
nescitus
Posts : 46 Join date : 2020-12-01
Subject: Re: little fun with TSCP Sat Feb 13, 2021 10:29 pm
To use these tables in Rodent I would need to retune at least pawn table, because it compensates for the lack of passer evaluation. Come to think of it, it would be an interesting challenge.
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Sat Feb 13, 2021 10:36 pm
Maybe it's useful for search decisions also.
Mclane
Posts : 2951 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: little fun with TSCP Sun Feb 14, 2021 12:13 am
I need executables. So that I can put all the versions in my arena tournament.
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: little fun with TSCP Sun Feb 14, 2021 10:57 am
Ed, have you ever thought about really improving Pro Deo? I see that you are reticent even to implement ideas that are widely in use. On the other hand, Pro Deo offers a lot of infos, utilities, comfort not present in other engines. It is perhaps only me, but I would rather use a modernized Pro Deo or even a NN trained on Pro Deo eval, even with a slower search, than a SF 3500.
Inspired by the FF2 controversy, I made a tour of the available commercial offers, and it seems that most users do, in fact, use engines the way I do: for analysis, or weakened for training games.
I would not be surprised if most of the Komodo sales went through ChessBase (I do not see how a standalone Komodo + 1y of subscriptions is worth the CB Komodo + GUI + opening book + sample database), and I absolutely do not see what SF can add to a mid-level player (from 1800 to 2400 Elo) compared to any other top engine.
So, why not implement the already common knowledge in Pro Deo, while keeping all its functionalities, personalities, book options, which are all things that matter a lot for a chess player?
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Mon Feb 15, 2021 1:58 pm
For a change this is going to be long
Your are 28, happily married for 4 years, 2 lovely daughters and a good job, nice house. Then your wife says: You, look for a hobby! And so I did. It was the time (1978) the first dedicated chess computers entered the warehouses. I bought one because chess was the hobby of my youth which abruptly ended at age 18 because of other interests (girls).
I forgot the name of my first chess computer (can't find it on the internet!), it played horrible and soon replaced it for a Fidelity, maybe not CC1 as stated elsewhere but a later version that looked exactly the same as the picture elsewhere. Anyway, these things got my fascination, I was back in my youth again and my fascination for chess.
Next purchase was the Voice Challenger, costed 1000 dutch guilders, comparable with (say) $2000 nowadays. My hobby became an expensive one. And so I decided to buy one of the first personal computers (TRS-80), bought Sargon, Gambit-80 by Wim Rens. Much cheaper on the long term.
Not so long later I wondered if I could not write a chess program myself, I was a reasonable chess player and I could program and if I managed I never had to buy anything new and expensive In 1982 I participated in the Dutch championship and to my surprise ended as shared second together with Gambit-81.
About 2 years later I was approached by Hegener & Glaser (Mephisto) to work for them with the demand that I would work for them as a full job. So there you are, you are offered a dream job, making your hobby your job but at the same time you are forced to give up your good job, career. And... with a great risk you could fail because playing strength was all. It happened to Ulf Rathsman (MM2), he was replaced by me. That's how it goes in this business, if you don't deliver you are out. And with 5 mouths to feed in the between time that's something carefully to consider before making a decision.
I decided to go for the adventure anyway because Hegener & Glaser offered me good conditions and excellent financial support till I was able to produce a good enough program that they could sell, which was the Rebell 5.0 module in 1986. Thereafter I was on my own, with the pressure to perform and deliver a stronger engine. So in the years after (till 1992/93) twice a year I got a phone call from Munich if I had a stronger engine because they wanted to release a new chess computer. In all these years I (fortunately) could deliver.
Disaster... Hegener & Glaser went bankrupt and I seriously considered to stop, look for a normal job and life. Sales stopped with a significant financial loss as result and on top of that I recently hired a good friend of mine (Rob Kemper) on the payroll which I then had to fire again, something I could not bear. And so we talked and decided to put all our energy on the PC. I converted the RISC engine (ChessMachine) to the PC and Rob would program an interface for the PC, which took about a whole year for us both, resulting in Mephisto Gideon and later a new interface with Rebel 6.0. Meanwhile 2 more years of poor sales and financial losses, but I could have it and with Rebel 8.0 the sun started to shine again up to a point I had hire a third person on the payroll for administration, handling and shipping orders. But still, which is actually the main point of this post, with the thought in mind that every new year could be the last year.
Finally in 2003 I had enough, I got a burn-out, kept a promise by delivering a Windows version (Rebel 12 by Lokasoft) and (publicly) resigned from the competition. That's what 20 years of pressing yourself did to me. No more. And the good new is, I got my hobby back.
So I guess this more or less answers your questions, improving the engine lost its priority and plundering open sources is not my style. I retired just at the right moment, the Fruit 2.1 open source release in 2004 and the elo explosion that followed in many engines. Not my cup of tea.
Instead I (still) love to do other computer chess related things.
Off my soap box.
Mclane, Brendan, TheSelfImprover, Piman, mwyoung, matejst and Damir Desevac like this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: little fun with TSCP Mon Feb 15, 2021 4:06 pm
Ed,
It was my best shot to try to convince you (and it was not the first time), and, unfortunately, it did not work. Nonetheless, I would like to emphasize that I am very grateful for Pro Deo the way it is now. Just the book option with its stats is very helpful in creating my own repertoire.
And, when I have a look at the CCRL list, Pro Deo is already 250 Elo stronger than Rebel. Years by years, experiment by experiment, it becomes stronger. So, there are still original paths.
I was also curious to know how long it would be needed for an experienced programmer to improve its engine with ideas that are now common knowledge. And what is the floor now? Is it 2900, 3000, 3100? We see that a good engineer, but without experience, like Andreas Matthies made a very strong and pleasant engine in a very short time. He needed time to learn the fundamentals, but since then he did quite well.
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Mon Feb 15, 2021 7:25 pm
Rubichess is great, the latest example is Chris, just look what he did in a very short time!
Mclane likes this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: little fun with TSCP Mon Feb 15, 2021 9:13 pm
Ed,
Andreas did a great job with Rubi -- it is one of the few "modern" engine I really like (though lately he is a bit too much focused on ratings and rankings). And yes, I was thinking about Chris too. I think that for an experienced author it is enough to know some technical fundamentals (e.g. bitboards) and the principles used in the last ten years to easily achieve 3000 Elo.
I guess it was the way John Stanback also took, and somehow, he managed to make a full circle with Wasp: it started like a faster Zarkov 6.5, to become unpalatable around the 3.xx versions, and return to a similar -- I find it very reasonable -- evaluation now, with Wasp 4.5. John described the way he tuned the engine parameters and it could be an important, unnoticed revelation.
Unfortunately, when I pit it against Komodo, it is obvious that Komodo's eval is much more elaborated: Wasp often had a clear advantage in the opening to lose in the middlegame, or the ending. I saw him exchange rooks and go from a drawn position in a rook ending to a lost position in a pawn ending.
The obvious problem with the evaluation is that one has to code complex chess knowledge, and it is much easier to use TBs or NNs. But I am still a fan of knowledge based engines -- and I do not include NNs in it, because we seem not to know how they work.
That's the reason I would like to see some of my old programs, like HIARCS Bareev, Pro Deo, or the recently released Gandalf 7 with modern search and tuned evaluations. Zarkov got 400-500 Elo points, I think they would also get the same amount in a very short lapse of time, and that they would be a pleasure to use. Thorsten often rants about new engine being just calculators, and, while I do not agree completely, I guess there is some truth in his complaints.
Mclane
Posts : 2951 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: little fun with TSCP Mon Feb 15, 2021 9:59 pm
Ed is a legend IMO. I will never forget when I first saw him with jan in cologne with the hair-dryer trying to give the hardware a little fresh cold air. What he made out of this 6502 platform was really a new level after Rathsman was overtaken.
As he said, he was always capable to give more strength / progress in his versions. The DOS PC programs had such a cool GUI, perfect without windows.
TheSelfImprover, Piman and matejst like this post
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Mon Feb 15, 2021 11:11 pm
nescitus wrote:
To use these tables in Rodent I would need to retune at least pawn table, because it compensates for the lack of passer evaluation. Come to think of it, it would be an interesting challenge.
I implemented the code in ProDeo in a raw way without paying attention to speed and search that needs to be modified as well. It turns ProDeo into a gambit player. A couple of examples ProDeo always has (had) trouble and now solves almost instantly. Unfortunately PGN4WEB is not working at this moment.
Example-1 r1b2rk1/p2nbqpp/p3p3/2ppPpB1/N2P1N1P/8/PPP2PP1/R2Q1RK1 w - - bm c4;
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Tue Feb 16, 2021 11:41 am
Alright, of course there was a bug, fixed it and playing a 5000 game match at 40/40, the new version after 500+ games is leading
Code:
[ProDeo Pesto] vs ProDeo [5000] [40/40] 195-181-156 (532) match score 285.5 - 246.5 (53.7%) (+24 elo) Won-loss 195-156 = 39 (532 games) draws 34.0% LOS = 98.1%
This while Pesto currently is 30% slower in nodes and several search adaptions need to be made. Seems this new PST approach has a future.
Pawel, Chris, are you reading?
matejst likes this post
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Tue Feb 16, 2021 12:47 pm
matejst wrote:
Ed,
Andreas did a great job with Rubi -- it is one of the few "modern" engine I really like (though lately he is a bit too much focused on ratings and rankings). And yes, I was thinking about Chris too. I think that for an experienced author it is enough to know some technical fundamentals (e.g. bitboards) and the principles used in the last ten years to easily achieve 3000 Elo.
I guess it was the way John Stanback also took, and somehow, he managed to make a full circle with Wasp: it started like a faster Zarkov 6.5, to become unpalatable around the 3.xx versions, and return to a similar -- I find it very reasonable -- evaluation now, with Wasp 4.5. John described the way he tuned the engine parameters and it could be an important, unnoticed revelation.
Unfortunately, when I pit it against Komodo, it is obvious that Komodo's eval is much more elaborated: Wasp often had a clear advantage in the opening to lose in the middlegame, or the ending. I saw him exchange rooks and go from a drawn position in a rook ending to a lost position in a pawn ending.
The obvious problem with the evaluation is that one has to code complex chess knowledge, and it is much easier to use TBs or NNs. But I am still a fan of knowledge based engines -- and I do not include NNs in it, because we seem not to know how they work.
That's the reason I would like to see some of my old programs, like HIARCS Bareev, Pro Deo, or the recently released Gandalf 7 with modern search and tuned evaluations. Zarkov got 400-500 Elo points, I think they would also get the same amount in a very short lapse of time, and that they would be a pleasure to use. Thorsten often rants about new engine being just calculators, and, while I do not agree completely, I guess there is some truth in his complaints.
Well, there is still progress possible with the old school approach, from the CCRL 40/15
Pro Deo 3.0 2588 15 15 1203 42.5% 2645 37.6% Pro Deo 2.2 2538 10 10 2626 48.2% 2552 33.5% Pro Deo 2.0 2517 13 13 1554 47.1% 2539 33.7%
matejst likes this post
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: little fun with TSCP Tue Feb 16, 2021 1:36 pm
Admin wrote:
Alright, of course there was a bug, fixed it and playing a 5000 game match at 40/40, the new version after 500+ games is leading
Code:
[ProDeo Pesto] vs ProDeo [5000] [40/40] 195-181-156 (532) match score 285.5 - 246.5 (53.7%) (+24 elo) Won-loss 195-156 = 39 (532 games) draws 34.0% LOS = 98.1%
This while Pesto currently is 30% slower in nodes and several search adaptions need to be made. Seems this new PST approach has a future.
Pawel, Chris, are you reading?
Yup. My eval is at the point where any change to it is a regression
At the moment am experimenting with Ferdy/lakas GitHub code for search tuning
matejst likes this post
nescitus
Posts : 46 Join date : 2020-12-01
Subject: Re: little fun with TSCP Tue Feb 16, 2021 5:42 pm
I am reading. Two tests with "pesto" tables (and pawn rables retuned) finished 15 Elo below default. I guess that with serious tuning of all pawn-related parameters I would gain something. My current guess is that "pesto" tables mess up also with king's pawn shield, because new version got mated much more often. Third test is in progress, using my old tables for pawns and "pesto" for everything else.
EDIT: This comes out as almost equal
Last edited by nescitus on Tue Feb 16, 2021 8:15 pm; edited 1 time in total
matejst likes this post
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Tue Feb 16, 2021 8:08 pm
Chris Whittington wrote:
Admin wrote:
Alright, of course there was a bug, fixed it and playing a 5000 game match at 40/40, the new version after 500+ games is leading
Code:
[ProDeo Pesto] vs ProDeo [5000] [40/40] 195-181-156 (532) match score 285.5 - 246.5 (53.7%) (+24 elo) Won-loss 195-156 = 39 (532 games) draws 34.0% LOS = 98.1%
This while Pesto currently is 30% slower in nodes and several search adaptions need to be made. Seems this new PST approach has a future.
Pawel, Chris, are you reading?
Yup. My eval is at the point where any change to it is a regression
At the moment am experimenting with Ferdy/lakas GitHub code for search tuning
My eval is tuned as well, no elo gain possible, but novelty can do that.
Laskas is very interesting, I am awaiting the results.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: little fun with TSCP Tue Feb 16, 2021 8:29 pm
Admin wrote:
Chris Whittington wrote:
Admin wrote:
Alright, of course there was a bug, fixed it and playing a 5000 game match at 40/40, the new version after 500+ games is leading
Code:
[ProDeo Pesto] vs ProDeo [5000] [40/40] 195-181-156 (532) match score 285.5 - 246.5 (53.7%) (+24 elo) Won-loss 195-156 = 39 (532 games) draws 34.0% LOS = 98.1%
This while Pesto currently is 30% slower in nodes and several search adaptions need to be made. Seems this new PST approach has a future.
Pawel, Chris, are you reading?
Yup. My eval is at the point where any change to it is a regression
At the moment am experimenting with Ferdy/lakas GitHub code for search tuning
My eval is tuned as well, no elo gain possible, but novelty can do that.
Laskas is very interesting, I am awaiting the results.
Lakas. It’s not to be confused with Laskos! The vanilla version seems only in effect to play batches of games with master and then selects the batch that performed best.
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Wed Feb 17, 2021 12:19 am
nescitus wrote:
I am reading. Two tests with "pesto" tables (and pawn rables retuned) finished 15 Elo below default. I guess that with serious tuning of all pawn-related parameters I would gain something. My current guess is that "pesto" tables mess up also with king's pawn shield, because new version got mated much more often. Third test is in progress, using my old tables for pawns and "pesto" for everything else.
EDIT: This comes out as almost equal
Currently I am doing it double, my own PST plus PESTO, just to watch the effects.
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Wed Feb 17, 2021 12:21 am
Chris Whittington wrote:
Admin wrote:
Chris Whittington wrote:
Admin wrote:
Alright, of course there was a bug, fixed it and playing a 5000 game match at 40/40, the new version after 500+ games is leading
Code:
[ProDeo Pesto] vs ProDeo [5000] [40/40] 195-181-156 (532) match score 285.5 - 246.5 (53.7%) (+24 elo) Won-loss 195-156 = 39 (532 games) draws 34.0% LOS = 98.1%
This while Pesto currently is 30% slower in nodes and several search adaptions need to be made. Seems this new PST approach has a future.
Pawel, Chris, are you reading?
Yup. My eval is at the point where any change to it is a regression
At the moment am experimenting with Ferdy/lakas GitHub code for search tuning
My eval is tuned as well, no elo gain possible, but novelty can do that.
Laskas is very interesting, I am awaiting the results.
Lakas. It’s not to be confused with Laskos! The vanilla version seems only in effect to play batches of games with master and then selects the batch that performed best.
Already possitive results?
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: little fun with TSCP Wed Feb 17, 2021 10:16 am
Admin wrote:
Chris Whittington wrote:
Admin wrote:
Chris Whittington wrote:
Admin wrote:
Alright, of course there was a bug, fixed it and playing a 5000 game match at 40/40, the new version after 500+ games is leading
Code:
[ProDeo Pesto] vs ProDeo [5000] [40/40] 195-181-156 (532) match score 285.5 - 246.5 (53.7%) (+24 elo) Won-loss 195-156 = 39 (532 games) draws 34.0% LOS = 98.1%
This while Pesto currently is 30% slower in nodes and several search adaptions need to be made. Seems this new PST approach has a future.
Pawel, Chris, are you reading?
Yup. My eval is at the point where any change to it is a regression
At the moment am experimenting with Ferdy/lakas GitHub code for search tuning
My eval is tuned as well, no elo gain possible, but novelty can do that.
Laskas is very interesting, I am awaiting the results.
Lakas. It’s not to be confused with Laskos! The vanilla version seems only in effect to play batches of games with master and then selects the batch that performed best.
Already possitive results?
Yes, yesterday, with my modified version of it. Got 51% from about two hours work including coding, against prior master 60,000 games with the most ridiculous weightings imaginable. Just make no sense at all. Still, if they work, they work. This was all of qsearch btw, which had previously been very resistant to tuning improvements.
Admin Admin
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Wed Feb 17, 2021 3:52 pm
Posts : 2538 Join date : 2020-11-17 Location : Netherlands
Subject: Re: little fun with TSCP Fri Feb 19, 2021 8:15 am
Code:
Score of ProDeo 3.1 vs ProDeo 3.0: 2009 - 1271 - 1720 [0.574] 5000 ... ProDeo 3.1 playing White: 1046 - 587 - 867 [0.592] 2500 ... ProDeo 3.1 playing Black: 963 - 684 - 853 [0.556] 2500 ... White vs Black: 1730 - 1550 - 1720 [0.518] 5000 Elo difference: 51.7 +/- 7.8, LOS: 100.0 %, DrawRatio: 34.4 % Finished match
Finished testing, will make preparations for release with the advice for every programmer to give it a 2-3 day work try.
Mclane, matejst and nescitus like this post
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: little fun with TSCP Fri Feb 19, 2021 11:23 am
I got myself into a huge developmental mess as a result of a) working on data in the giga-number range (processing often takes time measured in days) and b) (don’t get envious) having three dev machines, two with 6 cores, 32Gb and 2080 and one with 64 cores and 256Gb.
This means one or better two of them processing data and or testing, and one writing software to use the data or process it some more. Which in turns means I’m working on three sub-projects at once, and have to keep switching between them as more data cones down the pipeline.
Compare to old way of work, which was one thing at a time, it’s quite difficult to keep three things at various stages in your head at the same time. So, I now instituted a project notebook which tells me what I am doing, what needs doing next and so on. That seems to work but more problem arising based around which directories on which machines contain which data and at which stage. That’s a whole new element of chaos.