what is known about CSTal's inner workings?

[PK]

People that remember Chess System Tal tend to write about it from time to time, expressing their regret that no such engine exists today. Perhaps it would be a good idea to create a modern, open-source replacement of this marvel. In fact, I am toying with a thought of splitting Rodent developement into two paths: normal, competitive engine and its slower 'rabid' version. ('rabid' version has to be slower, since lazy eval conflicts with the notions of compensation and speculative sacrifices).

Right now I have a 'rabid' prototype that, with normal settings, scores a bit below Rodent 1.4 (thus being still in 2700s), and with extreme settings is capable of playing speculative sacrifices. But I'm not quite there yet.

So the big question is: what is known about how CSTal operated? I have found a lot of prose, sometimes inspierd prose (by Fernando Villegas), and some vague hints about "using eval for pruning decisions", "knowledge about constructing mating nets" and search asymmetries causing it to prefer sacrificial line even if there exists better, albeit quiet alternative. There was also a spirited rant against pre-processing engines, meaningful in its day, but painfully obvious right now. Is this really all that is known?

[F. Bluemers]

I've read somewhere it had no Qsearch,that would make sense.
I guess it had an assymetric eval otherwise it might play
defensive seeing attacking bonusses for the opponent...
Good luck with your project

[S.Taylor]

Why did people keep thinking Chris Wittington was daft, if he did such ingenius things?

[bob]

Why did people keep thinking Chris Wittington was daft, if he did such ingenius things?

It was a highly speculative program but had some serious tactical weaknesses. I've never seen anyone call him "daft" regarding his programming skills on that project. He went off the deep end once or twice dealing with people on CCC, and then the rybka issue, but it has nothing to do with his programming ability.

[PK]

how much of that search deficiency might have actually been eval deficiency? it's easy to make engine hang pieces based on overvaluating phantom compensation.

[bob]

how much of that search deficiency might have actually been eval deficiency? it's easy to make engine hang pieces based on overvaluating phantom compensation.

Since we have/had no source to look at, it is impossible to guess. If I had to use one expression to describe its behavior I would say "steer toward whatever is the most unclear". Against humans it was both entertaining and successful. Against computers, not so much. You can dig up old games on CCC by searching for cstalx, which was the account CW used to play there.

[carldaman]

People that remember Chess System Tal tend to write about it from time to time, expressing their regret that no such engine exists today. Perhaps it would be a good idea to create a modern, open-source replacement of this marvel. In fact, I am toying with a thought of splitting Rodent developement into two paths: normal, competitive engine and its slower 'rabid' version. ('rabid' version has to be slower, since lazy eval conflicts with the notions of compensation and speculative sacrifices).

Right now I have a 'rabid' prototype that, with normal settings, scores a bit below Rodent 1.4 (thus being still in 2700s), and with extreme settings is capable of playing speculative sacrifices. But I'm not quite there yet.

So the big question is: what is known about how CSTal operated? I have found a lot of prose, sometimes inspierd prose (by Fernando Villegas), and some vague hints about "using eval for pruning decisions", "knowledge about constructing mating nets" and search asymmetries causing it to prefer sacrificial line even if there exists better, albeit quiet alternative. There was also a spirited rant against pre-processing engines, meaningful in its day, but painfully obvious right now. Is this really all that is known?

This is a great idea, Pawel, and not that surprising coming from you. Having this "rabid" version should unshackle you from the "ratings rat race". It was also ChrisW's philosophy not to seek the highest possible Elo at all cost, but focus on playing style primarily.

Too many developers sacrifice style for "strength", even when they're nowhere near the top of the proverbial pile.

Regards,
CL

[S.Taylor]

Why did people keep thinking Chris Wittington was daft, if he did such ingenius things?

It was a highly speculative program but had some serious tactical weaknesses. I've never seen anyone call him "daft" regarding his programming skills on that project. He went off the deep end once or twice dealing with people on CCC, and then the rybka issue, but it has nothing to do with his programming ability.

Oh yes! I remember you saying he went "off the deep end", those many years ago!
(In a way, those were the good old days!)

[Kempelen]

I had similar ideas in the past, but it is difficult to make a "stylist" version of a engine, mainly because it is impossible to measure how this style is "attractive" or not. Only subjetive impressions......

[PK]

Yes, measuring attractiveness of a playing style is difficult. right now I'm toying with the settings, looking at the games and trying to collect some ideas, decide which factors are interesting etc. Results so far are:

- evaluating distance of pieces to enemy king (GambitFruit style) changes the flow of the game without hurting playing strength too much. This is probably an old news. What is interesting is the fact that current "tournament default" combining this tropism with a more "modern" king safety eval (with lowered weight) performs better than "modern" king safety alone. In other words, these two terms don't evaluate quite the same thing.

- increased bonus for attacking enemy pieces changes the game flow a lot, leading to risky piece placement and messy positions. This modification makes an engine quite a bit weaker, even though its play doesn't look weak. This seems a good candidate for additional knob. I have also tried to remove this factor completely, as its default contribution is rather low - removing it weakens the engine too. There might even be a gain from tuning this.

- single bits and pieces of pattern knowledge, even relatively frequent (i.e. fianchetto) don't seem to change the impression enough to warrant adding a switch. it remains to be seen whether a set of 10+ patterns would make a noticeable difference in playing style, but I am sceptical about that. nevertheless it must be tested.

- increasing weights for enemy mobility and own attack creates the playing style that seems to want something from the game and pushes its own agenda. of course it isn't Tal-like at all, but with right opening set it would make a passable... let me guess... Robert Huebner?