New commercial engine soon: Onno

[Dr.Wael Deeb]

The "size" of the seed has absolutely no meaning. Onno plays the same style with a seed of 0 and a seed of 10000. It is not possible to turn randomness off.

For me, this is the most interesting detail in technical info.
When search is randomized and un-reproducible, it is impossible to prove is the program a clone or not, based on it's moves.

Now that's what I call an ugly irony....you don't know if it's a clone or a cow....you don't have the engine yet let alone it's source code I assume....
Dr.D

[Matthias Gemuh]

For me, this is the most interesting detail in technical info.
When search is randomized and un-reproducible, it is impossible to prove is the program a clone or not, based on its moves.

That is indeed very interesting !
Yet more obfuscation will have to follow.

With the right tuning, Strelka, Toga, Cyclone and Stockfish can really kick butt.

Matthias.

[Onno Garms]

When search is randomized and un-reproducible, it is impossible to prove is the program a clone or not, based on its moves.

According to my tests, turning randomness off makes the engine weaker. Not much, but slightly.

As there seems to be demand for the option to disable randomness, I will add it. Doesn't quite fit into my concept not yet to make the adjusting screws public, but anyway.

By the way, it had been my intention to obfuscate the "real evaluation", it would have been easy to add randomness silently by just omitting the option "random seed" As the engine behaves deterministically with a fixed seed the presence of randomization could not be detected from outside.

[Onno Garms]

When search is randomized and un-reproducible,

Search is reproducable (unless you set the seed to -1). I tried to explain that on my website. Apparently that was not sucessfull.

[Leto]

According to CEGT Blitz Onno 0.12 is just as strong as Rybka 2.1 in single processor mode (both rated 2893). Not bad. You'd really want to start working on that multiprocessor coding if you want to challenge Rybka's supremacy.

[Onno Garms]

What I don't understand is why there isn't a certain seed value which produces the best results in the long run.

Seed is something different then random intensity. I see absolutely no reason why the seed value should make a difference for the strength, so I made only one test in that direction which confirmed that the seed does not matter.

[Onno Garms]

Onno requires about 1 MB on the hard disk
Jouni

Oddly, so does Fruit

That's plain wrong (as far as you refer to the opensource version of fruit). Fruit 2.1 exe file is little above 100k.

[kranium]

i wish you the best with your new engine...
and look forward to purchasing it as soon as it's released.

i envision: when (if) you catch the fish...it's gonna try to escape upstream screaming onnooooo!

at which point you'll respond oyesssss!

(maybe you chose the wrong name?)

Norm

[mcostalba]

Moreover my engine outperforms some commercial engines (on single CPU). As a freeware engine it would be among the very top freeware engines and thus contribute to the recent increase of freeware engine's level, which is a threshold at the entry level for commercial engines. I believe that this increase is harmful, because it reduces the motivation for quite a few people who have promissing engines to improve it further. So after all it reduces the chances that an engine will catch up with Rybka.

Opensource engines, namely the Fruit branches, will never catch up with Rybka IMHO because they share all their secrets but Rybka doesn't.

Motivation for open sources engines is not competiton as I had stated in a previous post, so I completely agree with you that open source engines will never catch up closed ones.

In the previous post I said is like to play poker with only your cards on the table...there is no chance to win.

But I don't agree it is harmful. I think, on the contrary, that open sources engines (not closed sources free ones that I don't distinguish from commercial ones by this point of view) can give people an higher trampoline from where to jump on.

Without good open sources engines that serve as a reference it is almost impossible for a newbie starting from scratch hoping to catch up with the best commercial's in less then a life.

As you also have written in your site, open sources engines have been helpful for you to create your own Onno engine and I really hope that it will reach its goals. So that open sources although not competitive in itselfs can be instrumental in helping new engines aimed at competition to quickly gain strength and can be a good base upon which an engine author can add its new ideas and its added value.

[Onno Garms]

Then why does random seed make the engine slightly stronger?

The theory, why randomness (not random seed) makes an engine stronger is as follows:

Static evaluation isn't exact. There might be a systematic error that over- or underestimates all positions that are examined in a search from the current position.

As the evaluation function is inherently inaccurate, a player is better off if there are more good positions he can reach. (He has several good moves rather then one.) So we want to add the number of good moves to the evaluation function.

As counting such moves is difficult to integrate in a search, instead the position values are randomly slightly modified. The more good moves a player has, the better are the chances that he can profit from the random modification.

This idea isn't new. It was discussed eithere here in the programmers forum or on the Winboard forum. (Also it is known for years that an engine with completely random evaluation beats a random mover.)

P.S. Now you can see what a mistake it was to mention this seed thing among the engine options.

Maybe it is in deed a too complicated option with only moderate user benefit.