Moderators: hgm, Rebel, chrisw
You cannot separate the two.gladius wrote:[...]But, reducing it down to ELO terms, how much comes from the idea vs the implementation choices?
jefk wrote:gladius wrote: I think there are definitely ideas out there remaining to be found that will be very impactful. Even within in the current alpha-beta framework, not talking about things like MC searches or something.Monte Carlo Tree Search is definitely not out of the question, amazing progress in Go has been a result of this. But it is far from obvious how to make it work in Chess. You cannot just copy how it's done in Go.
well you mention M(onte)C(arlo) type of searches.
Maybe such search methods within the context of an engine
can be (further?) enhanced and being used in positions
for which it will be useful; and thus leading to better moves
and higher rating(s) again.
any ideas/opinions (about this) ?
jef
I tried a very basic and naive go-like approach a few years ago and it was awful. And the current approach is working so well that it's difficult to get motivated to spend a lot of energy on this.
I can tell you a little about the characteristics of such a program. It would play a very natural and intuitive game - appearing to be a very strong positional player and perhaps not so strong tactically. In closed positions it would be far stronger and in positions where the pawns are locked up but you are a queen up the program would not believe it was wining as modern programs do - it would just appear to have a much more intuitive feel for the game. If there were any hope for winning by giving up a piece to open up the pawn structure, it would naturally do what was required with no fuss.
Although I say it would not be particularly strong tactically it would have no hesitation for sacrificial play if this increased it's odds of winning. Such a program done well would be a joy to watch play. It would also be a fighter when it was losing - giving the opponent a great deal of resistance and picking moves which made the opponents win much more difficult. On the other hand when it was dead lost with no hope of winning, it would tend to play a bit aimlessly as these types of programs need to have something to work with. As long as there is some non-trivial chance of drawing or winning they will play great.
Larry and I want to experiment with this soon - we feel that even if it's not as strong as Komodo it might be very appealing to users.
We have some very interesting plans, but we have learned that it's counter-productive to reveal them too soon and in fact we don't have a release date.Werewolf wrote:It's great to hear that you two are still full of ideas and have lots of things left to try.
I'm hoping to start some really big (100 cores) IDeA projects soon but it'd be completely pointless to start before Komodo 6 comes out.
What's the latest estimate on its release?
Twice per year is a very good move, you've really got that one right.Don wrote:
What I can say is that in the future we do plan to move to a regular twice per year predictable release schedule combined with continuous updates to paying customers and other value-added features to make it a lot more work for pirates.
I see Stockfish on all the rating lists even though it's updated every day and this does not seem to be a problem for them. How would this be any different?Werewolf wrote:Twice per year is a very good move, you've really got that one right.Don wrote:
What I can say is that in the future we do plan to move to a regular twice per year predictable release schedule combined with continuous updates to paying customers and other value-added features to make it a lot more work for pirates.
I'm not sure about the 'continuous updates' bit just because unless they're bug fixes that's a lot of testing to do on rating lists.
Another option we are considering is a subscription service to Komodo which means any time you purchase you get 6 months or 12 months of updates starting from the day they subscribe. But we still need official release versions for the testers and if we did it this way we would probably have official releases every 4 months.
But twice a year is a far, far better model than once a year. Even three times a year is possible, if you make progress.
There is obviously more to it than just Open Source. The GnuChess program has been open source for much longer but was merely mediocre. I think development and test methodology are much more important for a top level chess program. The recent distributed Stockfish test framework is a far more likely key to its success, along with the Stockfish team core principles of keeping the codebase clean and actively paring search and evaluation down to essentials.rtitle wrote:As to why Stockfish has risen to near the top and stayed there, while previously top programs (e.g. Rybka) have fallen behind, I think the answer is simply that the open-source development model is superior to the closed-source commercial programming model. In the software development world we refer to this as the "many eyeballs effect" - if you have a lot of people looking at / thinking about / contributing to a given piece of software, you ultimately win against a private company trying to compete with closed source and just a few people funded from its own limited revenues. Even once-mighty Microsoft is losing in the Windows vs Linux war, for a similar reason.
Reading this forum over the years, it's often occurred to me that most of the chess engines being developed are so similar to one another that they're basically variants of the same program. A forum like this simply would not be possible in other softeware fields. E.g., imagine if an IBM DB2 engineer posted "I'm struggling with an issue in DB2" and an Oracle engineer replied "Oh, here's how I solved the same problem in Oracle". Even if company rules allowed such a conversation to occur, it wouldn't be possible because the designs of those 2 programs are so different that the Oracle guy wouldn't know what the IBM guy is talking about. But I *do* see that sort of discussion happening all the time in the "Technical" section of this forum. So I claim you're all really developing variations of the same program. And I've often thought as a follow-up thought, "why not form an open-source project and collaborate instead?". The Stockfish team has finally done that, and I predict as a result that program will prevail vs its competitors. Open source wins in the long run.
Rich Title