Search found 610 matches

by petero2
Wed Aug 10, 2011 7:55 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Defect of a node?

I think it is easier to resolve this issue by running tests. Can you run a test one with your default acceptance probability of 1/32 ,and another with Peter's modification of selecting exactly one random move and multiply by the number of legal moves? The test should use a fixed time control select...
by petero2
Tue Aug 09, 2011 6:56 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Defect of a node?

"cost" is defined as sqrt(mean^2 + variance) My conjecture is that the correct measure to decide splitting is what I call the "defect". sqrt(mean^2 + variance)-mean I tested this method and a method that has the much simpler node expansion criteria "ply<5". This means that the in-memory tree will v...
by petero2
Mon Aug 08, 2011 9:49 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Defect of a node?

My algorithm is quite similar to UCT but it is not exactly like UCT. Here is what I do: 1. Start with an in-memory tree with only one node representing the root position. 2. Do a random walk starting from the root, until a leaf node is reached. The probability controlling which child to visit are ca...
by petero2
Mon Aug 08, 2011 5:49 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Defect of a node?

With my method, http://talkchess.com/forum/viewtopic.php?topic_view=threads&p=417425&t=39678, I get a standard deviation of 3.9826e+12 after 100e6 random walks. If I split the data in 100 chunks with 1e6 random walks in each chunk and compute the standard deviation for each chunk, I get: This shows...
by petero2
Sun Aug 07, 2011 10:57 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Defect of a node?

This has become the new number one ! It improved the leader of both the limited splitting version and the free splitting version by the tiniest of margins possible :) But it clearly makes the point that the accepted concept that "picking the node with biggest variance" is wrong. It should have been...
by petero2
Fri Aug 05, 2011 11:21 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Another piece of perft(13)

For the new readers, here some explanation of my new KnockOut commands (no "standardization candidates", definitely): "uperft" calculates the exact perft(1) upper bounds for K plies. "uperftr" estimates these by finding lower bounds for them, playing random games. "randgame" uses the (calculated or...
by petero2
Sat Jul 30, 2011 8:43 am
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: A test point for the Monte Carlo practitioners

A test point for the Monte Carlo practitioners: I don't have perft(13) yet, but I've got parts of it. [D]r1bqkbnr/pppp1ppp/2n5/4p3/4P3/5N2/PPPP1PPP/RNBQKB1R w KQkq - 2 3 For the above ply 4 position, perft(9) is 23,814,128,415,915. How well do your approximation algorithms do with this? And for ano...
by petero2
Sat Jul 30, 2011 7:25 am
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: A test point for the Monte Carlo practitioners

A test point for the Monte Carlo practitioners: I don't have perft(13) yet, but I've got parts of it. [D]r1bqkbnr/pppp1ppp/2n5/4p3/4P3/5N2/PPPP1PPP/RNBQKB1R w KQkq - 2 3 For the above ply 4 position, perft(9) is 23,814,128,415,915. How well do your approximation algorithms do with this? Using a Mon...
by petero2
Thu Jul 28, 2011 6:47 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Perft(13) betting pool

This was just an experiment to see how far you could get with non-uniform sampling. For the record, I ran a longer test over night and with a bigger hash table. The standard deviation for the average-of-1000 walks seems to have stabilized at about 3.75e16, compared to 1.05e17 for uniform sampling. ...
by petero2
Thu Jul 28, 2011 9:48 am
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: Perft(13) betting pool
Replies: 576
Views: 800079

Re: Perft(13) betting pool

If you look at the code I posted you can see that I call perfTDynWeight(13, 0), so the number of fullDepth plies is 0 Ok sorry. I don't have much time today so I only read your description (which was quite clear) and the function headers. Do you have an idea of the cpu implications of using the has...