It is currently Mon Jun 26, 2017 3:22 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 4 posts ] 
Author Message
Offline
 Post subject: jury+ swim programming help wanted
Post #1 Posted: Wed Mar 01, 2017 8:55 pm 
Lives with ko
User avatar

Posts: 195
Liked others: 6
Was liked: 27
Rank: NR
it might be a good idea to build on top of what has already been built.

Two heads are better than one, so what i propose is a jury of bots:

each bot produces its set of suggested moves with associated adjudged win probabilities. the jury foreman can simply choose the most popular suggestion that has a win probability > 50%.


as far as playing strength is concerned, it probably wouldn't be better than massively parallel virtual supercomputer bots like alphago simply because of her superior search capability, but it should be at least as good as any of the pc bots individually. it could run either standalone, or on a mini-network via internet, one pc per bot.


Last edited by djhbrown on Tue May 02, 2017 12:28 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: jury+ swim programming help wanted
Post #2 Posted: Thu Mar 02, 2017 2:31 am 
Lives with ko

Posts: 199
Liked others: 6
Was liked: 55
Rank: KGS 3 kyu
I'm not a big fan of consensus approaches as my experience says that either they require manually tuning weights, or they lead to always playing the move of the strongest bot. Would you rather play the 9d move, or the move that 10 6d bots recommend?

Nevertheless, this got me thinking: bots could be modified to provide and incorporate what is usually called (in other settings) a "witness". E.g., given a move suggestion by one bot, along with what it thinks is the most likely follow-up, other bots (who don't consider this the best move/sequence) could easily provide their best refutation, since that refutation is only a new move sequence. This new move sequence could be incorporated in the first bot's search tree. Something like this, at the limit, would imply that they would consider the same move sequences and the only difference would be on the evaluation function - they disagree on their evaluation of a position, but there are no "blind spots".


This post by uPWarrior was liked by: dfan
Top
 Profile  
 
Offline
 Post subject: Re: jury+ swim programming help wanted
Post #3 Posted: Thu Mar 02, 2017 4:05 am 
Lives with ko
User avatar

Posts: 195
Liked others: 6
Was liked: 27
Rank: NR
UpWarrior's suggestion is sound and could be extended to a several-round 'progressive deepening' multi-party search.


Last edited by djhbrown on Tue May 02, 2017 12:29 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: jury+ swim programming help wanted
Post #4 Posted: Thu Mar 02, 2017 8:23 am 
Lives in sente

Posts: 845
Liked others: 0
Was liked: 135
djhbrown wrote:

either way, it would be interesting to see just how much consensus there is between different bots of similar strength - but it can't be expected to eliminate blind spots.


I am not so sure about that. It might depend VERY much on why they did or did not decide on different moves. And perhaps the "jury" is being called in at the wrong time.

Suppose there is engine A which operates by coming up with a set of candidate moves and then applies MCTS analysis to select the best from that set. Suppose there is engine B which does the same thing. Since different methods used (to create the set of candidate moves) these sets might be different.

Now suppose the process proceeded the way first described (each bot proceeds to analyze ITS candidate set using MCTS) but because the sets were different, could come up with different choices for best. Now that looks to me like a "jury" tied 1:1. If one of the "best"s is better than the other we have a 50% chance of selecting it.

But suppose we brought the "jury" in earlier, before applying MCTS. Instead, we form the union of those two candidate sets and apply MCTS to that. We do end up picking the best move.

If we assume that whenever the two bots don't have the same candidate set it is 50% for either of them to have the better move contained in the difference (see note below) with the original proposal we would have a 50% chance of ending up with that from a fair jury. But with the "jury" brought in earlier, would find that move. To me that seems to be finding/eliminating SOME "blind spots".

NOTE: If when there is a difference in candidate sets one of these engines always contains the better move, simply use that engine. The other would add NOTHING to the process.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 4 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 3 guests


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group