State of the Art Monte Carlo simulations and Bridgify 104
#1
Posted 2012-January-26, 00:50
#2
Posted 2012-January-28, 14:30
The reason humans can process this hand quickly, is that we "know" that 3-2 split is greater percentage than a finesse. With enough simulations, robots will figure this out. If you have east throw in some kind of crazy bid on the way to 3NT, all or nearly all of the simulations will place the heart king with east and then GIB will go back to going down as it will find two heart finesses attractive. Same thing happens in the real world with real people. Imagine if east dealt and opened 1♦,. That would surely alter your play. The bots take bids into account in their simulations.
#3
Posted 2012-January-29, 20:36
Appreciate your analysis of the monte-carlo approach: in my language this is akin to saying humans use a priori probabilities and the robots use a posteriori probabilities which require random trials.
I started this series of posts with the robots on their default settings and have now set them all to the slowest, strongest play level.
Would you agree that at their best the robots are very good, at their worst unbelievably bad? I started with the belief I could easily rank these robots in order of strength, now I do not think so; they seem equally prone to blunder or brilliance.
#4
Posted 2012-January-30, 07:54
Scarabin, on 2012-January-29, 20:36, said:
Appreciate your analysis of the monte-carlo approach: in my language this is akin to saying humans use a priori probabilities and the robots use a posteriori probabilities which require random trials.
I would argue that the difference is how the humans / robots generate their priors...
#5
Posted 2012-January-30, 16:58
Scarabin, on 2012-January-29, 20:36, said:
Would you agree that at their best the robots are very good, at their worst unbelievably bad?
Yes
#7
Posted 2012-January-30, 23:49
I think there is an essential difference to the way humans and robots approach bridge play (obvious but bear with me for just a moment, please).
The human approach is pragmatic or logical and takes into account, or should take into account, all available information.
The robot approach is restricted-random: a series of random double dummy layouts restricted by taking into account the known distribution of declarer's and dummy's hands and any information about the other hands revealed by the bidding.
This is a bit tortuous, what if any are the practical consequences?
First, the only criterion the robot uses is the double dummy question: "How many tricks will I win if I lead this card?".
Second a human player can say both these lines of play may fulfill the contract but one is superior to the other (in all circumstances). The robot will choose the line of play which figures in the majority of his samples.
Next the human will evolve a plan and keep to it until it is found to be faulty, the robot seems to do a new simulation for every card played and hence may change to a new plan merely because he has generated a different set of samples.
Another point which interests me but may have no relevance: computer random usually means pseudo random?
Having said all that I hope it's not a hopeless muddle, perhaps SoA3 will help to clarify my thinking.
#8
Posted 2012-January-31, 00:04
Quote
#11
Posted 2012-February-01, 01:13
Scarabin, on 2012-January-31, 22:13, said:
#12
Posted 2012-February-01, 09:31
#14
Posted 2012-March-16, 08:43
ahydra, on 2012-March-16, 08:32, said:
ahydra
his interpretation of instructions I gave him. i didn't want someone (anyone) to come and promote one computer bridge program over all others -- and especially over bbo's GIB program. Also what a computer program does, depends a large part on what settings it is on (slow bidding and play -- the better in theory it does). One could run hands through all the programs and then show only ones that your favorite robot shows. He decided not to list what each program did, so he uses underscores. I am not sure it had to go that far, but his way is the safest to make sure the points he wants to make gets left alone after he posted it.
#15
Posted 2012-March-16, 10:19
inquiry, on 2012-March-16, 08:43, said:
Makes sense. But then the post just looks silly - why not just omit the underscores, or (better) say "Robot A threw his king under the ace, while Robot B played perfectly right up until the point it cashed its last stop in the opponents' suit before exiting".
Then of course there's Human A, who signals with the 8 of hearts only to find himself unable to overruff dummy's 6 a trick later. Oh wait, that was me.
ahydra
#16
Posted 2012-March-16, 17:50
Apologies if these underscores appeared like a petulant protest. I find the prohibition on promoting other robots very reasonable.
Perhaps I should have gone back and edited these posts but I thought it probably was not worthwhile.
#17
Posted 2012-April-15, 13:40
Scarabin, on 2012-January-30, 23:49, said:
That's not entirely true. It's possible to incorporate different types of scoring. It's still based on the amount of tricks it will make, but it can make overtricks less (and down tricks more) important. For example, if you calculate your average tricks, then one card can make an average of 9.5 tricks while another will have an average of 9.2 tricks. However, if line one goes down in 50% of the cases and makes 2 overtricks in the other 50% of the cases, then the average score will be quite low. Compare that with line two, which makes 9 tricks all the time and makes an overtrick in about 20% of the cases. The average score will now be much higher. Based on number of tricks (for example MP play) a computer may choose line one, based on average score (for example imp play) the computer will choose line two.
Scarabin, on 2012-January-30, 23:49, said:
Next the human will evolve a plan and keep to it until it is found to be faulty, the robot seems to do a new simulation for every card played and hence may change to a new plan merely because he has generated a different set of samples.
The best computer player would generate all possible hands, analyze the DD result of each card in each of these deals, and come up with the percentage line of play. In practice that's impossible, so it has to stick to some number of generated deals. The more deals it generates and analyzes, the more accurate (and better) the results. That's also a reason why computer players are much better in the end game rather than in planning the play of the entire hand, because they can analyze hands a lot faster in later stages of the game.
Also note that it's quite difficult to tell a computer what is a possible distribution and what is not. Some people may overcall AQxxx and out, while others won't. Against certain opponents the computer should analyze this hand, against others it shouldn't. When generating a small amount of deals this may have a big influence on the result, but when analyzing all deals this would have a negligible influence.
Scarabin, on 2012-January-30, 23:49, said:
The amount of deals and possible distributions is large enough for this to be completely irrelevant.
#18
Posted 2012-April-16, 22:04
I have now discovered another program, Oxford bridge, which incorporates pragmatic reasoning and propose to attempt a statistical analysis of these 5 programs performance on hands from educative software.
#19
Posted 2012-October-04, 00:26
Piotr has also written a single dummy solver which I have downloaded but cannot run because it cannot find some DLL. I would ask Piotr for help but I do not speak Polish and cannot find his email address.
#20
Posted 2012-October-04, 01:38