Monte Carlo vs. Law of Vacant Places

[virgosrock]

[hv=http://www.bridgebase.com/tools/handviewer.html?lin=pn%7CVirgosRock%2C1565566382%2CRobotN%2CRobotE%7Cst%7C%7Cmd%7C3S79TJH5679DJAC5TA%2CS68H4QKD3578C348J%2CS245KAH23TJD4QKC6%2C%7Crh%7C%7Cah%7CBoard+1%7Csv%7Co%7Cmb%7C1S%7Can%7CMajor+suit+opening+--+5%2B+%21S%3B+11-21+HCP%3B+12-22+total+points+%7Cmb%7Cp%7Cmb%7C3S%7Can%7CLimit+major+raise+--+4%2B+%21S%3B+10-12+total+%7Cmb%7Cp%7Cmb%7C4S%7Can%7C5%2B+%21S%3B+14%2B+HCP%3B+15-20+total+points+%7Cmb%7Cp%7Cmb%7Cp%7Cmb%7Cp%7Cpg%7C%7Cpc%7CHA%7Cpc%7CH9%7Cpc%7CH4%7Cpc%7CH2%7Cpg%7C%7Cpc%7CH8%7Cpc%7CH5%7Cpc%7CHQ%7Cpc%7CH3%7Cpg%7C%7Cpc%7CHK%7Cpc%7CHT%7Cpc%7CC2%7Cpc%7CH6%7Cpg%7C%7Cpc%7CS6%7Cpc%7CSA%7Cpc%7CS3%7Cpc%7CS7%7Cpg%7C%7Cpc%7CD4%7Cpc%7CD2%7Cpc%7CDJ%7Cpc%7CD3%7Cpg%7C%7Cpc%7CSJ%7Cpc%7CS8%7Cpc%7CS4%7Cpc%7CSQ%7Cpg%7C%7Cpc%7CD6%7Cpc%7CDA%7Cpc%7CD5%7Cpc%7CDQ%7Cpg%7C%7Cpc%7CCA%7Cpc%7CC3%7Cpc%7CC6%7Cpc%7CC9%7Cpg%7C%7Cpc%7CC5%7Cpc%7CC4%7Cpc%7CS5%7Cpc%7CC7%7Cpg%7C%7Cpc%7CDK%7Cpc%7CD9%7Cpc%7CCT%7Cpc%7CD7%7Cpg%7C%7Cpc%7CS2%7Cpc%7CDT%7Cpc%7CST%7Cpc%7CD8%7Cpg%7C%7Cpc%7CH7%7Cpc%7CC8%7Cpc%7CHJ%7Cpc%7CCQ%7Cpg%7C%7Cpc%7CSK%7Cpc%7CCK%7Cpc%7CS9%7Cpc%7CCJ%7Cpg%7C%7C%0A%0A]399|300[/hv]

 

What made GIBBO take finesse for sQ ? I would have thought Law of Vacant places would be intrinsic to Monte Carlo simulations.

 

vrock

[wbartley]

With East known to have 1 fewer heart than West, the simulations should indicate even more strongly playing for the drop. In other words, what I'm saying is that Monte Carlo simulations should be all that is required to make the correct play here. For example, if East did have a singleton spade and a doubleton heart, there are a fairly large number of hands with which pass would not be the logical call over 1♥. But even without that inference, the larger number of non-heart cards in the East hand make it more likely he holds two than that he holds 1.

[smerriman]

But even without that inference, the larger number of non-heart cards in the East hand make it more likely he holds two than that he holds 1.

That doesn't make any sense, since the hearts have already been played.. how many each player originally had is irrelevant. East had to discard something!

 

After the 8 of spades is played, East has 8 vacant spades to West's 7, so the drop is 53.3%. Monte Carlo simulations will therefore prove the drop works out better - in the long run.

 

But when it's close, that may need to be a very long run indeed. If you run 100 tests, the results will come out with the finesse leading over the drop about a quarter of the time. So GIB would get this right 3/4 of the time, but wrong 1/4 of the time. And I suspect much fewer simulations are run - if only 10, GIB would get it right only about 60% of the time.

 

(And when I say get it wrong, I mean make the suboptimal play - basing the results on 100 simulations it would still make the contract ~51.6% of the time, just less than the 53.3% humans would.)

[johnu]

To do a proper Monte Carlo simulation, GIB would have to run enough simulations to have some confidence in the results. If GIB were hooked up to a suit distribution split program, it could input the known cards and not have to do a simulation but that's not the case.

 

I'm sure you have all seen GIB make plays that make no allowance for bad splits when there was a no cost line to make a safety play. That's because GIB didn't run enough simulations to "see" there would be a problem with a bad split.

 

Lack of computer resources is the reason BBO says there are basic bots that use just enough CPU time to keep the game going, and advanced bots that use more resources (but how much more??? and what would the effect be if GIB took twice as much CPU time to run simulations, 3x, 4x, etc.) Should paying customers expect more from advanced bots???

[virgosrock]

To do a proper Monte Carlo simulation, GIB would have to run enough simulations to have some confidence in the results. If GIB were hooked up to a suit distribution split program, it could input the known cards and not have to do a simulation but that's not the case.

 

I'm sure you have all seen GIB make plays that make no allowance for bad splits when there was a no cost line to make a safety play. That's because GIB didn't run enough simulations to "see" there would be a problem with a bad split.

 

Lack of computer resources is the reason BBO says there are basic bots that use just enough CPU time to keep the game going, and advanced bots that use more resources (but how much more??? and what would the effect be if GIB took twice as much CPU time to run simulations, 3x, 4x, etc.) Should paying customers expect more from advanced bots???

 

I keep saying there is a resource problem and "they" or "others" keep pooh-poohing me.

 

The context here is money bridge so i suppose I am a paying customer though just table fee.

 

In my experience GIBBO plays for corner cases almost all the time. Is pure genius in doubled contracts.

 

vrock

[johnu]

I keep saying there is a resource problem and "they" or "others" keep pooh-poohing me.

 

Another example. Have you ever seen GIB in a no problem contract, randomly cash an ace or king in a side suit that did not need to be cashed before drawing trump? If the opponents had 6 cards in the side suit, they would be 6-0 about 1.5%. So in 50 simulations, you would expect to see a 6-0 split a little less than 1 time. By cashing a side suit winner, GIB is confirming that not enough simulations were done for a 6-0 split to show up in the simulations.

[wbartley]

If you just eliminate 7 card minors from the East hand, 1000 samples produces the correct answer very reliably. On my puny desktop it takes about 10 milliseconds to run 1000 samples and compute the statistics.

[virgosrock]

If you just eliminate 7 card minors from the East hand, 1000 samples produces the correct answer very reliably. On my puny desktop it takes about 10 milliseconds to run 1000 samples and compute the statistics.

 

by "correct answer" you mean playing for drop?

 

vrock