yet another all time greates declarer play

[2200]

Hi guys!

I have experienced numerous all time greatest play by GIB, but this one I cannot take it anymore.

The 99th time I checked, it is still a claimer at trick 1. Just knock out the ♣A.

What did GIB do?

It cashed out diamonds first, creating a length winner for me, then started to knock out the ♣A.

 

[hv=https://www.bridgebase.com/tools/handviewer.html?bbo=y&lin=pn%7C2200%2CRobot%2CRobot%2CRobot%7Cst%7C%7Cmd%7C3S842HJ7D96532CA43%2CSQJT9HQ93DAKQ8C86%2CS765HT852DT74CT92%2CSAK3HAK64DJCKQJ75%7Csv%7CB%7Cah%7CBoard%2013%7Cmb%7CP%7Cmb%7C1C%7Can%7CMinor%20suit%20opening%20--%203%2B%20%21C%3B%2011-21%20HCP%3B%2012-22%20total%20points%7Cmb%7CP%7Cmb%7C1S%7Can%7COne%20over%20one%20--%204%2B%20%21S%3B%206%2B%20total%20points%7Cmb%7CP%7Cmb%7C2H%7Can%7COpener%20reverse%20--%205%2B%20%21C%3B%204%2B%20%21H%3B%203-%20%21S%3B%2021-%20HCP%3B%2018-22%20total%20points%7Cmb%7CP%7Cmb%7C4N%7Can%7CQuantitative%20invite%20to%206NT%20--%2012-14%20HCP%7Cmb%7CP%7Cmb%7C6N%7Can%7C5%2B%20%21C%3B%204%2B%20%21H%3B%203-%20%21S%3B%2021%20HCP%3B%2022-%20total%20points%7Cmb%7CP%7Cmb%7CP%7Cmb%7CP%7Cpc%7CD4%7Cpc%7CDJ%7Cpc%7CD2%7Cpc%7CD8%7Cpc%7CH6%7Cpc%7CH7%7Cpc%7CHQ%7Cpc%7CH8%7Cpc%7CH9%7Cpc%7CH2%7Cpc%7CHA%7Cpc%7CHJ%7Cpc%7CS3%7Cpc%7CS2%7Cpc%7CST%7Cpc%7CS5%7Cpc%7CDA%7Cpc%7CD7%7Cpc%7CC5%7Cpc%7CD3%7Cpc%7CDK%7Cpc%7CDT%7Cpc%7CC7%7Cpc%7CD5%7Cpc%7CDQ%7Cpc%7CC2%7Cpc%7CH4%7Cpc%7CD6%7Cpc%7CC8%7Cpc%7CCT%7Cpc%7CCK%7Cpc%7CCA%7Cpc%7CD9%7Cpc%7CC6%7Cpc%7CS7%7Cpc%7CCJ%7Cmc%7C11%7C ]400|300|[/hv]

[smerriman]

Was this basic GIB or advanced GIB? Makes total sense if it were basic GIB due to the simplistic way it works, and you get what you pay for, but shocking if it were the paid robot which has proper declarer play.

[nullve]

Me (imitating 2200 as far as possible) and three basic robots:

 

[hv=https://www.bridgebase.com/tools/handviewer.html?lin=st||pn|nullve,~~M8031b0j,~~M117001d,~~M11726of|md|3S842HJ7D96532CA43,SQJT9HQ93DAKQ8C86,S765HT852DT74CT92,SAK3HAK64DJCKQJ75|sv|b|rh||ah|Board%2013|mb|P|mb|1C|an|Minor%20suit%20opening%20--%203+%20!C;%2011-21%20HCP;%2012-22%20total%20points%20|mb|P|mb|1S|an|One%20over%20one%20--%204+%20!S;%206+%20total%20points%20|mb|P|mb|2H|an|Opener%20reverse%20--%205+%20!C;%204+%20!H;%203-%20!S;%2021-%20HCP;%2018-22%20total%20points%20|mb|P|mb|4N|an|Quantitative%20invite%20to%206NT%20--%2012-14%20HCP%20|mb|P|mb|6N|an|5+%20!C;%204+%20!H;%203-%20!S;%2021%20HCP;%2022-%20total%20points%20|mb|P|mb|P|mb|P|pc|D4|pc|DJ|pc|D2|pc|D8|pc|H6|pc|H7|pc|HQ|pc|H8|pc|H3|pc|H2|pc|HA|pc|HJ|pc|S3|pc|S2|pc|SQ|pc|S5|pc|DA|pc|D7|pc|SK|pc|D3|pc|DK|pc|DT|pc|SA|pc|D5|pc|SJ|pc|S6|pc|C7|pc|S4|pc|S9|pc|S7|pc|C5|pc|S8|pc|ST|pc|C2|pc|H4|pc|C3|pc|C8|pc|C9|pc|CK|pc|CA|pc|D6|pc|DQ|pc|H5|pc|CJ|pc|H9|pc|HT|pc|HK|pc|C4|pc|CQ|pc|D9|pc|C6|pc|CT|]400|300[/hv]

[smerriman]

Actually, this is more interesting than I thought, and it looks like it does affect advanced GIB too.

 

At the point GIB has to choose a discard on the king of diamonds, there is still a 100% single dummy line (the one nullve showed, discarding both spade honors). So even though there were easier routes, it's fine for even advanced GIB to get to this position.

 

If I recreate that situation multiple times with the older version that I'm using to test - with the single dummy GIBson engine enabled - it's discarding the K about half the time and making the contract.. but throwing the club the other half and going down.

 

But it's not because the sample is too small and it believes throwing the club will always work - it understand that line of makes if the diamonds are 4-4 or the club ace is with the short diamonds, so that logic is right.

 

It's doing so because it doesn't believe throwing the king is a 100% line - when it lists out holdings that its chosen line works for, there is always one where it believes it will get 11 tricks. So when those 'failing' cases outnumber the above failing cases, it chooses the club.

 

Every time I re-run it it's coming up with different holdings where it believes it can go down.. why it can never seem to find the right line I have no idea; you'd think it'd come up with it once in a while. (Actually, eventually on about the 20th attempt it came up with SK being 100%).

 

So easy to replicate too, so would be straightforward to debug :( Not that that would make it trivial to fix; the single dummy algorithm described in Matt Ginsberg's paper takes a fair bit of understanding, and I could understand it getting stuck in local minimums etc. But still, definitely improveable.