Leaderboard

I think this is not correct. Showing hand type is a means to an end, the end being reaching the optimal contract. It is not clear what it means for a hand to belong to a type (the standard take is that this should just mean distribution, i.e. the four suit lengths, along with a general description of strength. But why not specific stoppers, situational stoppers, conditional strength descriptions and other?). More importantly, more space should be assigned to situations where it is unclear what our likely best contract is, while we are relatively safe in bidding higher when we have a clearer picture of possible best contracts. In particular, each subsequent bid should not cut hand types in two, but cut the uncertainty in final contracts in two, i.e. halve the entropy of the conditional distribution of best final contracts. Hands that want to bid to a specific contract but are too weak to bid any higher are among the least complex ones and provide the greatest certainty. As a result these are among the best candidates for jump bids. As a rule of thumb, any bid that suggests itself as a final contract eliminates most possible bidding sequences that go through that bid, and significantly reduces the amount of information that can be communicated by the system. The rest looks really good. I like the 'width' of a cluster as the average IMP cost if partner cannot tell individual hands in a cluster apart. If only I had read this first, I could have spared us both my comment.

It's been a while since we have discussed AI-generated bidding systems. I am currently toying with a fairly simple idea: Assume that opps aren't going to interfere (ok, we will relax that assumption at some point but walk before you run blah blah. Also, for optimizing sequences that take place after both opps have passed it may be a reasonable simplification). We know that the optimal relay system allocates F(i) hand types to the i'th most expensive bid where F is the Fibonacci sequence. But then you have to think about relay breaks and captaincy appointment so I prefer to optimize the system for 2-way information exchange. So the proportion of hand types allocated to the bids are 1/2, 1/4, 1/8 etc. This is nice because we can then just organize the hand types in a balanced binary tree using hierarchical clustering. So in 1st seat, for example, it could be Pass = Left daughter 1♣ = Right daughter of right daughter 1♦ = Right daughter of left daughter of right daughter etc. We have some decisions to make: - When to branch left and when to branch right? In a non-GF situation maybe we should allocate bids so as to allow partner to pass as often as possible, especially when we are in danger of coming too high. But this is difficult to formalize in a way that can be implemented efficiently so I thought that a simple approximation would be to go low with the subtree that minimizes the expectation of PAR^p, where p might be 0.5 or thereabouts - Criteria for passing In general we should pass on some hands where the last bid may not be the optimal contract (when in a hole, stop digging). I am not sure how to formalize it. For now, I think I will just tell it to pass whenever the expected loss relative to PAR is less than say 1 IMP. Maybe the expectation of LostIMP^q where q is 2 or thereabouts? - Distance metric for clustering Here I will use the same metric that I used in my previous attempts (where I used an induction algorithm, i.e. top-down rather than the bottom-up approach of hierarchical clustering): the distance between two hands is the expected IMP loss if partner has to place the contract without being able to tell the two hands apart. - Aggregation of distances to produce subtree-distances There's mean and median and minimax etc. I always just use whatever the software gives me by default, but if there's someone here who is into machine learning maybe they can suggest something more intelligent? Extending the "partner-can't-tell-them-apart" definition to subtrees would be ideal but probably too computationally demanding. - Presentation of the system I have generated 1800 clusters of hands a priori which means that in 1st seat, pass will be defined as (either of 900 hand types), 1♣ as (either of 450 hand types) etc. It would be nice to try to find natural-language summaries of those hand type clusters, such as "unbalanced without a 6-card major, 10-16 points" or whatever. Anybody knows of a package (preferably for R or C++) that can do something like that? - Benchmarking Csaba has a script that can bid and play hands using GIB so that would be a reasonable comparator. Any thoughts?

[hv=https://www.bridgebase.com/tools/handviewer.html?lin=st||pn|helene_t,~~v3fakebot,~~v3fakebot,~~v3fakebot|md|1S6HKJTD8642CAKQJ8,SQJ432HAQ9DT75C62,S8H843DAKQ9C97543,SAKT975H7652DJ3CT|sv|b|rh||ah|Board%207|mb|1C|an|4+!C;%20HCP%2011-23;%20natural|mb|P|mb|1D|an|1+!D;%20HCP%204+;%20natural%20or%20fit%20in%20!C%20-%20Forcing|mb|1S|an|5+!S;%20HCP%208-17;%20natural|mb|2D|mb|4S|an|4+!S;%20HCP%203-15;%20fit|mb|D|an|1+!D;%20HCP%204+;%20penalty|mb|P|mb|5C|an|6+!C;%204-5!D;%20HCP%2011-16;%20natural|mb|P|mb|P|mb|P|pc|SQ|pc|S8|pc|S9|pc|S6|pc|HA|pc|H3|pc|H2|pc|HT|pc|S3|pc|H4|pc|SK|pc|C8|pc|CA|pc|C2|pc|C3|pc|CT|pc|CK|pc|C6|pc|C4|pc|S5|pc|D2|pc|D5|pc|DA|pc|D3|pc|DK|pc|DJ|pc|D4|pc|D7|mc|11|]400|300| Just to make it easier to see...[/hv]

If nothing ever went wrong games wouldn't need Directors.

I'd expect partner to pull to 5♠ with that hand. Our defensive values are limited by a failure to double 2♥, and partner knows we are hoping to take a spade trick (or perhaps even two) and we will not.

We have a defensive hand and partner didn't bid 4♥ to show a good raise last round. I think double is fine, we likely won't make 5♠ unless partner has very good support.

3♠ shows approximately this, would like a little bit more but wasn't dealt it

[hv=handviewer.html?nn=Argine&n=SQHA95DAT6532CQ96&d=w&a=p1d(4+!d; HCP 11-23; 4-card !d opener)P1S(4+!S; HCP 4+; natural - Forcing)P2D(5+!D; HCP 11-16; unbalanced)P3H(4+!D; 4+!S; HCP 13+; Splinter, shortness in this suit and big fit in last suit bid - Game forcing)P4H(5+!D; 3-4!H; HCP 11-15; to play)]450|200[/hv] Robots just can't get splinters right, can they? Admittedly, this may be yet another case of the logic not matching up with the description, given 4♥ seems a reasonable bid otherwise.

In automated tourneys without director you can't table chat but you should be able to chat to opps. For example, I am not sure how else to tell them what kind of leads and carding we play.

Is that true? Actually you are probably right, I don't recall ever having seen an alert. Maybe that explains why we do so well playing Moscito. Opps think we play 2/1. Just kidding. But it's fairly ridiculous. When you join the tournament you are encouraged to alert your bids.

Go into your account under convention cards and pick you + partner's card. Then verify. The alert function in red should be used to self=alert- let someone know if it is not working.

also, if you alert a bid it is NOT visible to opponents

Guess what i discovered When you play with a regular partner your CC is still set to GIB 2/1 so not only can you not tell opponents you are playing weak NTs or your carding they cant even check your CC