[bab9]

tysen2k, on Jun 29 2010, 07:27 PM, said:

Most of this concept comes from Matt Ginsberg, and he and I describe it in several RGB posts several years ago.  I also used a similar concept in my bidding system design contest I held, but I’ve modified it slightly. 

I am unfamiliar with RGB posts. What are they, and where can I find them?

In the system design contest, how did you determine a winner?

[CamHenry]

bab9, on Jun 30 2010, 03:32 AM, said:

tysen2k, on Jun 29 2010, 07:27 PM, said:

Most of this concept comes from Matt Ginsberg, and he and I describe it in several RGB posts several years ago.  I also used a similar concept in my bidding system design contest I held, but I’ve modified it slightly. 

I am unfamiliar with RGB posts. What are they, and where can I find them?

In the system design contest, how did you determine a winner?

RGB is rec.games.bridge; you can find it on various newsservers or on Google Groups. Searching the archives is possible but not trivial on Google.

[hotShot]

Nice work Tysen!

Anybody feel competent to say at what level of event such a system would be allowed?

I guess that, if you add a restriction like: "You need at least 3 HCP to open."
the information contained in "pass" drops a lot and "Silent Spades" would be less attractive.

[helene_t]

Very impressive work, Tyssen!

[hotShot]

bab9, on Jun 30 2010, 09:32 AM, said:

I am unfamiliar with RGB posts.  What are they, and where can I find them?
Try

Try Biddig system design contest and for further reading:

Matt Ginsberg's post on system evaluation

[mgoetze]

tysen2k, on Jun 30 2010, 01:27 AM, said:

An example. Start with a large sample of double dummy hands. Before any bidding begins and you have a random hand, your distribution of perfect contracts looks like this:

6C	0.9%
6D	1.1%
6H	1.5%
6S	1.5%
6N	1.7%
7C	0.2%
7D	0.3%
7H	0.3%
7S	0.5%
7N	0.7% 

You're saying I should be in a grand every 50th hand, and in a slam every 15th?

[helene_t]

Doesn't seem too unreasonable to me. Some of those slams will be sacrifices.

[gwnn]

Tysen I wonder if you could show some responses to opening bids. Maybe you talked about this somewhere but I didn't read properly.

[helene_t]

No, so far he only optimized the opening scheme I think. The responses could then be optimized using the same principles.

[NickRW]

mgoetze, on Jun 30 2010, 12:17 PM, said:

You're saying I should be in a grand every 50th hand, and in a slam every 15th?

That's probably not that far off reality. You should remember that DD results are roughly right around the 4M mark. Above that they slighly favour the declaring side (declarer usually has enough controls that he is in charge and, DD, will always finesse or drop or squeeze or end play correctly). Below the game level, DD analysis slightly favours the defending side (they have enough cards to do some damage and will always get off to the best opening lead etc).

However, over a large number of hands, DD analysis is in the right sort of ball park compared to single dummy results and doesn't have the other problems that single dummy results have - so they're good for this sort of work.

Nick

[NickRW]

tysen2k, on Jun 30 2010, 03:25 AM, said:

NickRW, on Jun 29 2010, 05:54 PM, said:

but what happens if you up the efficiency to 20%

Actually I started off with a 20% efficiency. It had the opposite effect of what you are hoping for. It wanted to keep its opening bids even lower because it thought that the responses would be much better. It didn't want to eat up any bidding space at all and reluctantly bid up through 1S.

Tysen

Yes. I slept on it and woke up anticipating that was exactly what you were going to say, recalling the comments that were made in the earlier work. However, I am still sceptical - very.

1) As a sanity check, how does this system compare with the others that were in the bidding system design contest you held some years back? If this system is truly a good system, surely it should win easily.

2) Even if it does win, there is something about a flat efficiency rate that seems wrong to me. What I mean is that

a] A Benji 2♦, 23+ opener, although you've preempted your side a lot, is really quite unlikely to be interfered with by opps.
b] An Acol 2♣, 23+ opener is only marginally more likely to be overcalled
c] A Standard 2♣, 22+ is, again, just marginally more likely to be overcalled.
d] A forcing 1NT (say) 20+ opener is a bit more likely to be overcalled
e] A forcing 1♦ at about 18+ is a fair bit more likely to be overcalled
f] A forcing 1♣ at about 16+ is more likely to be overcalled still - in particular a cheap 1♠ overcall hits it quite badly
g] A forcing pass at about 14+ or whatever, is really asking to be hit and is hit badly by cheap overcalls of 1♥ and 1♠ - indeed it is no longer strong enough to effectively dictate to opps that their 1NT overcall can't be natural and to play anymore.
h] A pass showing spades and no strength guarantee at all - well - it really has to be at the bottom of the pile in terms of being overcallable - indeed, apart from deterring opps from calling spades naturally, it really has no effect whatsoever on 2nd seat.

The above surely must have some impact on the efficiency that can be expected from different strengths and levels of opening. And in turn this must have a dramatic effect on what bidding system would be recommended.

Nick

[NickRW]

hotShot, on Jun 30 2010, 08:59 AM, said:

Anybody feel competent to say at what level of event such a system would be allowed?

I think the 3hcp is probably either 0-3 hcp or Tysen has just sampled the different exact hcp counts in his results in order to make a small enough graphic to be readable.

Given that pass does not guarantee any strength, it might be playable at EBU level 4 - though I read in another thread that they're bringing back level 5 and level 4 will have 1♥ and 1♠ restricted to natural openings - no transfer openings

Nick

[tysen2k]

I too am still skeptical about the process, but I thought it was cool enough to post anyway.

BTW, can everyone still see the graphic that describes the system? I could see it when I posted it at home, but now that I'm at work it doesn't show up. We do have lots of filters here, so it could easily be blocked for just me. [Edit: Helene, thanks for the confirmation it still shows!]

The HCP indicated are just sampled exact figures, so if there's a transition between two HCP squares, it happens somewhere in between those amounts.

The posted system is for 10% efficiency. Nick, I totally see your point about efficiency, but I'm not sure if there's a fair way to handle it. Let me think about it some more. Actually, maybe trying to generate responses to these openers might be a good validation. I'm worried that there are too many hands that pass and bid 1C and it may not be able to come up with good responses. I could try that next.

Tysen

[tysen2k]

helene_t, on Jun 30 2010, 05:20 AM, said:

Doesn't seem too unreasonable to me. Some of those slams will be sacrifices.

Right, which is actually good because it's another component of competative bidding the model isn't ignoring. If it were to only focus on constructive bidding, it might want to stop in 2S when the opponents can make 3H.

Tysen

[NickRW]

tysen2k, on Jun 30 2010, 04:40 PM, said:

I too am still skeptical about the process, but I thought it was cool enough to post anyway.

Yes I did say it was sexy!

Quote

The posted system is for 10% efficiency.  Nick, I totally see your point about efficiency, but I'm not sure if there's a fair way to handle it.  Let me think about it some more.  Actually, maybe trying to generate responses to these openers might be a good validation.  I'm worried that there are too many hands that pass and bid 1C and it may not be able to come up with good responses.  I could try that next.

Tysen

I too don't know of a "fair" way other than to play about with your code and some different numbers and see what happens by trial and error. The fact that you've chosen a neural net doesn't help a lot either as they can be difficult to decipher what they are doing, never mind why - they just react to the stimulus provided.

It MAY be a little easier if you had a set of masks - so many spades, hearts, diams and hcp = do X and randomly change the Xes until you get improvement. It would be easier to determine the minimums (if any) guaranteed - though a little more hit and miss about finding an optimum set of responses.

If it were my code, I'd want to play with:

Base efficiency assumed to be, say, 10%
Pass = +0%
1♣ = +1%
1♦ = +2%
1♥ = +2.5%
1♠ = +3%
1NT = +3.5%
2 ♣ = +4%
2 ♦ = +4.5%

Minimum hcp guaranteed:
<=10 = +0%
23+ = +5%
Other exact counts pro rata.

Whether the above is any more "fair" (i.e. based on reality) is anyone's guess - but I think it is worth taking a look at something of this nature. Put it this way, if the Silent Spade is the right way to bid, how can system designers have been so wrong all these years? That question really begs some sort of answer - even if it is that humans are genuinely stupid after all.

Nick

[NickRW]

I have another question - how do you determine which hand (or hands) will be the opening hand on any given board in the DD database?

You see, rather than pester you to try this that and the other, I was thinking of doing my own analysis... You can, potentially, take all 4 hands on any given board and analyse what it should open (or not) (as if it were the dealer - even if it wasn't). But on some boards, one side has no makeable contract and no sacrifice. On other boards, one side has a makeable contract, but opps are going to (constructively) outbid you anyway and there is no worthwhile sacrifice. Such (pairs of) hands are fairly common within any set of random boards, but including them in the analysis could skew things in favour of an opening pass.

Also, though your side may have no (or a low level) makeable contract, and no sacrifice, an opening preempt could have a fairly huge effect on their auction - but if you're just concentrating on makeable contracts and sacrifices where they should be bid, then (low level) preempts aren't going to figure too much in the equation - are they not?

Just musing here on potential methods and their implications.... Especially as the silent spade is not recommending any preempts and not recommending any forcing from strength opening.

Nick

[bab9]

tysen2k, on Jun 29 2010, 04:38 PM, said:

I made a bit of progress into having a system that's completely computer designed.  I designed it by using some information theory concepts and used a neural network to define bids that minimize the entropy of perfect contract distributions.  I know that's kind of a high-level description... I can go into more detail if desired.

What size was the training data set?

What type of neural network did you use? How many nodes?

[tysen2k]

bab9, on Jul 2 2010, 06:01 AM, said:

What size was the training data set?

What type of neural network did you use?  How many nodes?

I'm running off of 20k training cases. I've got 20 nodes that are being fed into a softmax function to determine the opening bid. I've tried increasing the number of nodes and it doesn't seem to make much difference.

I've recently started trying to fully incorporate competition/preemptive bidding. I'm now also looking into what the opponent's best contract is in addition to our own. So I'm now trying to maximize the chance that we find our best contract minus the chance they find theirs. I let it run overnight...

And I'm still getting results similar to when I had no competition. It never wants to bid higher than 1N and the bid definitions are somewhat similar to the original Silent Spade, though they use more HCP info.

Pass is something like 4+ spades (unbalanced only) or 15+ HCP and both minors
1C is 0-12 HCP, no singletons/voids, 2-4 spades, 2-5 hearts
1D is clubs
1H is diamonds
1S is hearts
1N is 0-10 HCP, semi balanced, usually has diamonds and only 2-3 in both majors

It's really strange... I'm going to look at it some more and see if I can find out why it's not preempting at all (except that 1N bid is a preempt). And I don't know why it likes "1 over" transfers either.

Tysen

[hrothgar]

tysen2k, on Jul 2 2010, 07:48 PM, said:

It's really strange...  I'm going to look at it some more and see if I can find out why it's not preempting at all (except that 1N bid is a preempt).  And I don't know why it likes "1 over" transfers either. 

Any chance that you can tweak the starting conditions?

Seed the initial population with some standard methods and see whether the learner evolves away from what is (presumably) a local minima

[NickRW]

tysen2k, on Jul 2 2010, 04:48 PM, said:

It's really strange...

Looks to me like it saying there are more spade contracts in the database than anything else, therefore when I have spades I want to bid cautiously - hence the spade pass - then it is chucking, essentially, the fert hands into 1C and, then, 1 over transfers are sort of the logical thing to do with the remaining suits/bids.

Are you looking at only those (pairs of) hands where that side should declare? Maybe not of course, but that would skew things....

Nick