Re: AI for Bidding in Bridge

btlog wrote:
> "Randolph M. Jones" <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>...
>
> > btlog wrote:
> > I'm not up on a lot of research in this area, but the main work I'm 
> > aware of is Ginsberg's Intelligent Bridgeplayer:
> >    http://www.cirl.uoregon.edu/ginsberg/gibresearch.html
> >
> > Are you really talking about *learning* a bidding system, though?  What 
> > kind of learning do you mean?  Certainly human bridge players learn 
> > bidding systems by instruction (and then perhaps tuning their memories 
> > of the instructions), but I'd think it would be rare for a bridge player 
> > to try to induce bidding rules just by observing bidding behavior.
> >
> > I'm pretty sure Bridge Baron just has a set of large flow-charts 
> > (basically a rule-based system) for bidding.  GIB seems to have 
> > something similar but also apparently adds a lookahead search of some 
> > sort.  Sorry but that's the extent of my knowledge about it.
>
>
> Thanks for your reply
>
> I am aware of Ginsbergs GIB program and some of the focus of the
> research that has been involved in the application.
>
> At the moment I am still keen on the system inducing the bidding rules
> with minor tweaking by instruction. I am hoping that this method will
> allow the system to be more flexible in its bidding responses.
>
> I would like to implement a fusion of the range of bidding methods
> that have been implemented at this stage. I will put a disclaimer that
> I am only starting the planning for my project proposal and I may find
> some interesting research that may point me in another direction.
>
> It is both disappointing and encouraging that there is not a lot of
> reasearch (that I can find published) that is in this area.
>
> My long term goal is to provide a more flexible bidding engine to
> commercial applications. I would like to couple this with a top of the
> line play engine.
>
> Here is hoping
>
> Regards
>
> Craig

I'll just add a couple of cents worth of advice about inducing bridge 
rules (and remember you're getting what you pay for here :-)).  Maybe 
you've already thought about these things, so you can just ignore me if 
so :-).  In my experience, two of the hardest parts of getting a system 
to learn anything interesting are getting the blame/credit assignment 
right, and getting the feature-language right.  So first of all, make 
sure that your system comes with a feature language to describe 
situations that already includes every potential concept that might need 
to be used in the ultimate induced bidding rules.  So you'll want the 
system to know about point-counts, sure-tricks, what signals mean, 
vulnerability, and *everything* else that influences the right bid. 
It's okay to have too many features, because any reasonalbe learning 
algorithm will learn to ignore the irrelevant ones.  But if you are 
missing any important features, there just aren't good learning 
algorithms out there that will discover them.

Second, I personally don't think you're going to get anywhere if you 
have your system just passively look at the hands and watch the bidding 
(although maybe that's not what you had in mind).  It should be a fairly 
spoon-fed process, where each time you give the system an example of a 
bid, you give it clear information about why that particular bid is the 
right one at that time (using the feature language you built above). 
Then you can leave it to the learning system to create general rules 
from the specific examples.  I think there's a lot of promise to that 
approach, but if you're thinking of any more ambitious kind of rule 
induction, I imagine you'll run into a lot of problems (which everybody 
who has tried this kind of thing has run into).

Anyway, good luck...sound like a fun (and potentially lucrative) project.

Randy