[goffster]

There are many obvious places where running simulations are bad. GIB could improve its bidding immensely if
it simply followed the most obvious of bidding rules. Why does it insist on running simulations for
cut and dried situations?

[smerriman]

goffster, on 2016-July-06, 17:05, said:

There are many obvious places where running simulations are bad.

Just for argument's sake, can you name one? While the results of a simulation-based bid can look odd / plain wrong to a human, its bid is made for a good reason. The only way it can know if a situation is clearcut is to simulate; if its simulation says a bid is clearcut then it makes it. If it is making a wrong assumption in the simulation, that's what needs resolving, not the actual logic behind the simulation itself.

[goffster]

I do not have hard examples. They almost always revolve around bidding its hand twice, rebidding 5 card suits, when no action was required. Usually robot fails to realize that the consequences of being wrong include being doubled.

[1eyedjack]

Point is, GIB DOES have a set of rules, and it DOES run simulations only where those rules run out. Perhaps there are not ENOUGH situations defined by rules, and perhaps some of the rules as programmed are plain wrong, but maybe that is a work in progress with an undefined conclusion.

[goffster]

If it were to instead rely on a more expansive set of rules, its bidding would improve.

[Stephen Tu]

It doesn't do any good to make this complaint without specific examples. AFAIK GIB marks certain rules as "you should simulate here", and others as should make book bids. Which is as it should be IMO, in general it should tend to stick to book bids early in the auction and tend to simulate when making calls that will tend to be final decisions.

Post specific auctions where GIB clearly #%$^*^'d up, then that reveals auctions where rules need to be added, or existing rules reprioritized, or existing bid definitions adjusted. The number of rules needed to bid effectively is just enormous especially when competition has to be taken into account. The problem is that just in lots of auctions GIB falls into somewhere not explicitly defined and falls back on default rules which don't work too well in the situation in question.

[goffster]

I'll try to tabulate some situations over a few hands.

I am talking about clear cut situations where the right bid is clearly defined.

I think any user of GIB is 100% familiar with what I am talking about.

[mycroft]

If you look through BBF history (I realize you would have to be incredibly lucky to find it, but it's there), you'd know that there are already more rules than any reasonable person would believe, and that they do increase all the time, and that they are incredibly complicated and difficult to keep from inconsistencies already. It's surprisingly hard to codify even a "simple" bidding system. It's pretty amazing how much happens in humans without "thinking".

Yes, please point out areas where things are obviously wrong - others do it all the time. I agree, any user of GIB knows what you're talking about - there are random places where the world caves in for no apparent reason. We don't know where *you* are falling down holes, though. "Fix it, it's broken, and it's obvious where it's broken" is the requirement that dooms all projects it's used in; so I'm pleased that it's not in the "GIB improvement project".

[Stefan_O]

goffster, on 2016-July-06, 17:05, said:

There are many obvious places where running simulations are bad.

How can you even tell, whether a robot-bid was based on simulation or not?

[Bbradley62]

Stefan_O, on 2016-July-09, 04:36, said:

How can you even tell, whether a robot-bid was based on simulation or not?

First, if GIB does something clearly bad, it should certainly be fixed if it was not based on simulations, so reporting it is helpful either way.

Second, if you report something, one of the first things that is likely to happen is that Lycier will rerun the hand using basic GIB, to see what happens without simulations. This should save the programmers a step in their analysis of the problem.

[goffster]

Ok, here was one that just played out.

http://tinyurl.com/hde97rs


Once robot bids 2D, it is done. Grandma would be done. Expert would be done.
There is nothing more to its hand than it has advertised in the bidding.

Yet, it chose to hang its partner GIB with the bid of 3H later in the bidding.

Perhaps a useful bidding tool for gib to have would be:

"Does my hand differ substantially from what I advertised?"
"Is partner asking me to bid more?"

No?
then "I am done, I am not going to run any simulations"

[1eyedjack]

It may be that when both robots in a partnership are bidding based on simulations, there is potential for both to deviate in the same direction with catastrophic consequences. ie if both decide marginally to overbid, you end up with a gross overbid.

West, on this hand, trusted partner to have the 15+ TP promised for the 3D bid.

That said, I neither think that 3D *should* show 15+ TP, nor that any simulation run by West should indicate bidding 3H, whatever reasonable definition you assign to 3D.

If West was bidding on sims, I suggest that the problem has more to do with the sim algorithm than the fact that a sim algorithm was used,

[johnu]

1eyedjack, on 2016-July-10, 22:21, said:

West, on this hand, trusted partner to have the 15+ TP promised for the 3D bid.

West should have trusted partner not to have 15+ after rebidding 1NT which showed 12-14 and then passing the 2♦ bid. I guess GIB forgot the previous 2 rounds of bidding.

[1eyedjack]

johnu, on 2016-July-11, 00:38, said:

West should have trusted partner not to have 15+ after rebidding 1NT which showed 12-14 and then passing the 2♦ bid. I guess GIB forgot the previous 2 rounds of bidding.

The 1NT rebid promised 12-14 HCP. I don't think that it made any promises regarding TP, although being balanced or semi-balanced probably places some automatic constraints. It is certainly possible to have 15 TP consistent with the earlier bidding, although more than that would be a stretch so the "+" is probably surplus to requirements.

Whether TP has any relevance to investigating 3N, the most likely game, is another issue, and one regularly aired.

[goffster]

Here was another disaster, just played out:

http://tinyurl.com/hn3oa33

Again after bidding 2N (fine)

It later bids 3S (not fine)

Again the logic:
"Does my hand differ substantially from what I advertised?"
"Is partner asking me to bid more?"

No?
then "I am done, I am not going to run any simulations"

[goffster]

You can argue about the logic all you want but if robot were able to say:

"Does my hand differ substantially from what I advertised?"
"Is partner asking me to bid more?"

No?
then "I am done, I am not going to run any simulations"


Then its bidding would improve immensely.

[Stephen Tu]

You really should just post each of these hands in its own thread. All of these are basically something like:
- need a new rule on this auction
- need new point ranges for this bid (or these bids) on this auction.

You have to remember that computers don't think the same as humans. You can't merely tell it "don't bid on if you don't have extras for your previous auction", "don't bid on if you already limited your hand and partner just bid an extra level competing for the partial, didn't make game try" , and then automatically the computer figures everything out and is magically better on all sequences. It is a dumb pattern matcher, it basically looks into database, and says "do I have a rule that matches my hand and this auction? yes, so I bid it". It has no concept of "do I have more than I previously shown?". You have to really explicitly give it thousands upon thousands of rules so that it emulates human behavior, you can't simply give it a common sense notion in natural language and then it "gets it". GIB has some basic rules basically making it have a bias towards "keep going if game is possible", it needs rules on these auctions tweaked so that east showing less just for competing, and west's non-passes showing more so that it can't choose them.

Your last one basically they need to lower the minimum HCP for the doubler's actions (so west knows east can be bidding on shape, less tempted to bid more), and increase the minimum to bid 3s after doubler took out the natural 2nt (basically shouldn't exist given the NF 2nt the previous round).

Even better if 2nt was two places to play IMO, but that's another discussion.

[goffster]

The logic posted can be translated into code.

A hand could be statistically measured as being a certain distance away from the average hand
rather easily using a hand dealer.

[Stephen Tu]

That won't matter if its view of what East holds is wrong from the rule database. It thinks East has 11 HCP minimum so it tends to deal hands where it has e.g. the Q of hearts in addition so 3nt makes so it feels passing 3c is too conservative.

It sounds to me like you are another non-computer programmer who thinks it's easy to teach a computer the "box principle", underestimates how much tonnage of code it takes to mimic the human idea of deciding whether you are min or max for your previous bidding and acting accordingly, and being able to know automatically whether partner's bid is signoff or not. An idea that can be explained to a human in a short paragraph can take many man years of effort to translate into code a computer can follow.

There are just so many bidding sequences and computers just have no way to intuit what bids "should" mean without practically having to spell out each and every sequence to them. How would west know east is signing off unless rules database specifically marked with that information? Why should it be a signoff if west can have 14 hcp for 2nt and east supposed to have 11? We as humans can intuit the point rules as nonsensical and decide the database is off and adjust and do the right thing and override the rules telling it to keep on going if game could make. Computers have to be led by the nose to that conclusion.

[1eyedjack]

I also am not a programmer, but as I see it there are approaches that might be taken:

1) Program in a logical test that is independent of the preceding sequence, whereby GIB can determine whether it is forced to pass, force to bid (and to what level) or something in between with varying degrees of encouragement. Sounds and probably is horrendously difficult and any solution would be error prone on occasions. Would probably need to build in the flexibility for future refinement of the algorithm.

2) Build a database of sequences in which the programmer defines for each sequence the above parameters, and GIB just interrogates that database. Not so difficult to implement in principle but a monumental task to populate the database, which would only gradually happen over a long period. In the meantime GIB would encounter two situations: Either the sequence is covered or it is not, and if not, do what it has always done in the past.

3) Some combination of (1) and (2), with (2) taking priority, and (1) taking over if (2) returns "undefined".

Then, "all" that would be required is that this routine is triggered whenever GIB is required to call.

I observe that there are some sequences coded into GIB at present which it absolutely takes as forcing, including to a specified level, so it would appear that some attempt at approach 2 is already underway.