[lycier]

---

Result :4♥W-2
My Gib is basic robot,but its evaluation is same with advanced robot. Here I will discuss its evaluation issue.

Gib CC says :
HCP vs Total Points
Gib uses both old fashioned HCP (A=4, K=3, Q=2, J=1)) and "Total points" (HCP+3 for void, 2 for singleton, 1 for doubleton, but short suits containing an honor are reduced by 1 point). It will usually force to game if it thinks it has 25 Total Points between the two hands.

For this hand,3♥ says 14+ TPs.
West Gib actually has 12 hcp,plus 3 distributional points ,15 TPs in total. This is evaluations approach of Total Points.
Really? Why does West Gib ignore worse singleton fit in ♠?

Assume that West Gib holds :
♠ void
♥KQJ42
♦932
♣KQJ63

According to Gib CC, its TPs is samely 15 TPs, obviously misrepresents its evaluation, in fact in this situation it should devaluate instead of value-added due to their ♠ misfit. So eventually such evaluation approach could easily lead to wrong results. It shows its over values when thee are no good suit fits .
I think such counting is one-sided and wrong, the premise of the evaluation application should be suit fit, in other words,only on the premise of suit fit , such evaluating count is reasonable. It just like dummy distributional points counting.
Any ideas?

[Bbradley62]

Yeah, GIB's evaluation method is what it is, and there have been many similar comments, to no avail. (Did East also take a shortness distribution point for the doubleton heart he was raising with?)

I also wonder why West thinks his heart suit qualifies as twice rebiddable, which the documentation says means 7 of them or 6 good ones.

[gwnn]

Yes,yes,you are correct.
What you said just showed a weakness of Gib TPs evaluation model.
What else?

[fromageGB]

Bbradley62, on 2015-December-30, 20:44, said:

(Did East also take a shortness distribution point for the doubleton heart he was raising with?)

I love this method. You have 4 card support - raise him. Similar hand but void in his suit - you are that much stronger, so raise to game.

[cloa513]

lycier, on 2015-December-30, 07:11, said:

---

Result :4♥W-2
My Gib is basic robot,but its evaluation is same with advanced robot. Here I will discuss its evaluation issue.

Gib CC says :
HCP vs Total Points
Gib uses both old fashioned HCP (A=4, K=3, Q=2, J=1)) and "Total points" (HCP+3 for void, 2 for singleton, 1 for doubleton, but short suits containing an honor are reduced by 1 point). It will usually force to game if it thinks it has 25 Total Points between the two hands.

For this hand,3♥ says 14+ TPs.
West Gib actually has 12 hcp,plus 3 distributional points ,15 TPs in total. This is evaluations approach of Total Points.
Really? Why does West Gib ignore worse singleton fit in ♠?

Assume that West Gib holds :
♠ void
♥KQJ42
♦932
♣KQJ63

According to Gib CC, its TPs is samely 15 TPs, obviously misrepresents its evaluation, in fact in this situation it should devaluate instead of value-added due to their ♠ misfit. So eventually such evaluation approach could easily lead to wrong results. It shows its over values when thee are no good suit fits .
I think such counting is one-sided and wrong, the premise of the evaluation application should be suit fit, in other words,only on the premise of suit fit , such evaluating count is reasonable. It just like dummy distributional points counting.
Any ideas?

That's a problem but in this case the problem is that they think 14TP(+) over an overcall is enough to virtually force to game especially with its definition of 9+TP for an overcall. Its bidding 3H with a non-rebiddable it suit- it needs 6 hearts with KQJ.

[lycier]

Bbradley62, on 2015-December-30, 20:44, said:

Yeah, GIB's evaluation method is what it is, and there have been many similar comments, to no avail. (Did East also take a shortness distribution point for the doubleton heart he was raising with?)

I also wonder why West thinks his heart suit qualifies as twice rebiddable, which the documentation says means 7 of them or 6 good ones.

The description of suits length counting have been worst part of Gib work,very very difficult to be improved.

[lycier]

To be honest, I really care about Stephen Tu's points of view at present, hoping he would tell us, including whether TPs evaluation need to be improved,etc.

[Stephen Tu]

Certainly GIB's evaluation engine ought to be improved, it's a long standing complaint but we haven't seen any movement to address this. I think it mainly needs to be more aggressive with good long suits (the shortness only algorithm not being able to fully capture the utility of 6/7+ suits IMO), and somehow be able to be smarter about devaluing honors in the opp's suit, and upgrading honors in partner's suits, upgrading/downgrading for fits/misfits which it doesn't really do. It's harder for them to change evaluation engine than simply define and shuffle around bidding rules which is easier.

For this problem, I mainly have issue with 14+ total points being considered enough to make a forcing 3♥ call. When overcaller can be 8+ I think this is nuts, the floor for 3♥ should be at least like 16+. 1♠-2♥- all pass seems like a way more reasonable way to bid this.

[lycier]

Stephen Tu, on 2016-January-02, 15:04, said:

When overcaller can be 8+ I think this is nuts......

It seems a little doubt for me, I think overcaller should promise 9+hcp in the most situations unless plays mp which overcaller might hold 8hcp,in another word, a qualified overcall in the most situations keep 9 hcp as possible as instead of 8hcp as a standard bidding for Gib system.

Thank you very much for your good reply.

[lycier]

Hi Stephen Tu :


May I ask you a question?
Of course,it is about evaluation counting, as we know that many experts only add length points,don't add short-suit points,for example Audrey and Eric Rodwell,Max Hardy etc,there are clearly documented about it on their books.
On the contrary,Gib adds distributional points in the own hand as evaluation counting,experts add destributional points only for dummy with fits.This is a popular method. How would you think of it?

[Stephen Tu]

Most of the experts counting length points also add short suit points after finding suit fit.

GIB on the other hand counts short suit points instead, and only at the beginning.

As a starting point for valuation, it really doesn't matter all that much what you use. That's because long suits are correlated with shortness, if you are long somewhere that makes you short somewhere else. It's usually within a point of each other, e.g. if you are going to count 5431 as +1 length points, if you count +2 singleton shortness points instead, that one point difference is really very rarely going to lead you to choose one bid vs. another. And when it does make a difference, these are always borderline hands where it's pretty random whether over/underbidding is going to work out well or badly at the end. After all partner has a range too, and sometimes he "has you covered" (he has a bit extra which cover your slight overbid). The problems tend to come when both hands underbid, or both hands overbid.

The main problem with GIB valuation is the total point thing is static AFAIK, rather than dynamically updating as the auction progresses. Most good players are reevaluating the hand at every turn to bid, based on opp's and partner's bidding, devaluing quacks in the opp's suits especially under the bidder, upgrading positional tenaces over an opponent's suit, upgrading quacks in partner's suits, upgrading for good fits found, downgrading for misfits/duplication of values. GIB doesn't do that reevaluation in the middle of the auction. In some cases esp. with the advanced bots it can recover to an extent via simulation on a "final bid" of the auction, when it is deciding to bid game or not, compete one more or not, etc., the simulation can in effect do a reevaluation. The problem is that sometimes it is partner making the final bid, and your bid had to fit a rule, and reevaluation was supposed to push your border hand into a different evaluation bucket. E.g. inv raise/cue instead of normal raise, or the other way around.

Also long suits may not be counted quite enough, and it probably should be counting A/K a little stronger vs. Qs/Js, and hopefully counting some for Ts/9s.

But really there are so many buggy / not well defined rules that its overall bidding performance is affected a lot more by perpetrating ridiculous auctions than these fine evaluation judgment matters, and they are right to work mainly on those before tackling the eval portion. The problem is just there are just innumerable possible bidding auctions, and it is very hard to write rule set to cover them all; computer cannot construct "common sense" rule from similar situation anywhere near as well as humans. When the auction hits gap in the rule set, uncommon competitive auction especially, it often does something kind of crazy.

[lycier]

Hi Stephen Tu :

I am glad to listen to your opinion.
In fact there are many of issues on Gib evaluation which are very difficult to be overcome at present,I suggest if you would have ability to establish a basic rule,no doubt,Gib will be great improved at once,that is to let Gib know a principle of " when the hand is a misfit, count HCP only.And TPs for suit contract,hcp for notrumph contract.", most of evaluation issues will be solved,however if you wouldn't establish this rule, that only shows TPs isn't a mature application of valuation system,so I think this is the crux of the problem, am I correct?

Here I have a curious question on Gib. I remembered that Deep Blue (America IBM company produced) had ever beated world first master of international chess,as we know there are two sevices for Gibs, one for bidding,one for playing,but for Deep Blue,there are 32 brain (microprocessor),so assume Gibs use Deep Blue to play against experts only with same Gib system, can Gibs samely beat human? With high respect.

Lycier

This post has been edited by lycier: 2016-January-04, 17:56

[Stephen Tu]

lycier, on 2016-January-03, 07:01, said:

In fact there are many of issues on Gib evaluation which are very difficult to be overcome at present,I suggest if you would have ability to establish a basic rule,no doubt,Gib will be great improved at once,that is to let Gib know a principle of " when the hand is a misfit, count HCP only.And TPs for suit contract,hcp for notrumph contract.", most of evaluation issues will be solved,however if you wouldn't establish this rule, that only shows TPs isn't a mature application of valuation system,so I think this is the crux of the problem, am I correct?

The principle you suggest is sound. But I don't know how hard this is to code into GIB. I also don't really think that this is necessarily a big drag on its bidding performance vs. just really bad rule priority and range definitions on some auctions.

Quote

Here I have a curious question on Gib. I remembered that Deep Blue (America IBM company produced) had ever beated world first master of international chess,as we know there are two sevices for Gib, one for bidding,one for playing,but for Deep Blue,there are 32 brain (microprocessor),so assume Gibs use Deep Blue to play against experts only with same Gib system, can Gibs samely beat human? With high respect.

Not as currently written. Gib has a huge handicap against experts in defense, given its inability to utilize defensive signaling. That alone would make it get thrashed. Then in bidding it is relying on a hugely buggy bidding database, where it gets thrashed even more. In contrast with chess they had an advantage of database of opening moves vetted by grand masters that had been refined over decades, so for first x moves of chess game if human follow normal "book" opening line the computer making absolutely zero errors. Gib it's just far easier to get it "out of book". Plus in chess everything is visible, if it see position it can understand what's going on, what the threatening moves are, calculate what to do next. In bridge the bids are visible, but if it doesn't have a rule previously written to define it, how can it figure out what the bid mean, because bid meanings are to some extent arbitrary?

Also in general for computer programs, you just don't throw more CPU at it and it automatically gets better. The program has to be specifically written to utilize multiprocessing, divide the work up into different chunk that can be calculated independently & simultaneously, then bring result together and combine them. If program not written with massively parallel processing in mind, it takes A LOT of work to rewrite to utilize the CPUs. Otherwise the extra CPU is useless, 1 is doing all the work and 31 just sit there idle.

Gib against expert for declarer play only (on non-basic robot, high time setting, if it understand opp auction), is pretty close to even. But other aspect it is too far behind.

[lycier]

Stephen Tu, I got it, thank you very much for your excellent reply.