[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] Wrong cube errors categories in analysis statistics pane

From: olivier croisille
Subject: Re: [Bug-gnubg] Wrong cube errors categories in analysis statistics panel
Date: Wed, 01 Oct 2003 14:07:55 +0200

I'm computer-litterate but hopelessly useless when it comes down to programming/coding, so I'll leave the 'technical' part of the discussion for you to discuss with Jorn et al.

So, back to practical things as far as I'm concerned. I think our discussion illustrates the proverb "le mieux est l'ennemi du bien" as they say in French, no idea of the exact translation, if any, means something like "wanting to do something even better might just end up doing harm to what was already quite good".

Obviously no offence intended, I have too much respect for gnubg developpers, but IMHO we should just set back to former analysis doubling categories, that were perfectly satisfactory for all *practical* purposes and were not buggy (doubling error wrongly classified, see first post of the thread and its attachment)

Granted, in theory, it lacked analyzing players' doubling strategy around CP. But, again, in practical terms, the CP information is of very limited value-added, and just adds confusion :

- first, as Holger correctly stated, there is no such thing as a a 'Wrong double around cash point', so this stats line should disappear.

- now, what about 'Missed double around cash point"? Again, from a practical point of view : why would a player carry on playing around his CP? Surely he knows he has a double - otherwise, I should urge him to read Robertie's 501 three times in a row :-)) - because around CP his position must make him a large favorite in the game. So, if he plays the position instead of cashing it, in his mind it HAS to be a 'too good' problematic and reasoning. OK, I realize my explanation is a tad tedious, but I hope you'll get the picture. My point is 'Missed double around cash point", even if this is theoretically incorrect, should be merged with 'Missed double around TG', as was the case before, because in essence they refer to exactly the same nature of decision by players. Actually, since it was the cashing decision that was correct and missed, it even should be the other way round : it is 'Missed double around TG" that should be merged into 'Missed double around CP'

So I suggest getting back to former categories...

                                     GNU                  FRENCHKISS
Missed doubles around DP              0                    0
Missed doubles around TG              0                    0
Wrong doubles around DP               0                    0
Wrong doubles around TG               0                    0

- and, yes, I chose on purpose a game where I made no mistake, wasn't that easy to find :-) -

... but renaming 'Missed doubles around TG' in 'Missed doubles around CP', because again this illustrates better the nature of the decision that had to be made. But for the same reason I would leave 'Wrong doubles around TG' untouched, because here the position was indeed too good.

French Kiss on GG

From: Jim Segrave <address@hidden>
Reply-To: address@hidden
To: Holger <address@hidden>
CC: olivier croisille <address@hidden>, address@hidden
Subject: Re: [Bug-gnubg] Wrong cube errors categories in analysis statistics panel
Date: Wed, 1 Oct 2003 02:09:16 +0200

Warning - from someone whose theory is less than awe-inspiring"

Imagine MSC as a line, increasing in value left to right:

You lose------------------------------------------------You win gammon
                DP               CP        TG
         1      |   2        3   |   4    5 | 6

The numbers represent regions of MWC.

So for each region, the limits and associated errors:

1 = MWC < DP, error:
        double:    wrong double around DP
        no double: no error
        take:      no error
        drop:      wrong drop round DP

2 = DP < MWC < .5 * ( DP + CP)
        double:    no error
        no double: missed double around DP
        take:      no error
        drop:      wrong drop round DP

3 = .5 * ( DP + CP) < MWC < CP
        double:    no error
        no double: missed double around CP
        take:      no error
        drop:      wrong drop round CP

4 = CP < MWC < .5 * (CP + TG)
        double:    no error
        no double: missed double around CP
        take:      wrong take around CP
        drop:      no error

5 = .5 * (CP + TG) < MWC < TG
        double:    no error
        no double: missed double around TG
        take:      wrong take around TG
        drop:      no error

6 = TG < MWC
        double:    wrong double around TG
        no double: no error
        take:      wrong take around TG
        drop:      no error

Re the other questions

> Does someone know why getMatchPoints() does this and whether this is
> intended?

I can't reproduce this.

From today's CVS, gdb output reformatted for ease of readability:

Breakpoint 2, TheoryUpdated (pw=0x899c740, ptw=0x899b900) at gtktheory.c:331
1: aaarPointsMatch = {{{0.42115128, 0.328579843},
                      {0.589850724, 0.589850724},
                      {0.759432137, 0.767605424},
                      {0.947443664, 0.767605424}},

                     {{0.240567863, 0.232394576},
                      {0.410149246, 0.410149246},
                      {0.57884872,  0.671420157},
                      {0.808395922, 0.671420157}}}

And for player 0 in the Market Window, I see the values exactly as
reported above, converted to percentages. There's no swapping visible.

For those who want to test this, this is from:

    GNU Backgammon  Position ID: sLnJAEzYvg0ICA
                    Match ID   : cInxAAAACAAA

> To categorize cubes gnubg compares the arithmetic means of DP and
> CP, and CP and TG with the winning probabilities.  Could someone
> please explain me why (and confirm that) this is correct?

See above, I think it is correct.

> I'm wondering about the following: The winning probability
> (aarOutput[ 0 ][ OUTPUT_WIN ]) doesn't include any added value for
> gammon or bg chances (aarOutput[ 0 ][ OUTPUT_WINGAMMON ] and
> aarOutput[ 0 ][ OUTPUT_WINBACKGAMMON ]). Those match points (DP, CP,
> TG) have an entity of cubeful equity, don't they? And I think
> equities are calculated as the sum of winning chances, gammon
> chances and bg chances, thus include gammon and bg value. If all
> this was correct so far, then I suppose these 2 values can't be
> compared.

This seems a good point - I think picking the point needs to be done
in terms of MWC, it's not enough to do equity, since in some match
points, gammons are irre;evant.

Jim Segrave           address@hidden

MSN Search, le moteur de recherche qui pense comme vous ! http://search.msn.fr

reply via email to

[Prev in Thread] Current Thread [Next in Thread]