BBO Robot Hands

[tx10s]

Does anyone here really believe that the hands in Robot Bridge are really random? I know for a fact that the finesses are skewed 40% for the players and 60% for the robots, based on analysis of over 2600 finesses/open endings, and that is just the tip of the badly skewed hands iceburg.

[hrothgar]

Does anyone here really believe that the hands in Robot Bridge are really random? I know for a fact that the finesses are skewed 40% for the players and 60% for the robots, based on analysis of over 2600 finesses/open endings, and that is just the tip of the badly skewed hands iceburg.

 

Prove this. To start with, please specify

 

1. A sampling methodology for hands yet to be dealt

2. The test of statistical statistical significant you feel to be appropriate

3. What confidence intervals you are using

[1eyedjack]

Does anyone here really believe that the hands in Robot Bridge are really random? I know for a fact that the finesses are skewed 40% for the players and 60% for the robots, based on analysis of over 2600 finesses/open endings, and that is just the tip of the badly skewed hands iceburg.

How hard would it be for you to post your raw data, rather than just your conclusions? Do you have a qualification in statistics that would lend credibility to your interpretation of the evidence, such that it should be trusted without peer review?

[barmar]

We've been through this many times, most recently

 

http://www.bridgebase.com/forums/topic/66440-tournament-hands/page__p__794169__hl__finesse__fromsearch__1#entry794169

 

Even if we didn't insist we're not doing this, why would we WANT to do this? What would BBO gain from players losing finesses more often than real life?

[tx10s]

How hard would it be for you to post your raw data, rather than just your conclusions? Do you have a qualification in statistics that would lend credibility to your interpretation of the evidence, such that it should be trusted without peer review?

My testing method was to look at the previous 20 tournaments I played, nine different times. The finesse split in those nine different sets was:38.9%, 41.4%, 39.2%. 41.0%, 40.6%, 40.1%, 40.4%, 39.8%, and 38.3%, overall average 40.0% for 2522 finesses.(I eliminated a partial set of 8 tournaments from these statistics, which if included would lower the overall average to39.7%.) That is an average of 281 finesses per set. The odds of not getting more than 40% of the finesses in 280 is about 4000 to 1. I leave it to you to calculate that happening 9out of 9 times if the dealing was totally random. Other observed statistics, players got 4-0, 5-0 and 6-0 spits 5.9% of the time whereas the robots got the same bad splits 3.2% of the time. Players got 3-2 splits 66% of the time (close to the normal 67.8% of the time, yet the robots got 3-2 splits 80% of the time (96 out of 120), a long way from normal, and about 1 chance in 500 of occurring in totally random deals.

 

Posting the entire set of spreadsheet is rather difficult, but I would be willing to E-mail it to you if you really want to see it.

[1eyedjack]

I am having difficulty understanding just what you have done.

My testing method was to look at the previous 20 tournaments I played, nine different times.

Given that you can only play in a specific tournament once, why would you want to look at it more than once?

The finesse split in those nine different sets was ...

How did you identify which two cards in any given deal qualifies for the test of whether it fell within the parameters for a finesse?

 

I could go on, but the reality is that I am probably not the right person to validate your findings (although there will be plenty of readers of this forum who would be capable of doing so, and I hope that one of them steps up to the plate). If your method is sound, it would be a bit of a bombshell.

 

I have a few hundred tax returns to complete between now and the end of the month, so I will certainly not be looking at any data before then. I have downloaded into pbn all of the hands that you played since 02 Dec 2014, all of which seem to have been in robot tournaments. 354 hands in December and 382 (to date) in January. What I can do with this data I am not quite sure. I might send it to Bob Richardson to see if he has time to play with it.

 

As for sharing the data, it should be fairly trivial to upload a spreadsheet to dropbox or google drive - certainly no harder than emailing, and that way you open it up to peer review by a load of individuals, not just me.

 

 

 

[Bbradley62]

My testing method was to look at the previous 20 tournaments I played, nine different times.

I am having difficulty understanding just what you have done. Given that you can only play in a specific tournament once, why would you want to look at it more than once?

It seems pretty clear that OP, on 9 different occasions, possibly separated by several weeks, looked at his most recent 20 tournaments, then aggregated the data for those 180 tournaments.

[1eyedjack]

Thanks. Now you spell it out, it is clear.

[tx10s]

We've been through this many times, most recently

 

http://www.bridgebase.com/forums/topic/66440-tournament-hands/page__p__794169__hl__finesse__fromsearch__1#entry794169

 

Even if we didn't insist we're not doing this, why would we WANT to do this? What would BBO gain from players losing finesses more often than real life?

I do not know why it is the way it is, just that it is. My guess is whoever wrote your randomizer dealing program put the bias in and did not tell anyone. That being said, the data is very conclusive that there is bias in your "random" dealing program.

[tx10s]

I am having difficulty understanding just what you have done.Given that you can only play in a specific tournament once, why would you want to look at it more than once?How did you identify which two cards in any given deal qualifies for the test of whether it fell within the parameters for a finesse?

 

I could go on, but the reality is that I am probably not the right person to validate your findings (although there will be plenty of readers of this forum who would be capable of doing so, and I hope that one of them steps up to the plate). If your method is sound, it would be a bit of a bombshell.

 

I have a few hundred tax returns to complete between now and the end of the month, so I will certainly not be looking at any data before then. I have downloaded into pbn all of the hands that you played since 02 Dec 2014, all of which seem to have been in robot tournaments. 354 hands in December and 382 (to date) in January. What I can do with this data I am not quite sure. I might send it to Bob Richardson to see if he has time to play with it.

 

As for sharing the data, it should be fairly trivial to upload a spreadsheet to dropbox or google drive - certainly no harder than emailing, and that way you open it up to peer review by a load of individuals, not just me.

I looked at all cases of what would be considered standard finesses, like is the K in from or behind the AQ, is the A in front or behind a K, is the Q in from or behind an AKJ, etc. I did not count cases where a Q could be finessed either way, just those cases where the finesse or open ending could only be finesses from one direction.

[tx10s]

I am having difficulty understanding just what you have done.Given that you can only play in a specific tournament once, why would you want to look at it more than once?How did you identify which two cards in any given deal qualifies for the test of whether it fell within the parameters for a finesse?

 

I could go on, but the reality is that I am probably not the right person to validate your findings (although there will be plenty of readers of this forum who would be capable of doing so, and I hope that one of them steps up to the plate). If your method is sound, it would be a bit of a bombshell.

 

I have a few hundred tax returns to complete between now and the end of the month, so I will certainly not be looking at any data before then. I have downloaded into pbn all of the hands that you played since 02 Dec 2014, all of which seem to have been in robot tournaments. 354 hands in December and 382 (to date) in January. What I can do with this data I am not quite sure. I might send it to Bob Richardson to see if he has time to play with it.

 

As for sharing the data, it should be fairly trivial to upload a spreadsheet to dropbox or google drive - certainly no harder than emailing, and that way you open it up to peer review by a load of individuals, not just me.

BTW, I only looked at Robot hands. I do not play pairs that often, and identifying the best hand and evaluating finesses is even more time consuming for pairs games than the ridiculous amount of time I have spent already.

[1eyedjack]

Did you do all this manually, hand by hand, or did you have some automated method?

[tx10s]

Prove this. To start with, please specify

 

1. A sampling methodology for hands yet to be dealt

2. The test of statistical statistical significant you feel to be appropriate

3. What confidence intervals you are using

If you read the post, the sampling was done on hands that were already dealt. The fact that the finesse split ranged between 38.9% and 41% for sets of twenty tournaments averaging 281 finesses/ open endings with an overall average of 40.0% I felt was significant enough

I do not understand question 3 Are you asking my confidence that it is exactly 40%? I make that assumption as writing a bias of 40.2% or 39.8% would be overdoing it, but I can not say that for sure.

[tx10s]

Did you do all this manually, hand by hand, or did you have some automated method?

Sadly, my computer programming skills are insufficient to develop an automated method, so the analysis was done by hand. To further clarify my previous explanation, I evaluated finesses irrespective of who the bidder was. (I did keep records by who was declarer, but since that could vary for the same hand, I felt the only reliable statistic was finesses not regarding declarer.) In other words, for example, if there was an AQ in either direction, I credited the finesse to the robot if the K favored the robot and credited the finesse to the player if is favored the player. It is pretty straight forward identifying finesses, so I feel my statistics are reasonable. BTW, the nine sets of 20 tournaments are spaces over the past 4 years.

 

There seems to be a lot of negative push back from BBO. I have no ax to grind other that hoping the randomizer could become truly random. My concern also has nothing to do with my success rate in the game. I am currently averaging scoring master points over 70% of the time, which actually surprises me as being that high.

[barmar]

Could this be some unexpected consequence of the "best hand" feature?

 

The way this works is that we first deal all the hands normally and calculate each hand's HCP. If South doesn't have the best hand, we simple swap his hand with the one that does. If there's a tie for the best hand among the other seats, we pick one of them at random.

 

I don't see how this could result in a finesse bias, though.

[barmar]

Did you count 2-way finesses in your statistics? If you guess wrong, I don't see how that's the software's fault.

[1eyedjack]

Did you count 2-way finesses in your statistics?

Apparently not::

I did not count cases where a Q could be finessed either way, just those cases where the finesse or open ending could only be finesses from one direction.

[tx10s]

Did you count 2-way finesses in your statistics? If you guess wrong, I don't see how that's the software's fault.

No, as I stated earlier, I did not count two way finesses, only finesses that could only be played in one direction

[Bbradley62]

Since I was puzzled as to how your statistics are so different that the ones I recently posted, I looked at the 11 tournaments that you played 1/14-1/19. Being a little lazy, I only kept track of which finesses worked/didn't for you when you declared (105 of the 132 hands). I also only kept track of missing Aces, Kings and Queens, so that my spreadsheet would be a little easier to follow if you want to look at it (although I don't see how to post it). There were certainly many opportunities to finesse for missing Jacks and Tens, but I did not tabulate those.

 

On those 105 hands, you had 49 opportunities to finesse for a missing Ace (when you had the K but not the Q of the suit in question). 28 of the 49 Aces were onside.

 

On those 105 hands, you had 51 opportunities to finesse for a missing King. Sometimes you held AQJ of the suit between the two hands (in which case "onside" meant in front of the Ace) and sometimes you held A in one and Q in the other without the J (in which case "onside" meant in front of the Q. 31 of the 51 missing Kings were onside.

 

On those 105 hands, you had one-way finesses for missing Queens 36 times, and the missing Queen was onside 13 times.

 

So, on those 105 hands you had 136 opportunities to finesse for missing A/K/Q (not that you took all those finesses, or even that it was a good idea to do so) and the missing honor was onside 72 times (53%).

[Bbradley62]

Since I don't see how to post the spreadsheet, and my dataset isn't all that big, I'll copy and paste it here, although I'd still appreciate it if someone explained how to post a file. Well, I can't get it to format nice, and I can't find the recent post about how to make nice columns, so I'll try again tomorrow...

[Antrax]

You can use Google Docs, they have a decent spreadsheet.

[ArtK78]

It astonishes me that players continue to think that computer generated hands are biased against a player. Quite frankly, even if a bias existed, everyone still plays the same hands.

 

If there really is a bias in favor of finesses not working, one should incorporate this into one's overall strategy of play. Let me know if you suddenly become a master of the robot games by playing for all finesses to fail.

 

When I was a new player (back in the 70's) people complained all the time about computer hands. I rarely hear that anymore. But every so often someone complains about a bad break by breaking out the old "darn computer hands" complaint.

[tx10s]

Since I was puzzled as to how your statistics are so different that the ones I recently posted, I looked at the 11 tournaments that you played 1/14-1/19. Being a little lazy, I only kept track of which finesses worked/didn't for you when you declared (105 of the 132 hands). I also only kept track of missing Aces, Kings and Queens, so that my spreadsheet would be a little easier to follow if you want to look at it (although I don't see how to post it). There were certainly many opportunities to finesse for missing Jacks and Tens, but I did not tabulate those.

 

On those 105 hands, you had 49 opportunities to finesse for a missing Ace (when you had the K but not the Q of the suit in question). 28 of the 49 Aces were onside.

 

On those 105 hands, you had 51 opportunities to finesse for a missing King. Sometimes you held AQJ of the suit between the two hands (in which case "onside" meant in front of the Ace) and sometimes you held A in one and Q in the other without the J (in which case "onside" meant in front of the Q. 31 of the 51 missing Kings were onside.

 

On those 105 hands, you had one-way finesses for missing Queens 36 times, and the missing Queen was onside 13 times.

 

So, on those 105 hands you had 136 opportunities to finesse for missing A/K/Q (not that you took all those finesses, or even that it was a good idea to do so) and the missing honor was onside 72 times (53%).

I looked at the same hands, and my count was similar, 73 out of 134, actually slightly more optimistic than your count. I then included all of the tournaments since I made my posting, and the finesses for the players were 53% successful, the highest I have ever seen for that many tournaments in all of my reviews. I think we can agree that there is no significant difference in how we evaluate finesses. Below is a copy of the spreadsheet of the 20 tournaments I evaluated before I posted. (It is interesting that I had the highest success rate after posting my comment, but I will leave conspiracy theory to others.) If you have time to check the previous 20, I would like to know what numbers you came up with.

 

Finesses

Player Robots Total

Date Tourn Win Lose Win Lose Win Lose

31-Dec 5357 4 6 1 4 7

1-Jan 9466 5 6 2 3 8 8

1-Jan 9705 6 7 1 7 7

1-Jan 864 2 3 2 3

2-Jan 7203 4 6 1 5 6

3-Jan 3698 3 9 3 1 4 12

4-Jan 9091 7 10 1 2 9 11

4-Jan 92 6 5 4 3 9 9

5-Jan 5308 4 5 7 1 5 12

5-Jan 6810 4 12 1 1 5 13

6-Jan 1798 4 1 1 5 1

6-Jan 3864 3 7 1 3 8

6-Jan 4027 7 1 1 7 2

8-Jan 3839 7 12 7 12

8-Jan 4589 3 11 2 2 5 13

10-Jan 1077 2 7 2 2 9

11-Jan 5532 3 9 2 1 4 11

11-Jan 6086 4 3 1 4 4

12-Jan 1855 4 5 5 1 5 10

12-Jan 2753 3 7 4 2 5 11

Totals 85 132 37 20 105 169

 

Pct 39.2% 64.9% 38.3%

[tx10s]

I looked at the same hands, and my count was similar, 73 out of 134, actually slightly more optimistic than your count. I then included all of the tournaments since I made my posting, and the finesses for the players were 53% successful, the highest I have ever seen for that many tournaments in all of my reviews. I think we can agree that there is no significant difference in how we evaluate finesses. Below is a copy of the spreadsheet of the 20 tournaments I evaluated before I posted. (It is interesting that I had the highest success rate after posting my comment, but I will leave conspiracy theory to others.) If you have time to check the previous 20, I would like to know what numbers you came up with.

 

Finesses

Player Robots Total

Date Tourn Win Lose Win Lose Win Lose

31-Dec 5357 4 6 1 4 7

1-Jan 9466 5 6 2 3 8 8

1-Jan 9705 6 7 1 7 7

1-Jan 864 2 3 2 3

2-Jan 7203 4 6 1 5 6

3-Jan 3698 3 9 3 1 4 12

4-Jan 9091 7 10 1 2 9 11

4-Jan 92 6 5 4 3 9 9

5-Jan 5308 4 5 7 1 5 12

5-Jan 6810 4 12 1 1 5 13

6-Jan 1798 4 1 1 5 1

6-Jan 3864 3 7 1 3 8

6-Jan 4027 7 1 1 7 2

8-Jan 3839 7 12 7 12

8-Jan 4589 3 11 2 2 5 13

10-Jan 1077 2 7 2 2 9

11-Jan 5532 3 9 2 1 4 11

11-Jan 6086 4 3 1 4 4

12-Jan 1855 4 5 5 1 5 10

12-Jan 2753 3 7 4 2 5 11

Totals 85 132 37 20 105 169

 

Pct 39.2% 64.9% 38.3%

When I posted this, the data columns lined up nicely. When I went back to add a comment, I see that the data does not present well. In each row, the first item is the date, the second is the tournament number. The third is finesses onside for the players, the fourth is finesses offside for the player, the fifth is finesses onside for the robots, the sixth is finesses offside for the robot, the seventh is the sum of finesses onside for the players and offside for the robots (in other words the finesses that favor the player) and the eights data point is the sum of finesses offside for the player and onside for the robots.

[tx10s]

It astonishes me that players continue to think that computer generated hands are biased against a player. Quite frankly, even if a bias existed, everyone still plays the same hands.

 

If there really is a bias in favor of finesses not working, one should incorporate this into one's overall strategy of play. Let me know if you suddenly become a master of the robot games by playing for all finesses to fail.

 

When I was a new player (back in the 70's) people complained all the time about computer hands. I rarely hear that anymore. But every so often someone complains about a bad break by breaking out the old "darn computer hands" complaint.

That was great sarcasm, but ignores the data. Obviously, if all finesses were offside, the game would be easier. Obviously, everyone plays the same hand, but I'm sure the experts (not me) would know how to adjust their strategy for the bias built into a dealing program. I have already seen that at BBO by all the people who bid the frequent "cheap" slams that the robot program puts out. When I say cheap slams, the average slam hand that I have looked at (about 30) has an average of 29 HCP and barely over 30 total points. Additionally, I have looked for hands with over 30 points that do not make slam, and they are very few and far between.