Join Bridge Winners
All comments by Veronique Ventos
You are ignoring the author of this comment. Click to temporarily show the comment.
Tx :) I went to Boston last year to present our paper related to the boosting of Wbridge5 but I will return there with pleasure ;)
Sept. 19, 2018
You are ignoring the author of this comment. Click to temporarily show the comment.
Very interesting Richard. You should definitely visit us in paris for at least a week
Sept. 19, 2018
You are ignoring the author of this comment. Click to temporarily show the comment.
Just come back with my NukkAI team from the International Logic Programming conference where I presented as first speaker our paper related to bridge and Artificial Intelligence “The game of Bridge: a challenge for ILP”. There were some “star researchers” very interested by our work concerning a supervised learning problem in Bridge: given a ‘limit hand’ (11HCP), should a player bid or not, only considering his hand and the context of his decision.

If you are interested, our paper is pp 72 to 87 in the Publication of the ILP 2018 regular papers, LNAI 11105 Downloadable for free until October 1st 2018 :
https://link.springer.com/content/pdf/10.1007%2F978-3-319-99960-9.pdf
Sept. 10, 2018
Veronique Ventos edited this comment Sept. 10, 2018
You are ignoring the author of this comment. Click to temporarily show the comment.
I love this stuff !!!! thanks
It's very useful for humans of course but also for the designing of a Bridge Artificial Intelligence based on human expertise like the one under construction in our AlphaBridge project :)
Aug. 25, 2018
You are ignoring the author of this comment. Click to temporarily show the comment.
Tx Richard :) the current work is far from the seed approach about which we had a lively discussion :) by the way the insights you gave us were very useful
June 30, 2018
You are ignoring the author of this comment. Click to temporarily show the comment.
Sure :) I'm an AI researcher and since 3 years I focus my research on the game of Bridge. The first step was to convince my community that bridge is a killer application for AI. It was difficult but I have been lucky enough to bring together top level researchers and my friend Y. Costel (Mr WBridge5) to work on this project.
There are different challenges related to bridge and we have created the Alphabridge project whose aim is to address these challenges thanks to an innovative hybrid architecture involving several modules based on different AI paradigms(see my page for more details:
http://vvopenai.monsite-orange.fr/).
Alphabridge has considerably grown and has led to a corporate project: NukkAI, a private AI lab based in Paris but with international collaborations, was born on May 7 2018 !
The website will be online soon but you can already follow us on social media like https://www.facebook.com/NukkaiLab/
June 29, 2018
You are ignoring the author of this comment. Click to temporarily show the comment.
:)
Dec. 30, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Sorry for my answering late but I have asked my co-authors to answer your very interesting questions.

I happen to have had the same concerns as yours when O Teytaud proposed to adapt his « seed approach » to Bridge: I could not figure out how such a method could even work on previous games and was more then skeptical about bridge.

Finally it does seem to work but as for the other games it is completely empirical.

I agree with Michal statement “It is plausible that certain subsets of random data better capture the essential properties of a given game” , my intuition
is that in some datasets, generated boards are more different from each other (heterogeneous sample). Indeed, it seems that the best seed generates more unbalanced hands (bad seeds generate dull, balanced hands (4432 or 4333), please do not ask me why :) )

As described on my homepage, I am a bridge player and a specialist in several areas of AI like Knowledge Representation and Machine Learning and do not pretend to master every aspect of the project. Richard is much more qualified than me on the PRNG part, so is O Teytaud who unfortunately has been less responsive since the day he got hired by Google last year :) The idea of the project is to interact with specialists of different areas !

Here are some details in response to the different questions:

There are only 127 simulations max per decision: this small number is related to the limited processing power of computers used in robot competitions and time limit imposed. Systems must use anytime algorithms and give an answer in a limited time (like 10mn/board).

WBridge5 uses the Delphi Pascal (the language used by Yves Costel) PRNG. This confirms richard's hypothesis : “Alternatively, the developer might be using some kind of awful awful awful RNG algorithm that has properties that we would normally think of as ”bad“, but might prove helpful in this example.”

There is no offline treatment: when there are few constraint, the program generates hands without constraints (the 32 bits of the system are not enough to code every possible hand) and keeps the one consistent with the bidding sequence. This allows to respect the probability distribution. When there are more constraints it is more difficult because one has to generate under constraints (this part is certainly improbable with a statistic approach).

The seed is initialized once and for all at the beginning of each match (64 deals).

I recall that our goal was to test the seed approach on the current version of WBridge5, with all its strengths and flaws and not to globally improve WBridge5 that belongs to Y Costel who designed it more than 20 years ago and has been updating it on his own. For this reason we did not look into the code (like the other bridge robots, there is no open source code and neither Olivier nor me are allowed to see Yves' code). We did not even use the powerful machines available at our lab. The game conditions are the same as in the World Championships (Windows, etc).
I also recall that the goal of AlphaBridge is not to beat the actual bridge robots like Wbridge5 with whom we continue to collaborate. it will be unfair to compete with so different processing power. The double objective for our AI is : beating the best human players and providing tools to improve human practice

In recent months, we have moved to a phase where we were building modules of the new architecture AlphaBridge.
For instance, I am currently building a Rule-based machine learning system allowing to automatically update a set of hand-crafted rules according to expert decisions.
This work is totally apart from any existing Bridge AI; when this work will be published I will be much more comfortable answering questions about this part since Inductive Logic Programming is one of my research areas.

Thanks anyway for your interest Richard, Michal , Nick, Jim, Mike etc … Thank you kevin for posting my paper on BridgeWinners,
research is something that evolves thanks to interaction.
Dec. 30, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Hi Nick,
to avoid copyright pb, I created this Draft version
http://vvopenai.monsite-orange.fr/file/6d60eb48fbc683d829f1727dafb7e13e.pdf
Dec. 29, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
I must tell you one thing, despite my name I'm not native English speaker ;) joking aside if people want to help us to correct our next article from a grammatical point of view, I will be very grateful.
From a scientific point of view, I know exactly where we are going with this project which already has the merit of having attracted my fellow AI researchers who did not want to hear about it before :)

For more details here is my last webinar (sorry for my terrific french accent :p) :
https://www.youtube.com/watch?v=u8x4T9bwUmM&feature=youtu.be
Dec. 28, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Hi Richard, thanks for your interesting comments.
Concerning the cross-validation, protocol is the same
than the one in the paper of Olivier Teytaud:
https://www.lri.fr/~teytaud/loo.pdf
The general idea is to study the impact of the best seed keeping a dozen points out of the forty to form the test set (for K = 1, 2, 3,…, 30 with K the size of a randomly known submatrix).
Here is the algorithm :

Pseudo-code :
nbArms = param
M=costel's matrix
for K=1, 2, 3, …, 30
{
for i=1, 2, …, n (n at-least 500)
{
M=M(randompermutation(40), randompermutation(40))
submatrix=M( 1:K, 1:K )
testmatrix=M( 1:K, (K+1):40 )
score(1) = sum of line1 of M
score(2) = sum of line 2 of M

score(K) = sum of line K of M
Let I=(i1,…,inbArms) the list of index with the best scores
results(i) = mean of scores related to lines in I
}
performance(K)= mean of results(i)
}

\end{verbatim}
Dec. 28, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
The goal of these previous works was to study the impact of the seed on an existing AI without changing the conditions (number of simulations, PRNG etc …). We didn't even use the huge processing power of the computing center of Paris Saclay :)
Dec. 28, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Hi Nick, I can send you an author version of the paper.
Dec. 28, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
Ercan: the point is that numerical approaches used by AlphaGo are successful for many applications but they are black boxes. AlphaGo is able to “compute” step by step the best move according to its training but it does not learn explicit rules. Therefore, AlphaGo can't transfer its knowledge to humans or to another program. In the same way, if the learning task of a robot is to sweep in a room without knowing where are the different obstacles, such approaches (Reinforcement Learning for instance) allow the robot to sweep better in the room using trial errors. But in this case, the robot does not learn domain theory like “there is a chair in the middle of the room”. Now if you want the same robot to make a task like vacuuming in the same room, learning will be back to zero and the robot will repeat the same errors.
Jan. 24, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
You are right but before this crucial step, bridge programs can beneficit of new techniques developped in ML and not only from deep learning paradigm. Partition search and other ideas from Ginsberg were great but we can imagine using other new methods. For instance, I have worked since one year with Yves Costel and Olivier Teytaud on the boosting of Wbridge5 according to the choice of random seeds used for Monte-Carlo simulations. We adaptated a seed methods to bridge and the experiments allow us to find a better seed than the current one. The gain was estimated to 0.1 imp per board and Yves Costel decided to use this best seed in the 20th World Computer-Bridge Championship held in Wroclaw in september 2016. Wbridge 5 finally become world champion winning the final with 6 imps margin on 64 boards :)
Jack was not there but a new concurrent Xinrui was a serious opponent.
Now, I look for Lyon 2017 :)
Jan. 23, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
I'm an AI researcher (in the fields of Machine Learning)and I believe that bridge can be the future challenge for AI :) I wrote a paper on the description of different challenges inherent to bridge (don't ask me for the moment, the first article being written in French and the others in submission). The main point is that alphaGo can beat human but it cannot explain how. In bridge, you are judged on your explanations so … My main idea is to combine symbolic and numeric approaches in different components of the AI. Symbolic ML is called GOFAI (good old fashion AI) but I think it will soon come back as we need to really solve the game and not only to win.
Jan. 23, 2017
You are ignoring the author of this comment. Click to temporarily show the comment.
I am a researcher in Artificial Intelligence in a machine learning team in Paris and sometimes a bridge player ;).
I work on AI techniques related to numerous bridge challenges. For us, this game is like the goose that lays the golden egg (actually huge amount of robust data in a predefined format).
I think that computer bridge is the next challenge for AI and that AI can improve bridge's popularity ratings (Have you read about AlphaGo ? Do you think that google, facebook only need go or atari's games programs :) ? ).
I have worked since several months with Yves Costel on the boosting of his AI according to the choice of random seeds used for Monte-Carlo simulations. I went in september to Wroclaw to see a real championship of computer bridge and to support Wbridge5.
I just want to thank Al and computer bridge people I saw in Wroclaw in this small room without any advertisement. I saw passionate, talented people who are pleased by selfless actions.
The fact is that maybe it is not enough interesting since we didn't hear a bridge bot coughing neither putting its mouse vertically or horizontally :p
Sept. 27, 2016
.

Bottom Home Top