This blog provides supplementary thoughts and ideas to the site. If you haven't seen the main site, there is a lot there including the Martel and Rodwell interviews, photos, and articles. This blog is focused on advancing bridge theory by discussing the application of new ideas. All original content is copyright 2009 Glen Ashton.

Tuesday, November 03, 2009

The Robot Battles Part II - Disguise and the Bot Tell

This is the second of a series of posts on battling the BBO Robots.

In the late seventies at the University of Waterloo I attended a guest lecture by a world expert in artificial intelligence (AI). With incredible prescience he explained that the future of AI was not rule-based, by having a program judge a position by applying many rules (e.g. if the following twenty conditions are met, the best line will be a cross ruff), but by using processing power to examine end states (e.g. play out thousands of different lines, find out cross ruff works best in the majority of cases). Thirty years later, the BBO Robots with the GIB software use the processing power method to "think" - they are constantly running simulations of the current position to determine what will likely work best.

To run these simulations the BBO Robots take what they "know" - what information they have about the current layout - produce many example hands that match the knowledge, and then quickly play out each of the example hands to determine what seems to work best much or all of the time. It repeats this step ever time it has to play a card, and can even change its mind from card-to-card if new knowledge has been gained, or if the example set of hands becomes skewed in a new direction.

One of the best ways to combat the BBO Robots is to hinder them from knowing the true situation. If they create the wrong example hands, they will often play less than optimal. Thus one should try to disguise or hide the actual situation from the robots, and to only reveal key information when you need to.

This matches the situation in live bridge, where most players don't disguise their holdings enough. BBO Robot tourneys are a great place to practice your disguise techniques, given that Halloween is over and trying out disguises at your local bridge club may not be well received - okay, given that you want to try the techniques over and over again, to see when and how to use them.

Say you are declarer, and have 642 in hand in a suit, with decent length and strength in dummy. The vast majority of players will play the 2 first, the 4 next time, and the 6 last, even though saving the 6 and the 4 makes absolutely no difference. It can be better to play the 6 in some circumstances, so that your opponent, live or robot, doesn't know where the 4 and 2 is. For the robots, not knowing where the 4 and 2 is can change the example hands they produce to try to work out the best play to hurt you.

Against live opponents you want to use a mixture of disguise and truth - sometimes play the 2, sometimes the 4 and sometimes the 6. If you always played the 6 from 642, they would then know what you have or don't have when you play the 2 first. However Robots don't learn from hand-to-hand - they treat you as a new opponent. Thus you want to maximize your use of disguise against the Robot opponents, and that's why the tourneys are a great place to try out these techniques.

Say I opened 1NT with a six card major (more on this in the next post), and end there. If the Robots lead that suit, I will pick my cards to play the suit as if I had a nice three card suit, instead of true situation. The Robots will often be fooled by this disguise - they will assume, in their example layouts, that their partner is the one with the length in the suit, and defend, and mis-defend accordingly. Always look to assist the Robots in misreading the actual situation, by using disguise, and hiding information.

Say you are playing out a hand, and you will need to give up a trick to the Robots, and at that point they can hurt you with one play, and help with another. If you delay giving up the trick, the Robots gather more and more information, and then when you give up the trick their example layouts are very accurate. If you give you the trick early, and have hidden key information, the Robots will often form a wrong picture of the hand, and make the play that helps you.

This brings us to the "tell" and how it can help you. As Wikipedia tell us:

A tell in poker is a subtle but detectable change in a player's behavior or demeanor that gives clues to that player's assessment of his hand.
In live poker at the very cheap levels, most of your opponents will seemed to have gone to the Jessica Simpson school of acting. For the robots, early in the hand they don't know how to maintain a poker face, let alone read the song like Christopher Walken:

Early in the hand, if a robot has a number of choices to make, it generates examples and determines what seems to be the best play. If the robot has limited or no choice, it does not need to do this, and can play fast. Hence robots can quickly play singletons, or from cards that are equivalently the same (e.g. 345, or if the 7 has been played, 568). Thus the speed the robot plays a card is a "tell".

Say you have KT9764 in your hand and AJ3 in dummy. Early in the hand, you lead the 9, LHR (left hand robot) plays small quickly, you win the ace, and RHR plays small slightly slow. Now you led the jack, and RHR hitches (i.e. slow plays), before playing small in the suit. What is the robot thinking? If he didn't have the queen, there would be no need for robot thought, so let the jack ride.

Important warning 1: the same "tell" does not work later in the hand, since the robot does not have much cards left for thinking - the robot can play in tempo, and is quite capable of smooth ducks.

Important warning 2: the programmers can (and should) remove the "tell" by adding appropriate pauses into the robot play.

Watch for the bot "tell", and remember the longer the bot thinks, the more it has to think about.

The longest "tell" I've seen in the bidding is after 1H(bot)-P-1NT (forcing)-P;-?. If the robot opener has exactly 4Ss and 5Hs, he will now take a long time to rebid, sometimes rebidding 2H, sometimes passing, and sometimes bidding a minor. Most other hands provide for easy rebids following the structure, but this hand type does not fit into the rebid structure provided to opener.

Next up: hoggin' the hand.


  • At 10:44 AM, Blogger Memphis MOJO said…

    This is great stuff. I can't wait for part 3!

    Did you decide to play extensively for one month as an experiment?

  • At 12:34 PM, Blogger Memphis MOJO said…

    I forgot to ask this: Is there a place on BBO that explains what system the bots play?

  • At 8:48 PM, Anonymous Anonymous said…

    You can also hinder the bot by refusing to reveal cards that make other cards the robots hold equivalent, since it greatly reduces the number of nodes that GIB needs to expand to evaluate each sample deal.

  • At 9:09 PM, Anonymous Anonymous said…

    Another thing to keep in mind is that hands with voids are very difficult for the bots to analyze because the number of cards a player with a void can play on a given trick is much greater. This is really only significant early in the hand, the complexity of analyzing card play positions is exponential in the number of cards remaining (so most of the effort is expended at T1, with a lesser amount at T2). So you can expect the bot to play particularly badly on these hands, so when you hold a void, you really really want to hog.

  • At 12:12 AM, Blogger Paul said…

    The robots system file is available here

  • At 6:20 AM, Blogger Glen Ashton said…

    Thanks for the comments. Answer to a question: I only decided to play for all of October after I had played the first few days and found myself on the leader board, and was having a lot of fun. For myself, I figured out the bot system by clicking on their bids as a learn as you go approach. Sometimes they didn't have their bids - it goes 2NT(me)-3C(stayman);-3D-4D: clicking on 4D says 5+Ds, 8+ HCP - turns out it has a 1-4-5-3 with only two queens.

  • At 2:30 PM, Anonymous Chris Hasney said…

    This is a great series


Post a Comment

<< Home