And the results are in!

The results came through for the little (read: $50,000) match-up between Phil Laak, Ali Eslami and the poker playing AI program 'Polaris' (reported here)

First you might be glad to know that the humans won, but it was a close one. The players played four sessions each played simultaneously in separated rooms. Each of the players received the cards Polaris was getting in the other room and vise versa (so if Polaris got a A9o against Laak who got a KQs , at the same time Eslami would get an A9o against Polaris's KQs - Obviously neither the players nor Polaris was able to communicate to its counterpart in the other room).

The first session was declared a tie with Elsami up by $395 and Lakk down by $465. From the second session onward Laak found his rhythm and didn't lose another session (although he only won the fourth one by a mere $110), he finished with a total win of $2670. Eslami on the other hand lost the second session (down a staggering $2495) and the third one (down $635), he did finally manage to snatch a win on the last session but the $460 did not help his cause and he finished with a cumulative lost of $2265 which Laak's win only just covered.

In the post game comments, both Laak and Eslami agreed that they were forced to play the best heads-up game they could master and that they could not have continued much longer and still sustain a win. This is a big compliment to the Polaris programmers, since it establishes the fact that the program was not predictable and managed to 'play the players'.

In the same video (which you can find here along with the game's blog) Eslami states that he played extremely predictable and that anyone with knowledge of poker would have been able to find Eslami's range ("the best heads-up poker he can play"?). I find that odd, since I would imagine that the first thing you do when faced with a poker-bot (which can calculate everything that is happening in seconds) is play anything but predictable; by changing gears, representing hands and even sacrificing a bit of chips in order to tell the bot that you might call him all the way on Jack high. It appears that this is exactly what Laak did, which made Polaris fold more hands and sometimes pay with fourth pair when Laak had third pair (Laak even decided to make a shirt which says "Polaris Pays").

We'll still have to wait for the team from the University of Alberta to posted the detailed hand logs to see how Polaris truly plays but for now we can look at some highlight hands to draw some conclusions. It seems Polaris has the ability to minimize his loss in those hands you can't get away from such as when Laak turned a full house with his Q8o while Polaris turned trip 8's, while on the same hand Eslami lost quite a bit more. Polaris also caught a few of Laak's bluffs when he called Laak's Queen high river bet with King high. However, as mentioned above, Laak did manage to get Polaris to pay on those marginal hands.

The 'Polaris experiment' is still far from over, as even after four sessions the sample of hands played is not large enough to draw a conclusion from, especially in a game like poker, where you always need to account for variance (even more so since the win was by such a small margin). Still, the programing team definitely has a lot of material to help them understand how to improve and maybe (read: we hope) later pit Polaris v2 against Doyle Brunson and Phil Ivy.

- Random -