I wanted to participate on this year's AI Championship, THNL HU.
I even wanted to make some considerable bet on it as posted before!
After a tough time to create a sophisticated AI the testing phase with the University of Alberta started.
I did not face any issues before but a few days before the AIs actually should start playing against each other they were put on their server. Suddenly there was a massive amount of timeouts. After some investigation the reason was found: they work with single core machines. SINGLE CORE MACHINES!!!
I was able to accept the fact that the THNL HU is played with 200BB instead of 100BB which itself is ridiculous. I was also able to accept the fact that there is an average time per hand (not per decision) of 7 seconds which accumulates to a total time. In that total time you have to play the given number of hands. Totally unrealistic as well.
But single core - come on!
It really appears like the environment created strongly benefits their own AI.

I complained about it several times and tried to find a solution before I had to withdraw in the 11th hour. Here some extracts from my letter to the chair:

The stated aim of the contest is to encourage the development of a wide variety of new AI techniques and encourage exploration in this area. But the structure of the contest undermines this aim.
The servers provided to run the competitor's software are single-core machines on relatively low gigahertz processors, the kind of specification more reminiscent of the last century. Because of the 7 seconds per-hand time-limit used, this effectively bars more sophisticated programs from entry and biases the competition in favour of entries that use large amounts of pre-computation.
Average desktop machines have been more powerful than the Amazon medium ec2 machines used in the contest since the very early 2000s. We have been working on machines ten times as fast, but even our 'worst-case-scenario' testing was carried out on 2-core virtual machines. On our hardware our slowest time for a spot is around 2.4 seconds; on the Amazon server the same calculation takes in excess of 45 seconds. While we did not expect exactly state-of-the-art equipment in the competition, the magnitude of this discrepancy is almost unbelievable.
We offered to pay the costs of running the competition on a more up-to-date specification, but this offer was declined.
We are not the only competitors to have faced timeout problems, and to have faced the choice between crippling our AI or leaving the contest.

It looks like the overall effects of the constraints is to bias the competition in favour of the entry by the University of Alberta; I am sure this is not intentional, but it hardly encourages diversity.

you have a point. at the same time, isn't it fairly expensive to purchase enough computers with reasonable processing power ?

just be a boss and start your own contest?

I think a great training tool would be a 9 person ring game, with 8 of the players being advanced AI bots, and the 9th player being the trainee.

Anything like that available?

lol at limiting any computer based contest to a 1 core machine in 2012, especially AI.

I'm confused, you realize that a quad core desktop is like $350 right? (for real, not a lame tree-fiddy joke). Meanwhile 24 and 32 core linux boxes, while a lot more expensive that $350, are standard hardware in a lot of tech-intensive industries. Maybe I am misinterpreting your post.

bots are going to ruin online poker!

Going to? With all the sophisticated software available to help with real time decision making I think many full time online players essentially are bots.

Live play is the only pure form of poker. Online poker, the way it has evolved, is killing the game through its scams, fraud, and player dependence on sophisticated computer technology.

Partially playing devil's advocate but one could argue that limiting the processing power available forces the entrants to actually work on 'artificial intelligence' as opposed to more run of the mill solutions. What I mean is that AI problems can generally be approached in one of two ways.

1. Smart approximations of brute force. These can be your standard minimax solutions or something like an adaptive monte carlo solution.

2. More rule and state based systems that attempt to simulate how a person might think about the game.

The first solution much easier to develop than the second but it's effectiveness is almost exclusively dictated by the search space being considered and power of the hardware available. The second solution is vastly more challenging to develop but is much less constrained by the hardware specifications.

Chess, as always, is a great source for an analogy. Everybody is well aware of Deep Blue. It was a beast able to calculate in excess of 200,000,000 positions per second. For comparison the arguably strongest modern personal chess engine, Houdini, calculates about 2,000,000 positions per second running on my laptop. But it is a substantially stronger engine than Deep Blue even with only calculating 1% as many positions. It is much 'smarter'.

It seems by severely limiting the hardware available as well as increasing the search space well beyond average (by setting the stack depth to 200bb) they are making a concerted effort at encouraging more 'intelligent' solutions over solutions that are more about trying to search as much of the game space as possible. The winning solutions will almost undoubtedly be substantially weaker than the optimal poker AI available today, but I think it's also much more interesting from an AI point of view.