Yes there is an annual tournament with the University of Alberta group(http://poker.cs.ualberta.ca/) who is recognized as pretty much the leaders in the field. That said they only play limit I believe since their NL bots aren't yet near good enough to be profitable against the top pros.
Yes, the alberta HU bot plays limit. No limit is a vastly more complex game and the state-of-the-art in HU poker bots couldn't beat a competent low-stakes player.
I would think a million dollars worth of research would be enough to duplicate what the University of Alberta group has achieved. Then you could hire some low-wage people to sit and play limit-holdem online with your unbeatable bot.
I'd be kind of surprised if an elite poker player hasn't already done this.
No way. The members page is currently broken, but http://web.archive.org/web/20080805024322/http://poker.cs.ua... says they have four profs, two consultants, an adjunct researcher, 5 graduate students, and 3 programmer analysts. I can only speak to the costs at the University of Michigan, which I attend, but let's assume they're similar at Alberta. Professors' salaries are around $100k and graduate students cost ~~$50k-$60k before candidacy and probably ~$25k-$30k after. That's at least a half-million dollars per year for the professors and graduate students alone; add the programmer analysts and the consultants and you'll be hard-pressed to get much more than a year of work on the project for your million bucks.
I guess I was thinking that because the research has already been done and they have released the general ideas if not anything resembling an implementation, someone else could duplicate the research faster and cheaper.
If not one million, the same order of magnitude. Conservatively $4MM to do it in two years?
Yeah, I think it's possible (although unlikely) for someone to be able to privately build an AI that can compete at the pro level. As you say, they would have access to all the papers published by AI researchers.
There is always a chance that someone has come up with a new approach. For example, the introduction of Monte Carlo search trees http://senseis.xmp.net/?MonteCarlo to computer was hugely successful. An AI on a 9x9 board can now compete at the Dan level. Obviously that's a long way from top-level 19x19 strength but it does show there is always room for a completely new approach.
