Proposal for 6-player no-limit Texas Hold'em event at 2016 ACPC

[sganzfri]

Hi all,

  I'd like to propose a new event to start in 2016. It would be 6-player no-limit Texas Hold'em, 100 BBs deep. Here are some preliminary details and comments, following the template design outlined by Neil. Please feel free to comment, add clarifications, etc. Also, please reply if you are interested in participating in the event, and please send a link to this proposal to others whom you think would be interested and may not be on this list.

Rules:

100 big blinds. I don’t mind if they are 50-100 or 1-2; 50-100 would match the no-limit 2-player competition, but 1-2 might be a little easier for developing agents, and iscloser to more common human levels, so I suppose I will advocate 1-2. Then all agents would start with 200. 100 big blinds is most standard for humans. I think there is value in using the same variant as is typical for humans for several reasons. One is that, if we are successful, we can potentially receive significant media attention (and possibly funding by interested parties) for research/future competitions/workshops/etc. Second is that there would possibly be large amounts of human data available to learn from. Etc. There are already significant computational challenges requiring new techniques for going from 2-player to 6-player, and I think 100-bb variant will still be quite difficult computationally.

 

Interested competitors: please respond in this thread if interested. I believe at least four teams expressed interest already at the workshop.

Scientific merit:

It seems clear to me that significant algorithmic innovations would be needed to develop strong agents for 100 BB 6-player no-limit Texas Hold’em. My understanding is that the best 3-player agents either just run CFR, or fix preflop strategies and just run CFR for the sequences where two players make it to the flop. I think even to apply either of those approaches to 6-player would be highly nontrivial.

Theoretically, it would be very interesting to see how the equilibrium-based approaches can do, if CFR has interesting theoretical guarantees in the multi-player setting, if improvements can obtain theoretical guarantees, etc.

I also think there is room for significant innovations in terms of new abstraction/equilibrium algorithms, game decomposition, and real-time endgame solving beyond what is present for 2-player variants. I think Kevin’s functional regret minimization approach would be applicable, as we could potentially abstract away some information about what the specific history/players are in the hand. E.g., if utg+1 raises and cutoff calls, maybe that could get bucketed with situations where utg+1 raises and button calls. As far as I know, no teams currently use any sort of “player abstraction” for 3-player limit, which seems like an interesting new direction.

I also think that even doing the evaluation could lead to interesting research questions in tournament design/variance reduction that aren’t as relevant for 2-player. I imagine there will be 6+ entrants, but even if there were just 4 or 5, it would be interesting theoretically to determine the “optimal” way of performing the evaluation. E.g., which/how many permutations of the agents to use.

 

Resources:

I think a limit much lower than 7 seconds per hand (on average) would be ok. Perhaps 2 seconds per hand would be reasonable. Agents fold much more often preflop 6-handed than heads-up, and so there would be many fewer “difficult” postflop decisions to be made per agent.

 

Statistical significance:

I’m sure there is a lot of information available by searching forums such as 2+2 to determine variance. I am looking at a blog post now that says that “a fairly typical standard deviation” is 90 bb/100: http://www.nsdpoker.com/2011/03/nlhe_6m_pros/. Simple searches show that the SD for HU exceeds 100 bb/100 by comparison. I can ask some experts who have databases over large sample sizes for further information. But I think it’s fair to say that the variance is a bit lower for 6max than HU. Combining this with reducing the time per hand by 71% as described above could make it very realistic to obtain significance. There are questions like how often to permute the players, and whether you can integrate forms of duplicate scoring when doing so, etc. I think those are very interesting research questions. 

Win rates will depend heavily on the skill disparity, though they will likely be lower in general than for HU. I certainly expect the top agents will have an average winrate exceeding 10 bb/100, and maybe even 20.

-Sam

[Kevin Waugh]

In my mind, the two main concerns are how many competitors would enter, and how much computation is necessary to get statistical significance?

Regarding the first question, could past competitors please respond (or e-mail me directly) if they would be interested in submitting to a 4 to 6 player no-limit event?

I think discussion on the difference between 4 and 6 player would also be useful.  Jumping from 3 player limit even to just 3 player no-limit is in itself dramatic change in scale.

Regarding the latter question, would any competitors be interested in a test competition sometime in the fall to explore the variance of the game as played by bots?  If even 2 or 3 groups had very preliminary bots to play say 100,000 hands we should be able to get an idea of the variance and resource usage.

[pokerGeek]

Sounds great. I'm in.

Is there a protocol / code available for this new 6 player game ?

[Matej Moravčík]

Hi all,

We (the Nyx team from Prague) are almost definitely in, if there would be an six-max NL competition.
We have already some ideas for the AI, but we can't promise that we will be ready for the Autumn test.

Our suggestions for the competition:

1.) There should definitely be blinds 50/100 rather than 1/2 (if the betting is restricted to integers).
It has bigger scientific value because the optimization of the bet size would be possible and it would also allow to create continuous models of the game. It is also closer to the real games played.

2.) We don't mind any stack size but we would prefer 100+ big blinds for the players, since it would be near the size of real world problems.

3.) For the simplicity, there should be only bankroll competition

Matej and Martin

On Sun, Mar 15, 2015 at 10:12 PM, pokerGeek <wilson.da...@gmail.com> wrote:

Sounds great. I'm in.

Is there a protocol / code available for this new 6 player game ?

--
You received this message because you are subscribed to the Google Groups "Annual Computer Poker Competition" group.
To unsubscribe from this group and stop receiving emails from it, send an email to computer-poke...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/computer-poker/7310116c-e297-482d-8e53-8046272542a7%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

[Noam Brown]

The CMU team would also participate in a 4 or 6 player no-limit event. I would prefer 6 player, and it seems Sam would too. I would prefer 50/100 over 1/2.

[sganzfri]

Just a random comment -- someone recently pointed out to me, completely independently of this thread, that the "player abstraction" problem I mentioned has been studied before, though not for poker, e.g., http://web.eecs.umich.edu/srg/?page_id=220, http://web.eecs.umich.edu/srg/?page_id=1278).