## Preliminary Project Submissions

on April 12th, 2011 at 10:43 amPosted In: Announcements, Assignments

Remember that your preliminary bot submissions are due **Thursday, 14 April, 11:59pm**. You can submit it by posting a preliminary bot binary and PDF document describing your design and plans as a comment to this post.

The submission for the team including Ben R., Ben K. Samee, and Colin can be found at the following links

http://www.cs.virginia.edu/~bdr7fv/prelim_report.pdf

http://www.cs.virginia.edu/~bdr7fv/cardBlindBot

The submission for the Cheater team can be found at the following links:

Bot binary

Bot explanation

Documentation: http://www.cs.virginia.edu/~evans/poker/wp-content/uploads/2011/04/preliminary.pdf

Bot: http://www.cs.virginia.edu/~evans/poker/wp-content/uploads/2011/04/bot_e1.zip

Sounds like an interesting and ambitious strategy. I’m not sure I understand how you will combine CFR and MCTS, and it would be good to try a preliminary composition before going too far with the techniques separately to better understand how you will compose them usefully.

I like that you have a simple and straightforward strategy. The risk is that such a strategy could be easily exploited.

I’m not convinced the way you are calculating hand strength against 2 opponents makes much sense. It would make more sense to me to do a 3-player roll-out to evaluate the hand strength, but sampling the distribution if it is too much to do exhaustively. The other issue to consider is that one of the players may fold before the river. Then, you could do a 2-player hand strength estimation, but this probably overestimates the value of your hand against the opponent who stayed in.

I’m puzzled by your comments about the biased shuffling. The position should matter – the bot who acts last (dealer position) has an advantage, so should do better than average. In a 3-player game, the ordering among the 3 players also maters. The dealer cycles through the positions (moving the dealer button), so over a large number of hands, all that should matter is the relative ordering amount of hands. If you are finding something more than this, that would be interesting and does sound like a bug in the dealer (but it is necessary to confirm this more carefully).

Sounds like you have a good start on the bot.

Yes, I think the dealer program is already moving the dealer button for every match, so the position shouldn’t matter for a game of 3000 matches.

What do you mean by “relative ordering amount of hands”? Do you mean stacks?

Here is our plan regarding MCTS and CFR. First, we noted that CFR is a static strategy. We approximate an equilibrium strategy, and each move is simply a table lookup. MCTS, on the other hand, is a dynamic strategy that needs an opponent model.

When the game starts, we need to have some initial opponent model before the opponent has acted. We intend to build off of the implicit opponent model produced from CFR. Furthermore, we assume MCTS will perform poorly initially until we have a good opponent model. Therefore, we will use CFR as *our* strategy until our opponent model is suitably good, and then we can switch to MCTS.

What we are still looking for is a principled way to determine if an opponent model is “good.” We are also still deciding on what kind of opponent models are needed for MCTS. Currently, we are just planning on using something simple such as P(bet) P(fold) and P(raise) at each game step. Since a player has 4 rounds to make decisions(preflop, flop, turn, river), we just store 4*3 scalars as our opponent model. However, how many hands must we see before these probabilities are statistically significant? This question is complicated by the fact that our opponent is evolving over time as well(if they are not using a static strategy that is).

We are trying to limit the complexity of our project by initially choosing this simple opponent model; and then modifying it if it is inadequate. Furthermore, we hope to test our CFR results in a principled way, by seeing if our pre-flop predictions roughly agree with reasonable hand-tables from poker strategy guides.