Episode 79: The Computer Poker Research Group

On Episode #79 of the Thinking Poker Podcast, Nate and Andrew offer a warm welcome to newcomers of both the podcast and the iBus Media Network, Andrew gets leveled by an eight-year-old, and the pair breaks down a hand from an online MTT. They are then joined by Dr. Michael Bowling and PhD candidate Mike Johanson from the University of Alberta Computer Poker Research Group to discuss their ongoing efforts to build poker bots, the difficulties that come with constructing a decision tree in poker, the hidden complexities of bet sizes, and much more.

Timestamps

  • 0:00 About the Thinking Poker Podcast
  • 4:38 Teaching kids to play poker
  • 18:02 Strategy: Turn a straight, fold the river?
  • 36:23 Interview: Dr. Michael Bowling and Michael Johanson of the University of Alberta Computer Poker Research Group

Find Us at PokerNews!

The Thinking Poker Podcast won’t be available at this location much longer. Soon, the only place to hear us will be on the PokerNews podcast feed. We’ll still post announcements here when there’s a new episode available, and we’ll keep the current feed active for a few weeks to give everyone a chance to migrate, but please save yourself the trouble and subscribe now via iTunes or your favorite feed reader.

16 thoughts on “Episode 79: The Computer Poker Research Group

  1. Is there a way to download the podcast from PokerNews?

    Once you guys stop posting the podcast here, will there be a new place for us to go for comments? This post show discussion is very important. Even if you dont post the podcast here in the future, I hope you will post an announcement that it’s out and link to it so that we can continue out discussion.

    While listening to the strategy section, I agreed with everything that was said and would play it almost exactly like Hero did. So imagine my surprise when I ran the hand through Scenarios in PokerSnowie and it advised us to play the hand differently on almost every street.

    It advised us to fold to the flop bet, with raising small being a slight error and calling being a massive error.

    Then it says the turn should be bet 1/2 pot and, although still positive obviously, the EV drops as the turn bet gets bigger.

    Then it says that the river is a slam dunk call.

    I wish there was a way to get an explanation of the data, but this all seems wrong to me. This is one of the biggest challenges I run into while trying to learn poker from many different sources. I never know how to reconcile conflicting pieces of advice.

    • If you’re looking at http://www.pokernews.com/podcast/, there’s a big “download” button next to the episode!

      Piefarmer asked the same question about comments last week, so you can see my response to him. The short answer is that my plan is to continue posting an announcement here when there’s a new episode available on PN, so there will still be an opportunity to comment here.

    • I trust PokerSnowie less in multiway pots than I do in heads up pots. I wonder sometimes whether the human NLHE community is just that far from optimal in multiway pots, or if it’s just much harder for PokerSnowie to get convergent results. The U of A guys do indicate that they are much less close to building an AI that can play well multiway. So probably it’s a bit of both.

      Our analysis relies on a lot of exploitive assumptions about the Villain, and I think that among other things Villain’s flop bet is extremely difficult to balance and probably just a bad idea with any hand. PS suggests he bet just 2% of his range there, and it’s all bluffs – I don’t even know what to make of that, it seems bad to me, even though some of the bluffs are extremely strong (Js Ts).

      Regarding the turn sizing, PS thinks that Hero needs a very strong hand to call the flop bet (I agree with this) and consequently he has to bet small on the turn because he doesn’t have enough bluff candidates to balance a large bet. Take a look at SB’s turn calling range, according to PS.

      • This is admittedly lazy, as I can probably find it on PokerSnowie, but does the fact this is a tournament also explain part of the difference? At the very least, does stack depth matter? In this hand, Hero starts about 75bb deep. How does this compare with what PS has analyzed?

        Also, might I suggest a new strategy segment, a la Tony Dunst’s Raw Deal, where Andrew’s cousin (or newphew) breaks down a hand. Very cool segment AB.

        • Structurally, everything was exactly the same in PS. The big difference is that PS plays against other PS ranges. Since PS is generally tight from the SB, it would have a much stronger range for donking out on that flop than your average human player would. This is probably why PS suggests that hero folds to the flop bet. Also, PS plays way tighter in multiway pots. I’d imagine it would at least call if the third player was not involved.

          A side note to that point is that the researchers from this episode also talked about how difficult multiway pots are due them not being a zero sum game. This also leads me to believe that we humans play multiway pots way wrong. My guess is it’s probably best to avoid this at all unless you have position or an overly good price to do so. Even then I’m guessing you should probably tread extremely lightly post flop.

          For this reason, I think I like the idea of a raise-fold in the hand in question with the goal to get the odd man out and define the donker’s range. Usually we dont want to raise good draws because we dont want to get bet off of them, but maybe being OOP to a third player in a multiway pot on a paired board with a possible flush draw present, our straight draw isnt really worth as much.

          • Don’t forget that even HU pots are part of a multiplayer game. You might argue that, if both players know the ranges with which they reach the flop, it becomes a two player zero sum game, but those ranges are determined as part of the multiplayer preflop game. For example, you can’t treat the game ‘UTG raises, only BB calls’ as a two player zero sum game, because UTG has to choose his opening range based on the likely actions of all the remaining players, which is why opening ranges get wider as position gets later. The only true two player zero sum game occurs when it folds round to SB. I’d expect PokerSnowie would play that close to optimally, betsizing notwithstanding.

    • I don’t think it is. I noticed that when I click “Play” on the PN player, it shows 39:18 as the length of the show. However, it doesn’t seem to display an hours digit, and I believe the show is in fact 1:39:18 in length, so I think the hour is just getting dropped from that display. I clicked through the file that’s actually on there, and I don’t think they’ve taken anything out, certainly not an hour’s worth of material. Thanks for bringing this to my attention, though.

      • As far as I can tell the different segments of the show each reset the timer to zero. It’s a strange bug, or possibly it’s some sort of feature.

  2. Loved the interview. Confirmed everything I keep saying about multiplayer games and ‘GTO’ strategy. It was gratifying to hear that they don’t have much clue about multipayer game theory either, and I’m encouraged to keep going with my work on mutliplayer games.

  3. I’m sure that the CPRG guys could get their bot backed for heads-up NL matches against whatever competition they think it is ready for! Serious.

  4. As expected, I found this episode fascinating. The discussion of the min-betting strategy was eye-opening, as I’ve seen clips on YouTube where various bots make minbets in a manner that is completely contrary to what “strong human players” would do. (See, for example, https://www.youtube.com/watch?v=PlfEfKu5tZA . I believe hyperborean was the bot made by the Alberta team.).
    The “weirdness” of multiway pots was also illuminated nicely by your guests. Snowie frequently marks “standard” calls I’ve made multiway as blunders. For example, if I raise JJ in MP and get called by the button and big blind, Snowie often wants me to fold my overpair to a donkbet on a 9-high board. Now I’m kind of understanding that the fold *may* be correct in a game theory sense, since (if I understood your guests correctly) calling will actually benefit the player to my left (the button in this example) more than it benefits me. I guess this tallies with the well known idea that your range needs to be stronger (and have more chance of making the nuts) when you’re in a multiway pot. It’s really hard to fold an overpair to a single bet though. 😉

    Anyway, I loved this episode. Thumbs up to all involved.

  5. Hey Andrew and Nate,
    Great show. Fascinating stuff and you guys did a great job with the questions.

    I especially appreciated the content on the impact of non-optimal play on one’s opponents (with one opponent losing a lot and the other gaining a lot), and the follow up question on the possible role of this type of play in collusion. This concept really needs to be made more public as high stakes collusion when two players are chopping equity, or teams of bots that sit at the same table and use this strategy to destroy human players are real concerns. Nate said something about RayJ posting that he would do this to punish people who were bad for the game. I think this topic deserves a thread by Andrew on 2+2 with a cameo by the Alberta guys. The more people know about this the harder it is to pull it off.

    Anyway, Andrew can you explain in more detail what you and one of the Alberta team members meant when you said something about learning more about “what I should be thinking about at the poker table and not wanting to think about those things” (paraphrase). Do you mean feeling obligated to think very deeply about ranges/equity to play optimally, constantly focusing on the play of others when you’re not in the hand, etc.? Or was it something else?

Comments are closed.