Ok, the video can now be found in 5 parts at: (YouTube has a 10 minute limit per video so I had to break it in 5 clips.)

http://www.youtube.com/watch?v=-GIrELC0Agw
http://www.youtube.com/watch?v=PwrHLYanlWs
http://www.youtube.com/watch?v=HxfevYbY8AM
http://www.youtube.com/watch?v=Mq4Kn3fDF8I
http://www.youtube.com/watch?v=g2QFuZQDi-0

It turned out to be surprisingly difficult to try to explain the reasoning while playing. Part of the reason was probably that I hadn’t looked at my tables for ~36 hours before recording the video (with the exception of having played any preflop-hands to the flop). I think it turned out OK though, and it does give a pretty good and honest idea of my style and the reasons behind my moves.

It could have had even more variety in the types of hands that came up, but maybe that’s a good enough reason to record another video later that would cover the kind of strategic situations that weren’t seen in this one.

Anyways, please let me know what you think and give me any feedback you might have. I would like to know if these kind of videos are helpful to anyone, or if you disagree with my plays I am more than happy to hear the critique and discuss any alternative ways to play the same hands.

(And if anyone else if up for it, it would cool to see a screen capture video of your play too and hear what thought processes go into your decisions.)

Thanks for the feedback, Mixu!

I’m glad you’re thinking of doing a video too. You’re a tough opponent so it would be really interesting see/hear how you think of your plays.

And btw, I really need to import my hands to Holdem Manager soon too. I’m not very familiar with that software yet, but I believe it will have bunch of new stats about my play that I can then post here, just in case anyone wants to compare (or to see what weaknesses those stats will reveal about my game).

Hi Teppo, I watched all your videos yesterday and found them really interesting. They are so different from other poker training videos I have seen which have all concentrated on the idea of exploitation.

I have some questions about ranges, which seem to be implicit in your thought process rather than explicit. I assume that playing optimally you can imaginine that your opponent is also playing optimally, and thus you evaluate your relative hand strength on a given board versus his range given his actions (assuming he’s following a GTO strategy). It looks from the video like you are primarily evaluating your own relative hand strength for a given board and then deciding based upon the level of opponent aggression over number of streets etc the likelihood you are ahead/behind etc (i.e. does your hand need to be in the top x% to continue, top y% to raise etc), but that might not be the case. I guess this is probably equivalent to defining ranges for your opponent and then narrowing them based on actions etc, although there might be some material differences, for example (this is prob. too simple an example) if I’m button and raise and get called and the flop comes AK3, then its more likely a great flop for me – despite the fact that the BB range overall (preflop) is stronger, he has many of his A,K in the 3-bet part of his range, so when he doesnt I know that there is actually less chance he has an A – how does this relate to your way of playing? Is it that the ‘optimal’ approach is a far more mixed strategy than my simple reasoning takes account of, i.e. although he is 3betting AK preflop a lot its nowhere near 100%, so that although I’m often right, the huge pots he wins off me occasionally when I’m wrong compensate.

Since your approach works pretty well I guess I’d like to know what your estimates of the ‘GTO’ ranges are on different boards, or at least how your hand-strength thinking goes in more detail (since this implies ranges). There must be a lot more balancing of your play than was apparent from the videos you mentioned it a couple of times (so do you have that all worked out in advance or estimate it as you go).

Anyway – interesting videos – they got me thinking

Thanks for the feedback, MOU5ER! Seems like your observations on my play and way of thinking are pretty much right on.

I try not to care how my opponents play, and focus primarily on keeping my own hand ranges balanced enough that there wouldn’t be too obvious or too exploitable holes in my play. Unfortunately that is relatively easy only during the very  early rounds of play (e.g. the opening preflop action) or later too if the pot is small.

Later in the hand I will usually start interpreting my opponent’s actions more and try to understand what range of hands those actions would imply if they were trying to play game theory optimally as well. So even then my "hand reading" is very generic and not opponent specific, although the further the hand goes and the bigger the pot gets, the less confident I am about my abilities to keep my play unexploitable, and therefore I cannot help but let my reads about my opponents play some role in the final decision, especially if I suspect that they might have figured out how to exploit my play.

In small pots I usually don’t bother thinking about my opponent’s possible hand ranges much, and that’s mostly because I’ve played such hands more often and already have the feel for what kind of hand ranges they should be holding if playing optimally, and therefore it is accurate to say that I am implicitly reading their hands too.

(I haven’t watched my videos in a long time, so I hope what I wrote above doesn’t contradict with anything I did on the video. I know the way I think about the game hasn’t changed, but in practice I might have approached the decision making process slightly differently back when I recorded them.)


I’m not sure if I understood exactly your question regarding the AK3 example, but here are a few points that would go into my play in such scenario. (I’d actually generalize my commentary to apply to many different flops, so I won’t focus specifically on the AK3 flop, but all similar enough scenarios.)

So, the first thing I recognize after such a preflop line, just like you said, is that the button’s hand range has more very strong hands than BB’s hand range, although BB’s range still has to have some very strong hands too as he should mix up his play enough so there’s always some chance of hitting a monster no matter what the flop is.

Depending on the flop it can even be true that button is more likely to have a very strong hand, and therefore more likely to want to play for stacks, while BB could have a stronger average hand strength and therefore wanting to see the showdown more often (but without wanting to commit all of his chips).

This creates the dynamics for the rest of hand so that the button wants to play more big pots (wheter he flops something good or is just bluffing and continues to bluff for more streets than just the standard c-bet). BB on the other hand often tries to keep the pot small via passive play, and also has to play some of his rare but strong hands passively too so that it doesn’t become too obvious when he has a monster and when he doesn’t.

In practice this means that if I was BB in your example, I would always check the flop not matter what I hit, since that is the best way to start executing the plan of controlling the size of the pot in a case when I know my opponent is going to want to play bigger pots more often than I do. Following the same logic button should bet the flop more often he would have after a different preflop line, as that is the most effective way of executing a plan where you want to play a big pot more often than your opponent (or when you want to appear as if you wanted to play a big pot while in reality you are bluffing and want to maximize your fold-equity against an opponent who you know does not want to play big pots as often as you do).

P.S. I can break my HEM stats to SB and BB once I have a little more free time again. (Don’t have them in hand right now without exporting from CP, importing to HEM, etc.)

Thanks Teppo – very informative reply as ever, apologies for making such lengthy posts! But here goes another one!

What I meant by GTO ranges is simply the ‘assumed’ or implied ranges that your opponent has in order to justify your play. In theory we should be able to ‘guess’ what a GTO players range looks like on a particular board/acton sequence, and since its how you play anyway you might be better at making these ‘guesstimates’ than most. Obviously these are also your ranges in different situations. In exploititative play opponent range is key, you might make plays that differ hugely from what a simple hand-strength thought process would suggest, in GTO play I guess thats less likely, but I imagine that the ‘possible range’ on a board and the ‘likely range’ could still be very far apart.

I havent been trying to play anything like a GTO strategy myself, but I’ve been trying to figure out what one might look like using StoxEV – and its pretty hard (obviously LOL!). StoxEV is useful, but only evaluates conditional EV, it doesnt do much for devising strategy except in very simple/special situations. I’ve started with a generic pre-flop approach and trying to figure out how my ‘fledgling GTO guess’ fares against various rather simple counterstrategies, like ‘always calls’ and ‘only the nuts’, e.g. ‘always calls’ obviously sets upper limits on the bluffing/semi-bluffing frequency.

The other idea I had was to start from the low-blind case which has already been solved (i.e. 10bb stacks) and try to ‘guess’ (I’m afraid) which hands to add once you get to say 25bb where the game is a ‘two-bullet’ game rather than single-bullet jam/fold. At these stacks it appears that you can add a range of hands with the idea that its OK to check-fold some on the flop and shove others. In this game its definitely not OK for the SB to open-raise everything (because then stack-to-pot-ratio is again small enough that the BB can ‘reverse’ jam/fold him to death) – at first I thought this might be relevant to John Deers thread, but probably not since they are talking about deeper stacks there. The plan then was to move on towards 50bb stacks etc and see if these two approachs can meet in the middle (….not a chance…LOL), but I’m rapidly coming to the conclusion that its far too hard to do ‘by hand’. It might be poissible to get a sensible estimate with better analytical software though. Do you think this approach would sit well with your CFRM algorithm, since you could seed the 25bb case with the known 10bb solution [so your initial buckets would be pretty good] and the two-street nature of this game would also help simplify things, and then work upwards in blinds (be really useful for HUSNGs too!).

Roughly how much work was it for you to implement CFRM in C (I assume thats what you wrote it in)? I was considering giving it a go, but I’m still pondering how best to implement some bits of it. I tend to procrastinate a lot before writing any code 

PS have you thought about CUDA [for GPU acceleration] I have no idea if the algorithm is amenable to that?

MOU5ER: I think I understand what you meant now… The way I think of that is actually not by trying to come up with an exact distribution my opponent might hold, because I know I don’t know GTO strategies nearly well enough to do that, and therefore I read hands in a much “fuzzier” way than how I would play those situations myself.

In practice what I do is reduce the implied hand ranges from the size of the pot (thinking of it in terms of “standard bets” that to me are somewhere between half the pot and full pot), who did the betting (knowing that a player who bets is more likely to have a monster as well as more likely to have air compared to the player who calls), and also in what situations did most of the money go in (so that I recognize that the hand distributions are different if my opponent checks the flop and bets the pot on the turn after a flush card hits, compared to when he bets the pot on the flop and then checks the turn after the flush card hits, even though in both cases the river could be identical in terms of my own hand, board cards, and the size of the pot), and then I have certain relatively simple rules of thumb that I’ve come up with experience to remind me how big of a pot I want to play with each hand.

For example, if I flop a top pair in pot that was raised but not re-raised preflop, I’m usually happy to call standard sized bets on all 3 streets and see the showdown (or bet for value myself and try to get at least 2 bets in if my opponent starts checking), but I would usually not play it for 4 post-flop bets (unless they were small bets like half the pot or less). Though like I said it also depends on how much of that money goes in when the board is safe vs. dangerous to me, and of course my exact hand so flopping a TPTK w/ AK is much stronger that some other top pair, so sometimes I fold top pairs to a 3rd standard sized bet, and sometimes I play it for 4 bets. And if flop 2nd pair I may be calling only 2 bets, plus or minus one bet depending on the other factors mentioned.

So, I have relatively simple rules of thumb that I then adjust usually no more than plus or minus one bet, and that’s where my implicit hand reading comes from rather than from reading the hands explicitly. I do that because over-reading hands (relative to GTO strategy) is an exploitive strategy that may work if your reads are correct, but can also backfire if you are wrong and get outplayed, and personally I prefer to under-read than risk over-reading someone. And that’s why I also cannot assume that my opponents play exactly like I do, but can only assume that they play approximately similar sized pots with the same range of hands that I do, even if they get the money in using different betting lines than I do (for example they might re-raise more preflop, check-raise more on the flop, and so on, but if they make big enough mistakes by playing pots that are either way too big or way too small for their hand strength, my strategy is hopefully balanced enough to profit from both of those mistakes).

Roughly how much work was it for you to implement CFRM in C (I assume thats what you wrote it in)? I was considering giving it a go, but I’m still pondering how best to implement some bits of it. I tend to procrastinate a lot before writing any code 

I wrote it in C++ and it was actually surprisingly easy to do thanks to the awesome paper from the University of Alberta guys (I’d have to look it up to remember which exact paper it was) that described the method in a very clear way. Later I also found an open-source implementation of the algorithm that I think also showed how to apply it to real world Texas Holdem, so that might be helpful to you as well, even though I was not able to understand how that code worked at all so I ignored it myself. (My background is in mathematics and not in programming, so that’s how it usually goes for me.)

PS have you thought about CUDA [for GPU acceleration] I have no idea if the algorithm is amenable to that?

I have played around with that idea, but I’m not an expert in the topic nor do I have enough time to even seriously consider trying that at this time. I think the one obstacle I ran into when researching it was that my algorithm is recursive, and the current technology didn’t support running recursive functions on GPU’s (I may be wrong about that though). Either way I’ll have to wait until I have more free time until I can consider taking any advantage of the GPU technology in my algorithms. I do find it extremely fascinating concept though and hope to be able to start experimenting with it at some point.

Thats interesting, in more difficult situations, for example where a draw hits and your opponent bets do you try to figure out how frequently they *should* be bluffing in order to decide whether to call (I find this a tricky situation myself)?

Not exactly. I base this kind of decision on the size of the bet, and the estimated difference in our hand ranges.

For example, if a draw hits on the river and my opponent bets the pot, in a case where I thought our hand ranges were somewhat identical (and especially that we’d both have the same draws in our hand ranges), then I would try to call with exactly the top 50% of all the hands that were in my hand range and fold the rest (I’m ignoring the raise option for simplicity).

It gets more complicated when our hand ranges are significantly different. For example if I believe the opponent was much more likely to have the draw than I am (which I try to avoid though), then I would have no choice but to fold more than 50% of my hands and let my opponent profit from his bluffs. And on the other hand if I was more likely to have the draw and therefore had a stronger hand range, I would call more than 50% of my hands.

I can’t really explain it in more detail than that because I have no simple rules of thumb for those, and play more by my guts.

I’ve got the UofA paper with the pseudocode in it, so thats probably the one you mean. Where did you find the open source code, I had a quick look, but I couldnt get google to find it.

Unfortunately I cannot remember and I didn’t save the link. I might have found it on the pokerai.org forums, but I’m not sure.