Is anybody beating KGB's Dungeon level? Or should we be?
Even with Peek At End, I still can't seem to beat it consistently. I thought I had it with some heavy 3-betting, but it may have adjusted.
I am following the different "personalities" and playing styles of the AI's, but, really, it comes down to them being aggressive enough that I have to risk half my stack if I want to start reraising at all. I tried playing tight aggressive, but I seemed to just lose slowly.
It looks like you're new here. If you want to get involved, click one of these buttons!
I have won the last few times that I have played, but I have not yet played enough hands to feel like it would be statistically valid. From my small sample size it appears that the players are more selective in their bluffs, have a better sense of when to slow-play as well as trying the occasional slow-play bluff. That combination does make the game tougher.
I found your question interesting, so I ran some queries on our database. I threw out all sessions in which the user had frozen the button, or fixed which hands to deal (to prevent always dealing pocket Aces etc). Next I only considered users who had played at least 10,000 hands at KGB's dungeon.
The result: only 60 members are currently ahead, long-term, at KGB's dungeon. 59 if you don't count me...I wrote the software, a little unfair advantage.
At the top of the heap, one member (not me) has played 820,262 hands at KGB and won 81,202 big blinds. Nearly one big blind per 10 hands, quite a feat.
But the vast majority are losing to the bots at KGB. If I had to give out one secret to beating them, it would be to remember that they have no fear. You have to be tenacious and call them down with Ace-high sometimes, because they will try to pick up the pot if you show any sign of not wanting it.
You also have to let go of some of your biases. Take for example, a check-raise bluff on the river in a big pot. Against most low and medium stakes human opponents, that isn't even part of their arsenal. For one thing, it adds to their variance which is a problem if they're on a limited bankroll. They also look stupid when they get called, and nobody likes to look stupid. So most players aren't going to try a big river check-raise bluff. But of course the bots don't give a crap and will do it whenever their random number generator tells them to. So sometimes you have to call them even when your instincts are telling you, "That couldn't possibly be a bluff".
In the long run, it's really important to let go of your biases, because someday when you're playing high stakes against Tom Dwan, you can be sure the big check-raise bluff on the river IS part of his arsenal!
Thanks for the question.
Steve - funny you posted this because I was in the process of running the same types of queries.
think - it's interesting that you posted this because Steve had commented to me a few days ago how difficult it was to beat these bots in KGB. One additional piece of advice - you cannot be consistent in your play. They use a lot of analysis of your statistics and previous play, much more than at the easier levels. If you become predictable, you will get crushed. Switching up my usual tendencies has helped me play much better against them, and I'm sure that carries over to reality. Even though your normal opponents don't typically have the same memory that the computer has, they do have enough memory to notice obvious patterns in your game. So playing against these bots should help a lot in getting better at hiding your tendencies.
I have played 22K+ hands at KGB, currently ahead 22.42BB per hundred hands.
Play tighter than you are used to doing up front, do not be afraid to three bet CO and Button openers with any reasonable holding.
If they call a smallish flop bet, they will fold to a largish turn bet a lot of the time.
They will often fold to very large river bets.
General, thanks for the feedback. This is good to know. I've heard and noticed the river bet leak before, but I've never noticed the small/big flop/turn one. We're always working on improving them, so this gives us some more info to help make them better.
I switched strategies to calling a lot of blind steals and then check/calling the flop and saving the barreling (bluff, usually) for the turn (pot-level if I lead, or 3X if they lead). I just had two sessions where I got to over 500k. Then all of a sudden, I could feel that the whole table adjusted to the move. It's not "collusion," but it's like I'm playing 8 people who are sharing notes on my play.
And I don't say that like it's a bad thing -- I suppose there is an element of training oneself to "mix it up." Now I can tighten up and figure if I hit, they're going to bet into me and not be so ready to fold to medium raises.
And it is enjoyable to try to beat the game. But at some point, I start to read into the "tea leaves" of "how can I take advantage of the AI," which begins (?) to diverge from poker strategy or psychology. Or maybe it doesn't -- to Steve's point, maybe these moves make sense, they just don't make sense to "me" right now.
Are the AI adjustments lifetime-cumulative and individual (i.e.: do my AI opponents bring in the 'think' dossier when "playing" me) or on a session-to-session basis?
On another note, I am curious about the risk/return of the check-raise river bluff, especially in live 1/2 against a mix of opponents (as far as skill level and confidence/courage). Steve is 100% right about variance and risk being a negative for most players (the occasional player just loves the "big gamble," but I would say that they're the exception by far in poker -- you can find them by the roulette wheel). And this would be doubly true in live casino (as in, not home) play. Online, one could take a break and surf the net, do "home things," etc. A lot of your live opponents are playing while their significant other is playing slots, or having a "spa day," or whatever. Maybe they're solid players with a good grasp of EV concepts but who can't take that 60/40 bet because then there's a 40% chance that they will be stranded at the bar for two hours.
So there would be something to be said at first look for "variance equity," or really, "anti-variance fear equity."
If you've never tried advantage play at blackjack, you'd be surprised at the variance swings. It makes no limit look positively "slow and steady." A bankroll for $10 tables has to cover a $100-120 "big bet," and the bankroll should be way over 100 big bets -- so you're talking $10k to play $10 tables (at a 6 or 8 deck game). And, by the way, the swings won't make any sense, and they tend to be monstrous. A bankroll of this level might have a 5% ROR (risk of ruin) assuming perfect play, and if you try to bankroll a little less (say, 80 big bets), your ROR goes WAY up. That means your "statistically expected nadir," or whatever, represents losing much more than half of your entire bankroll.
At least in poker, if you got drawn out on on an all-in, you know that you had the best of it when you bet your money. So it may be painful on a psychological level (and hopefully won't put you on tilt or make you try to chase your losses), but it shouldn't rattle your confidence as a player. In blackjack, a 2%-4% EV advantage going into a given hand is a monster advantage. And statistically, playing basic strategy against a freshly shuffled deck, you will win 43 times, lose 47, and push 9 (decimals being dropped). Natural 21's (paying 1.5 to 1), and advantageous splits and double downs make up for this and bring the EV to -0.5% or so with basic strategy (the strategy on those little cards they sell at the casino gift shops -- funny how many times I've heard people say, "casinos print those so people will buy them and lose money!"). You could play the strategy on that card at a $10 table flat betting the minimum (which represents "perfect non-advantage play") and suck up free drinks and decent comps, casually and sharing company/conversation (and victories, when the dealer busts and pays off the table) with people who AREN'T out to take your stack (I know that's part of the fun of poker, but it does get awkward sometimes...), with an EV of somewhere around -$10 per hour (do NOT drink while pursuing advantage play!). And in Vegas, you can lay the card right on the table -- in AC, you can use it and share it, but you can't keep in "on" the table. I can't identify with casual vacation/tourist gamblers who mindlessly shove $20 and $100 bills into some video game with some "theme pictures" on it ("Medieval Princesses -- now with 3x Progressive Bonus and Wild Card round!") just so they can smoke their way through a pack of Pall Malls and hit some buttons. When I retire in 2060 or so, you can find me at the blackjack table, believe that...
And, by the way, the comps are at least a little substantial -- not this "I played poker all day, so I have enough in my Rewards account for a coffee and maybe a candy bar...or at least the candy bar" stuff.
Anyway, in blackjack, you are still having 10-15% more "losing" experiences than "winning" experiences. And when the deck is "hot," the dealer has a better chance of getting those big cards, too. There's nothing like putting down a big bet, seeing the dealer spike an ace upcard, and knowing that the insurance bet is an advantage bet (here) -- so you lay out another half a big bet...oops, no ten in the hole, there goes your insurance bet...and with the ace, the dealer will have two shots at getting a "made" hand, if they don't have one already...how is that K-8 "18" looking now against that ace? It's gonna lose you the equivalent of a whole 1/2 buy-in in about 20 seconds...
This is why I react a certain way when people bring up the idea of pair vs. two overcards (55/45) in a certain way ("coin flip"). A 10% advantage is a HUGE advantage -- it's one thing in a tournament, but I would take that bet anytime in a cash game. And there are 16 ways to make AK but 6 ways to make AA and 6 ways to make KK -- starting pairs just come about less often, period (although I realize 7-7 is more playable than A7 and K7).
I know that was a bit of a tangent. But where I am going with all of this is along the lines of: If there was a strategy which would help me leverage my potential "anti-variance fear equity" against intermediate-to-advanced 1/2 players by using large river bluffs or early-in-hand "non-game theory optimal" all-in bets to win one more buy-in than it lost me, I would be all about it.
The AI reacts to your all-time tendencies as well as your in-session tendencies.
A lot of the things you talk about in this post are just things to experiment with at the live tables. The bots are heartless and have no feelings. When people play with fake money, they also have fewer feelings and biases. It's the reason why people who run experiments on risk aversion, etc. always have monetary incentives for their participants. You really don't know how people are going to react when money is on the line. If we could rely on the data in the game, we'd be able to learn a lot about how people react in different situations with well over 100MM hands of data. However, it isn't real money involved, so players are more likely to make these river check-raises and there's really no way to know how an actual human with money on the table would react to it.
This concept of risk aversion that you talk about is not a stable trait. It depends a ton on whether your opponent is in a gain or a loss frame. People in loss frames (meaning, they see themselves as being down) are much more risk-seeking than people who see themselves as being in a gain frame. They value additional losses less than they value gains. It is possible that the opponent who just wants to stay at the table views himself as being in a gain frame - just being at the table is winning. That person is much more likely to fold to your c/r bluff than the person who sees his bet that your are raising as a loss, and consequently overvalues gains. That person wants to win his bet back and is more risk seeking and likely to call you down.
I definitely think a person who is not at all risk-averse can take advantage of others at the table. But like you said, you have to be willing to lose everything. You can learn a lot about opponent ranges, equity and the value of changing styles by playing KGB, but some tweaks need to be learned at the tables. Although wouldn't it be cool to simulate risk aversion and behavioral biases in the bots? It would be very hard to do well, but it is fun to think about. There's something called a Holt-Laury risk preference measure. It would be interesting to take the typical distribution of human poker players (although I have no idea what that would be - probably different from the known overall human distribution) and assign those values randomly to the bots and have them weigh in risk preferences to the EV. The problem is that would make them systematically worse players, especially since the human players would be unlikely to experience much risk aversion in a non-cash game.
I think the approach would be to come up with a template play that utilizes the strategy in an optimal way, and then hold real-world experiments. You can stake me if you like!
But seriously, what would be the right setup? Let's say you're going to take some low or medium SC's and rep it all the way down like it's high pair or big slick or something. It's already interesting as a "spec" hand, and there's a 20-25% chance that you'll hit 2 pair, trips, or an OESD or FD (please check these #'s), turning it into a very playable hand, with the added benefit of camouflage (and also with the added benefit that, assuming your opponents' holdings are biased toward high cards, then, by definition, if you hit the flop with your low/medium cards, there is less of a chance that they did). So, anyway, 20-25% of the time we have +EV from the flop onward, with the potential for some concealed monsters, which are the bread-and-butter of deep stack play. That is pretty interesting right there, and it's why Doyle says 67s is one of his favorite hands (in Super/System).
OK, but what about the 75+% of the time that we don't hit (or maybe catch low pair)? Note that I considered OESD or FD playable but not pairing up (OK, maybe if it was top pair on a low board)? Fire off the 60-75%% c-bet -- there's a 70% chance that they didn't hit the flop, and, even if they did, they might have a medium-strength holding that they're OK with folding to a 3x pre flop raiser. Of course, they might have a pocket pair, but what percentage of their range is pocket pairs, and how likely is it that they hit their set? If they didn't, and there are over cards on the board, you have some more fold equity. Of course, if they are "poker aware" at all, they know what a c-bet is. So you haven't totally narrowed their possible holdings too much. So the c-bet here, though not so much saying, "I hit the overpair," at least says, "don't you go thinking that I didn't hit the overpair!"
So, assuming a call, we are at ~20 BB going into the turn.
So we're still setting up the river bluff. My problem is what to do with the turn. Bet (what amount)? Check/call? Check/raise? Check fold and go back to ABC poker?
Let's just say you barrel the turn again (and get called). The pot just doubled and then some.
Now your river bluff pretty much has to be the rest of your stack (if it's a 100 BB buy-in), or it's not much of a bluff.
I guess if you go 3x/50%/50%, you'd be going to the river with pot ~30 BB. But does this give you fold equity along the way or do a good enough job of representing strength to set up the river?
Of course, then, you could do a short buy-in of 60-70 BB or so, which would mitigate risk a bit (and prevent a worst-case scenario river reraise).
So what, in essence are you betting on? If you were trying to check-raise the river and not just bet straight away, you would need 100+ BB to do it meaningfully. But, for instance, do you even count the 20-25% of the time that you flop a playable hand in your EV calculation? How much of the time will you get reraised out, maybe on the flop? How often is THAT a bluff on their part?
I mean, anyone pulling this move regularly would be spewing chips. But I am talking about your basic conservative TAG player trying it maybe once a session. And let's say you're laying ~2 to 1 against counting everything from pre flop to river (since, if they fold the river, they don't risk the call, but you risk the river bet either way). How much does your flop/turn fold equity and chance of flopping a playable hand make up for this? Does it get us to 60/40 or 50/50, and then, is there a >60% chance that they will fold to our river bet, either from fear equity or the fact that they just really think you're telling them that their non-nut hand will be second-best? If I can win one more dollar with this tactic than I lose, I would pretty much do it once a session. I think the impact on table image would be positive, too.
By the way, in the 1/2 live games I played, 3x pre flop is pretty much an invitation for a multi-way pot. You'd pretty much have to go $10-15 (5x +) to get people to fold out. And that definitely affects pot size, which tends to increase geometrically. Suddenly a 100 BB buy-in doesn't seem so big anymore...
Another thought: My first instinct is that this move as described is negative EV and a recipe for winning small pots and losing big ones. But how would you modify it to put the concepts in play for +EV? Where are the pressure points?