Last week an ex-Google friend of mine who also builds poker AIs directed me to a 2+2 thread about a new video poker machine from IGT that offers players 6-way fixed-limit hold’em against five bots. The bots play poorly, but the catch, of course, is really high rake — 25% up to 2.25 bets maximum, or 9 bets in the pot to reach max rake.
The game is called Texas Hold’em Fold Up, and it already has gotten a lot of attention.
In the 2+2 thread, John Mehaffey and Jimmy ”Gobbo” Fricke had a interesting discussion about the game, whether or not it’s beatable, and how it could be an effective way to earn Caesars’ Seven Star status in perhaps a less expensive sort of way than can be done via other casino games. Mehaffey wrote a good article about the game last week as well. Both the thread and Mehaffey’s article are worth checking out for anyone who wants details of the game as well as smart discussion about it.
Since there is still no Total Rewards in Brooklyn, I have not played the game as yet, but that’s not going to deter me from offering an opinion on this “6-max poker” game that stands as a clever take on the standard video poker machine. I know that many of my WSOP buddies enjoy a bit of video poker, so here is my attempt to answer some obvious questions that come to mind about Texas Hold’em Fold Up.
Is the game beatable?
Well, the casino certainly does not think so. We’ll get into the details of the five bots and 25% rake, but it’s worth noting that the game does give out casino loyalty points and is distributed by IGT which controls something like 95% of the video poker market and many of the patents. It’s safe to assume that IGT considers this primarily as yet another video poker game.
According to Mehaffey’s piece, Texas Hold’em Fold Up was created by Michael Baker, whose background is in life insurance and financial management — in other words, not in AI. These days to sell a casino game, a blackjack side-bet, or a slot machine concept to the casino industry, you’d need beforehand to demonstrate that the new game can’t be beaten, even by a very smart player. (If it did not exist already, I’m pretty sure that you could not get blackjack into a modern casino, except perhaps the ”blackjack pays 6-to-5” variety.)
But you’re pretty good at limit hold’em, you say. Can you gain an advantage with good play? Certainly.
As with table games and video poker, this game is designed so that the casino gets a minimum win from everybody, even from a very good player, and the casino has a chance to win even more when players make mistakes. According to the Wizard of Odds, you can play most video poker games within 0.5% of break even just by following a few rules, but most players don’t do that. The casino wins 1-2% more from actual video poker players than they would if those players played by the book.
Similarly, unless you’re playing with bad rules, following that pocket blackjack card at a Vegas table also gets you to within 0.5% of the house edge. Most players lose about 2% per bet instead, by making a few costly mistakes.
Going back to Texas Hold’em Fold Up, then, 25% rake is very high, and the casinos think you can’t beat it. I’m not tempted to say that they are wrong, and I doubt that you could beat that rake just with ”good poker.” But if we break down the structure of the game, perhaps there are some spots to exploit and others to avoid. We’ll take a look at that in just a bit.
Is Texas Hold’em Fold Up like the Heads-Up Hold’em machine?
The “Texas Hold’em Heads-Up Poker” machines, which have been had a niche presence in Vegas casinos since 2010, are also distruted by IGT. But those machines were created by a small company called G2 Game Design in an office park off of S. Rainbow Boulevard in suburban Las Vegas. The company purchased a neural network-based AI for heads-up limit hold’em developed by a Norwegian scientist in the late 1990’s.
First came the AI, which Frederik Dahl (an engineer with a background in AI) developed for his own personal interests, then a businessman bought the rights and they turned it into a slot machine. They convinced IGT and some of their customers that the AI plays so well that it can hold any limit hold’em player to a draw and beat most of them substantially, without the need for rake of any kind.
There’s more interesting backstory here, but after some players did play pretty well against the Texas Hold’em Heads-Up Poker machines, the casinos kept them but removed the player points. The heads-up machine indeed plays very good poker, and it is very hard to beat. While there is no rake, you also get nothing back.
This new 6-max machine is different. From what Mehaffey and Fricke have reported, the five bots play bad poker, but the rake is set so high the casinos are still willing to give you some of that rake back, in casino rewards points.
So how does a player get maximum value playing Texas Hold’em Fold Up? Much of this is speculation on my part, but there are a few factors in your favor:
- not all bets have the same value (you pay no additional rake above a 9-bet pot)
- the bots play badly, and predictably (with some logic, which I attempt to guess below)
- by state law, the bot cannot adjust to your play (every hand is played independently)
- the button moves, so position matters
- the goals of the casino and the game developer may not be the same
With some of these factors in mind, let’s think a little further about the game design, and about who is designing the game.
Bots playing badly — it’s the whole point
In an interview with The New York Times, Gregg Giuffria, owner of the Texas Hold’em Heads-Up Poker machines rolled out before, talks about the problems with running an AI that is too good and therefore beats the players too quickly. More to the point, poker players are drawn to opponents that play badly rather than to those who play close to game theory optimal poker (a topic I’ve explored here before).
The new 6-max game seems to solve this by offering five terrible opponents, but raking so much that you still can’t win. You see the bots making terrible bets and call downs, so you play more. If you make a side-bet that pays out a bonus when you make premium hands (bets for which the house has a significant edge), so much the better.
How would you go about designing an AI for a bad poker bot, albeit one that will still win given a 25% cushion? That’s an interesting question to consider.
As I learned from building a deep learning heads-up poker AI, reading board texture is hard for computers, much harder even than for a mediocre human player. It’s easy to calculate your odds for making a pair, for making a flush, and to approximate the odds of a one-pair hand winning on the river. But it’s much harder to know when your hand will have good implied odds, when your middle pair might be vulnerable, or when that same pair is pretty much the nuts given the way a hand has played out.
Exploiting these different spots when playing the new Texas Hold’em Fold Up game is your advantage, especially as the background of the game creator isn’t in math or computer science. I think it’s safe to assume that the bots play a rules-based, fixed strategy which is not sophisticated at all.
Mehaffey and Fricke’s examples of the bots calling down with king-high and seven-high on paired boards suggest that the bots may even not be reading the board texture at all, and instead are simply looking up the odds of winning with “a pair on the river,” even if the bot is playing the board. To be fair, the seven-high call involved a hand drawing to a straight and flush draw on the turn. It’s probably safe to assume that the bot will draw to any straight or flush, given some minimal pot odds.
I promise there will be specific suggestions — if you can’t wait, you can go ahead and scroll down below. First, though, I want to take a look at an example hand, just to see how devastating 25% rake can be.
Just like small-stakes hold’em?
A game with a high rake but with weak players who draw to the end and who never fold? Sounds a lot like actual limit hold’em games from back in the boom days. Or perhaps like small-stakes underground poker games, where players don’t come after work to fold, or to think too long about pot odds.
When I think about loose limit hold’em games, I think of the seminal book Small Stakes Hold’em by Ed Miller, David Sklansky, and Mason Malmuth. And when I think of that book, I think about playing a vulnerable top pair hand with multi-way action.
Suppose you hold on a flop. Your value against two random hands in such a situation is 61%, with the value of the other two being 19.5% apiece.
If you were to bet and both of your opponents called, with no rake your $1 bet is worth $0.83 in surplus value, since you always lose the $1 you bet, but you get $3 back 61% of the time .
Now if you’re paying 25% rake on that bet, it’s still costing you a dollar to bet, but you lose 25% of the whole $3 when you win. The value of your bet falls to $0.37 surplus value. That’s much more than a 25% cut in profits!
Let’s consider the same hand with 4-way action. Again, you are up against random opponents who are calling you blind, and you’ve got top pair with on a flop. In this case your hand has a value of 49%, with the other three random hands having a value of 17% each.
In the no-rake case, your surplus value here goes up to $0.96. (Of course it does, you want the extra caller.) In the 25% rake case, your value also goes up — to $0.47.
This example is against two or three completely random players. In reality, just eliminating some bad preflop hands from your opponents’ ranges means your bet loses almost all of its marginal value after the rake, even if you are still up against random hands and passive opponents.
I could not quite come up with a break-even scenario, but if you give one opponent , another , and the third one for middle pair, your surplus value goes down to just 17 cents, after 25% rake.
While the hand is getting raked, you are playing the ”penny auction” of poker. The competition is soft, but you’re paying a fee for every bid, whether you win or not. Thus does a great spot like top pair on the flop become a barely-profitable bet.
Meanwhile bluffing becomes especially expensive, since you’re paying a 25% tax when you win, while the bot gets full value to call you down. Calling a single bets, though, is still usually worth it. Getting 8-to-1 from the pot, compared to 6-to-1 after rake, is not really a major difference in most cases.
So what might be some good spots for the player?
I think it’s safe to assume that the bots are not explicitly colluding against you. I don’t know if that’s promised in the game details, but with a 25% rake cushion, the bots don’t need to collaborate to have an edge, and it’s really not that simple to code rules-based bots to team up on your hands, in any case.
That said, if two bots are raising each other until the street caps, this does put you in a tough spot with many a drawing hand, so it’s reasonable to think that the bots’ rules make sure that this happens often, without resorting to explicit collusion.
Mehaffey and Fricke mention the bots tend to cap once they bet and are raised. They do this both against the player and against each other. All of this suggests that the bot bets out with most of its good hands, and with some of its good draws, too. When it calls, it calls with weaker made hands, and weaker draws. It should actually be relatively easy to put the bot on a range of hands.
All of this means good spots for the player include:
- really strong hands (70%+ value), especially multi-way
- good draws (like big flush draws), since you will be called down
- the river, if you make it that far
- the button, as good hands will bet into you, and weak hands will check
While the pot is small, you pay a penalty to call, and a bigger one to bet. It’s better to get to a 70% winner cheaply, then make much more valuable bets. Even better would be making it past the 9-bet mark and stopping paying rake thereafter, then getting your flush paid off by middle pair.
I imagine that on the river, especially in position, it’s possible to get very good value, until and unless the bot’s algorithm is upgraded. Given that the bot cannot adjust to your play — and since the whole point of this poker game-slash-slot machine is to show you bad players you think you can beat — the river upgrade does not seem likely to happen anytime soon.
I’d either try to get great odds for a strong multi-way hand, or I’d try to see the river cheaply. Otherwise, you’ll probably want to fold a lot of marginal hands, or even good hands that just don’t make too many 70%+ boards. The bots can afford to look you up with middle pair, and they will. You can’t afford to call them off with similar hands in return.
The bots fold face up — does this help the player?
When the bots fold in the game, they fold face up (hence the “fold up” name). Does this provide an advantage to the player?
Not so much. It’s hard for the bot’s AI to adjust to board texture, and as mentioned before by state law it’s not allowed to adjust to your play. But it’s easy to account for the odds of making a straight or a flush based on cards folded face up. Use the information too, but this isn’t where the player’s edge comes.
You’re better off hoping that the bots hold on to their weak hands and longshot draws while their cards are still face down. Maybe they will make just enough of a hand to pay off on the river. Remember, the bots are tuned so that they defend against being taken off a hand, so bet good hands on the end that they will pay you off.
So we know casinos are not in the money-losing business. But could they have missed something? Possibly.
New casino games and slot machines are introduced in stages. Whoever creates a new game first sinks money into design and manufacturing, and also creates some math to show that a game cannot be beaten by a card-counting shark. Then, once a prototype is approved for testing, it gets introduced to sample casinos in Biloxi or wherever. There, the distributor gets a real idea for how popular the game might be, as well as how much it actually makes, against real players.
If the players don’t play, the game never makes it to Vegas. If the casino can’t beat the players, it also never makes it to Vegas. But if you put $2 million into a new game design, you really want to make sure that players voluntarily want to sit down and bet real money playing it. This is more important than anything else, especially to the game creator.
Given how much the 25% rake throws off the odds for even pretty straightforward spots like betting top pair in a loose 6-max limit game, it’s certainly possible that real players lost a tidy sum but also kept playing, given some of the terrible hands that the bots showed down. The adjustments you’d need to make to beat the game might be so unusual to a good limit hold’em player, that the test casino players of Texas Hold’em Fold Up never thought to try them.
Did I mention that there are no blinds? Instead, players post a small ante every hand. If position is as important as I suppose it could be, this is a good move by IGT to make it that much more painful to fold. And yet, it’s probably important to fold mediocre hands that make vulnerable top-pair, middle-pair type hands, while paying full rake.
Could you simulate this?
It wouldn’t be that hard to create players that compete independently, knowing that they need to beat a player who is paying 25% rake up to 2.25 bets. And in turn, to simulate a bot for the player’s side. But it’s not so simple that I could do it with my neural network AI in a few days.
I’d expect that the EV for the player’s hand is almost always negative, but often close to the value of posting an ante. Ante-only for this game really is a sharp piece of design. If I played, I think it would be too tempting to play too many hands, especially if you’re playing something like $2/$4 limit. Once you’re in, you get sucked into calling several bets, and not really in a position to bet anyone off. Not until you make a good hand, or make it to the river, could you get any real value from the bots.
I bet that a neural network simulation might be pretty good at figuring out what kinds of preflop hands lose the most, and which ones can get to those valuable late-game spots a little more often. Also, it would be nice to get an idea of how much position is worth.
How will the game change? Will it last?
Once a game hits the casino floor it never really goes away, but I don’t know if I see this one becoming too popular, as currently constituted. From what I’ve read, it sounds like the players lose every time. To make the game popular, you want the players to win sometimes, and that means decreasing the rake on the main game. Of course, lowering the rake means that the bots will have to play better, too.
That said, playing fast, small-stakes 6-max hold’em against opponents that fold face-up sounds like it could be a lot of fun. I can see this catching on, especially as part of a video poker mix, but not if the casino takes back 25% of every pot.
Also, if the rake goes down, some bluffs may become profitable, and it may not be necessary to nit it up as much as I suggest for this current version. You want players to have fun, and also to have a chance to win when they are having fun. If the fun style loses the most, this becomes a problem for the game’s longevity.
How many people will really want to nit it up at $2/$4 against a computer? Even for those who are happy to pay some rake to kill time after a tourney, will they keep playing if they never book a winning session?
Will there be a no-limit hold’em version anytime soon?
I don’t think so. Betting with a slider is just too complicated for a casino game, where you expect to play a hand every 45 seconds. I can see something with a “pot” or “max-bet” button, but the casino doesn’t want you to lose (or win) everything in one hand. They want you to make a lot of small bets, take a percentage of each bet, and occasionally to let you hit a longshot for a big payoff so that you win and come back.
Game design sounds in theory like it might be fun, but it’s actually pretty tedious. You have to please several different parties at once, and if either the players, the casino, or the game distributor is not satisfied, your game will never make it to the casino floor. There aren’t too many people who design games successfully, and if any of them are reading this, please let me know what I got wrong here.
Texas Hold’em Fold Up sounds like a fun game, and hopefully it will still be Vegas next summer for me to take a shot playing it during next year’s WSOP. Until then, those of you who do get a chance try it out, let me know what you think.
Nikolai Yakovenko is a professional poker player and software developer residing in Brooklyn, New York who helped create the ABC Open-Face Chinese Poker iPhone App.