IBM’s Deep Blue has apparently “solved” the game of chess – having defeated the reigning human world champion in 1997. Much more recently, Google’s AlphaGo defeated an 18-time world Go champion. Can it be long before computer programs are able to trounce humans in every sort of intellectual game? The developers of AlphaGo certainly think so, as the lead researcher just posted a paper in which he claimed to have created a program capable of playing poker at the level that “approached the performance of human experts and state-of-the-art methods.” Specifically, the paper asserts that the program, called NSFP, has “got within reach” of solving the game of Limit Texas Hold ‘em.
While I don’t doubt that the program was able to win a good many simulated poker games, it cannot “solve” Texas Hold ‘em any more than a powerful robot can solve boxing. Poker is not simply a game of odds, moves, and calculations. It is a game of controlled and exploited emotions – including greed, fear, over-confidence, and anger. Because a computer does not experience emotions, it cannot play poker any more than a motorcycle can run a marathon. Whatever NSFP was doing – and no matter how well it was done – it was only a facsimile of real poker.
As has often been said, you don’t actually play the cards in poker, you play your opponents. Properly understood, poker is a betting game, not a card game. If you don’t believe that, try playing for bottle caps or toothpicks. You will quickly find out that the game is pointless; no more strategic than War and involving less skill than Go Fish.
Poker only becomes meaningful when played for real money, at which point considerations of fear, greed, overconfidence, and self-doubt come into play. A good player is able to keep his or her emotions under control, while exploiting the weaknesses of opponents. There are players who can excel at low stakes, but who get progressively worse as the stakes increase – finally falling completely apart in no-limit games with everything on the line. A computer obviously does not care about money, however, and therefore has neither fear of losing nor anticipation of winning. It cannot be intimidated or tricked; it cannot falter at a crucial moment; it cannot worry about the rent; it cannot fall into debt. Thus the human and the computer are not playing the same game.
There is another reason that humans and computers simply do not play the same game when it comes to poker. The key strategy in poker is to convince your opponents to make bad choices – betting against you when they are sure to lose, and folding their hands when they are likely to win. As the poker theorist David Sklansky has explained, this is best done by showing strength when your hand is weak, and weakness when your hand is strong – but not always. To avoid predictable consistency, it is necessary to “vary your play,” for example by sometimes betting strongly on a good hand, or even allowing yourself to be bluffed.
This is not as easy as it sounds, and most players cannot avoid falling into patterns or showing “tells.” The computer program, however, can build in true randomization, which solves the problem of creating identifiable patterns, but only in the sense that a laser scope solves the problem or aiming a rifle or autotune solves the problem of singing on key. Errors are eliminated, but only by changing the nature of the activity itself.
The poker legend Jack Strauss was said to be the “master of the withering bluff,” who was able to intimidate opponents into folding good hands, for fear that Strauss somehow held a winning pocket pair. That tactic could not work against a computer program – which cannot feel a sense of loss, worry, or inadequacy. (Read more about Strauss and other poker legends such as Doyle Brunson, and their similarity to trial lawyers, in my book Lawyers’ Poker.)
A computer can be programmed to "want" to win, but it cannot "fear" losing, in the sense of suffering consequences beyond the loss of the game itself. Most games can be played simply as a contest of skills — chess and Go, for example — entirely within the parameter of the game. Winning can be its own reward, and the loss itself the only undesirable consequence. (You can bet on chess, of course, but that is not essential to the game.) Poker, however, is different, perhaps uniquely so, because each player is trying to take the others' money, not simply to achieve a defined goal or outcome. The objective in poker is not holding the best cards, or even winning the most hands — instead, it is manipulating your opponents into making the most bad decisions at the highest cost.
Playing human poker for real money is like walking on high wire — perhaps over Niagara Falls, if the stakes are great enough. For a computer, however, it is as though the tightrope is suspended only a foot or so above the ground. You may not "want" to fall off, but nothing bad will happen if you do.
Or to put it differently, poker is a game of nerves, and a computer doesn’t have any. A computer cannot fall prey to provocation, cajolery, manipulation, irritation, envy, embarrassment, or dashed expectations, which means that it cannot be outplayed. We can think of many other games that can be “solved” by eliminating the human element, but so what? If an engineer were to create an unhittable pitching machine, no one would suggest that it had solved the game of baseball. Likewise, a kid with night-vision goggles has not solved hide-and-go-seek.
So I wish only the best to the developers of NSFP. Perhaps their efforts can eventually teach us something about the process of complex decision-making with incomplete information. But they just aren’t playing poker.
(h/t Jim Schulz)
Haven't you in effect, and somewhat arbitrarily, just introduced a new "rule" for, or new element in the definition of, poker — namely, that it must be played by human beings with human emotions? We do have comparable rules for sporting contests, of course: either they must be played with no non-human physical elements at all (such as arm-wrestling or a barefoot race) or, more typically, the specific physical accoutrements involved (balls, nets, sticks, motorized vehicles) are regulated. I don't think we usually interpret the broad category of games quite the same way. But if you're right, why can't we use the same maneuver to deny that computer programs are "really" playing chess, Go . . . or tic-tac-toe, for that matter? And what would you make of an idiot savant who succeeds at poker precisely because they can calculate odds and vary strategy with little or none of the emotional investment normal players experience? In any event, this all sounds like a reprise of John Searle's "Chinese room." I'm inclined to think that if a computer program can win big pots in poker games, there's no meaningful way to deny it's playing poker.
As RQA notes, this critique really isn't all that dissimilar from comments made about chess computers like Deep Blue and early dismissals of chess playing programs. There is also a human element to chess playing; nerves, intimidation, overconfidence, intuition, etc. all play a role, especially at levels of play below that of grandmaster. Just ask anyone in the park with clock set to five minutes.
That there are human dimensions we consider essential to excellent play in a game does not mean that a computer cannot be designed in a way that consistently "beats" a human based on the parameters of "winning" the game without those essentials. In poker there is just as much of a defined goal or outcome as chess, the win parameter is taking all "the others' money." No one would walk away from a poker table empty-handed claiming that they had really "won" because they instilled more fear or bluffed better than the player with all the chips.
In other words, that a computer is successful at meeting the win parameter in a way that human players find unsatisfying, in large part because the computer is not subject to the same pressures as a human player, does not mean that the computer is not really "winning" the game. Our dissatisfaction may lead us to think differently about why we play the game, but that's another matter.
While not "solved" or close to being "solved" (this is a term of art that does not mean "beaten by a computer" or even "almost always beaten by a computer" as the post seems to imply), Chess has essentially hit that point. It is not surprising at all that limit poker may be next. I highly recommend Kasparov's 2010 piece in the New York Review of Books titled "The Chess Master and the Computer" which talks about this, and even discusses poker-playing computers at the end of the review.
Try playing poker for something that you do not value — say, potato chips or pebbles — and you will quickly learn that it is not the same game. It is not a question of "human dimensions that are essential to excellent play." It is a question of the nature of the game itself, which depends upon the experience, or risk, of loss that extends beyond simply being declared the loser.
I do not deny that NSFP is doing something impressive, but it is not playing poker.
Let me accept, then, that there is the distinction you've drawn between games "the nature of . . . which depends upon the experience, or risk, of loss that extends beyond simply being declared the loser" (like poker) and games for which that's not true (like chess and Go, I gather). Is that anything more than a distinction between games that "essentially" involve gambling and those that don't? Blackjack played for potato chips isn't much fun, either. To be sure, risk affects a game like poker with a community pot differently than it does one like blackjack played against the house, but "the experience, or risk, of loss" would seem to me to be an essential element of what you're defining as the game in either case. If all of that is so, I think I'd be comfortable agreeing that computers can't gamble and therefore can't be said to "play" what are inherently gambling games.
Here is the problem for most of us criminal practitioners. The deck is stacked against us. The State typically holds the best cards (the police) has the most cards (the money) and the wild card, Joker (sorry, I mean Judge).
I suppose I'm with RQA here. After all, the writers of NSFP could enter it into the world series of poker, and then it would be playing for real money. Would its win suddenly be worth more (if it did) because real money was at stake? If that answer is no, because the only way to win at poker is to do so while _also_ feeling human emotion, then we are definitely talking about two different things. Not that there's anything wrong with that – Five Thirty Eight has a cottage industry of comparing sports results between different eras (facing different types of adversity).
Michael
You're onto it.
Lubet writes: "The objective in poker is not holding the best cards, or even winning the most hands — instead, it is manipulating your opponents into making the most bad decisions at the highest cost."
Strange way of viewing the objective: which is, of course, to walk away with the most (or all of the) cash.
Poker is no different from the stock market, in this respect. Riddled with irrational and rational emotions, like fear, greed, etc. Winning requires predicting and "reading" the market properly (and, of course, an element of chance).
Perhaps needless to say, computers are now being used to play in the stock market casino very effectively.