Poker

Poker holds a specific place in the study of decision making: it is the game where the hard part is not search but not knowing. A chess engine sees the whole board and must look ahead; a poker engine sees a fraction of the state and must reason about what an opponent’s actions imply about the rest. The wager matters analytically because it prices a probability estimate immediately. A call, fold, bet, or raise converts private information asymmetry, bias, and adversarial updating into an observable cash flow. That is why poker belongs here as a measurement instrument for decision making under information asymmetry, not as a random game with cards attached.

Poker as an imperfect-information game

A game where each player observes only part of the state, so decisions must be made at information sets rather than at nodes.

Group all histories a player cannot distinguish into one information set, and require a single strategy across the group.

A player who cannot tell two situations apart must act identically in both, which rules out the node-by-node reasoning that perfect-information search relies on.

The big idea

Any setting with private information asymmetry, including auctions, negotiation, and security games.

In chess, a position is a node in a game tree, and a player choosing a move knows exactly which node they occupy. In poker, a player holding two private cards cannot distinguish between the histories in which the opponent holds one hand and those in which the opponent holds another. All of those histories collapse into a single information set, and the player must choose the same action across all of them.

That constraint breaks the decomposition that makes perfect-information search work. In chess, the value of a subtree can be computed independently of the rest of the tree. In poker it cannot, because the value of a subgame depends on the probability distribution over hands with which each player reached it, and that distribution is determined by strategy choices made elsewhere in the tree.

Information asymmetry as the state variable. Hole cards are private state. Position controls the order in which private state is acted on. A player on the button acts last on every postflop street, so the button’s policy conditions on a longer public history than any other seat’s, though preflop the blinds still act after it. The asymmetry is structural rather than psychological: one player has more observations from the same deal.

The measurement is expected value by seat, usually reported as big blinds per 100 hands. If $EV_s$ is the mean profit from seat $s$, the value of position is a contrast such as $EV_{\text{button}} - EV_{\text{cutoff}}$ or $EV_{\text{button}} - EV_{\text{big blind}}$. That contrast estimates the value of reducing information asymmetry by one decision.

Information asymmetry also creates the adversarial character of the probability problem. Nature does not choose a card to mislead an observer. An opponent chooses actions from a strategy that knows its actions update the observer’s posterior, so every likelihood term is strategic.

Information sets, and what stays hidden. An information set is a set of game histories a player cannot tell apart from what they have observed. A player who has been dealt an ace and a king faces exactly one dealt state, but they face many possible opponent hands, and every one of those pairings is a distinct history sitting inside the same information set. A strategy assigns action probabilities to the information set, not to the underlying history.

The consequence for algorithm design is decisive. A strategy cannot condition on private cards the player does not hold, and it must remain valid across every history compatible with the observations so far. Any method that reasons about a specific history rather than an information set is solving a different game than the one being played.

Game tree size and the no-limit continuum. Heads-up limit hold’em has on the order of $10^{17}$ game states and roughly $10^{14}$ decision points, which is large but finite and enumerable with sufficient storage. No-limit is larger by many orders of magnitude, and bet sizing is what drives the difference: with a continuous range of legal bet sizes at every decision, the action space is effectively unbounded.

Size is therefore not simply a bigger version of the same problem. Limit poker is a large discrete game; no-limit is a game whose action space must be reduced before it can be represented at all, and that reduction is itself a modeling decision with consequences that any downstream guarantee inherits.

Equilibrium play, and what solved means

A strategy that cannot be beaten in expectation by any opponent, which is a weaker guarantee than playing perfectly against a specific one.

Measure a strategy's exploitability, meaning how much a perfect adversary could win against it, and drive that number toward zero.

Equilibrium is a floor rather than a ceiling. It guarantees you cannot lose in expectation, not that you will win the most.

The big idea

Reading any claim that a game has been solved, and understanding what the claim covers.

A Nash equilibrium strategy in a two-player zero-sum game guarantees the value of the game against any opponent. In poker with symmetric stakes that value is zero minus the rake, so an equilibrium strategy guarantees breaking even before fees against any opponent, including a perfect one.

Exploitability measures the distance from equilibrium: it is the amount a best-responding adversary could win per hand against a given strategy, conventionally expressed in milli-big-blinds per game, abbreviated mbb/g. An exploitability of zero is equilibrium; a large number means an adversary who knew the strategy could profit substantially by exploiting its errors.

Nash equilibrium in a two-player zero-sum game. Equilibrium exists in every finite game by Nash’s theorem, and in two-player zero-sum games it enjoys three properties that make it the natural object of study. Every equilibrium yields the same value to each player, so the value of the game is well defined. Any equilibrium strategy is a maximin strategy, meaning it maximizes the worst-case payoff. And a pair of equilibrium strategies is exchangeable: play one player’s equilibrium against the other’s and the outcome is still an equilibrium.

None of those properties hold in games with three or more players, which is why equilibrium is a design target for heads-up poker and only a starting point for anything larger.

Essentially solved, and the precision the claim carries. Bowling, Burch, Johanson, and Tammelin reported in Science on January 9, 2015 that heads-up limit hold’em is essentially weakly solved. The qualifier is doing real work. Their strategy is not exactly an equilibrium; its exploitability is small enough that a human playing continuously for a lifetime could not distinguish it from one with statistical confidence.

That is a claim about detectability rather than about perfection, and it was chosen deliberately as the standard that matters for a game people actually play. A strategy with positive expectation against it still exists in principle, but no amount of human play could demonstrate the edge, because the residual exploitability sits below the noise floor of any measurement a human could gather.

Why equilibrium does not maximize winnings. An equilibrium strategy is unbeatable and also unmaximizing. Against a weak opponent making obvious errors, an equilibrium strategy collects only the value those errors give away by accident, not the much larger amount available to a strategy that targets them.

This is the central tension in applied poker. A strategy with the strongest guarantee is not a strategy with the highest expectation against any real opponent, and choosing between them is a decision about the opponent rather than about the game. The trade is the same one treated in game theory: a Nash equilibrium is unexploitable, while a best response maximizes against a particular model and becomes exploitable in turn.

Deviating from equilibrium to exploit an opponent

Maximizing expectation against a specific modeled opponent, which necessarily abandons the equilibrium guarantee.

Estimate the opponent's strategy from observed play, compute a best response to that estimate, and play it.

A best response to a specific non-equilibrium strategy deviates from equilibrium itself, so it is exploitable in turn by an opponent who notices.

The big idea

Deciding when to deviate, and understanding the risk that deviation creates.

A best response is a strategy that maximizes expectation against a fixed opponent strategy. If an opponent folds too often to a particular bet, the best response bluffs that spot at high frequency. The gain can be far larger than anything equilibrium play collects, but the conclusion is conditional on the opponent remaining static.

The cost is symmetric and immediate. A strategy that bluffs relentlessly in one spot is itself highly exploitable by anyone who notices, and the exposure created is often larger than the gain captured, particularly against an opponent capable of adjusting.

Best response, and the size of the deviation. A best response is well defined only against a fixed strategy, and real opponents are not fixed. As soon as an opponent adjusts, the best response is aimed at a distribution the opponent no longer plays, and it becomes worse than equilibrium against the new one. The deviation therefore carries two costs: the exposure it creates by being non-equilibrium, and the assumption it makes that the opponent’s strategy is stable long enough for the deviation to pay.

Serious opponent modeling programs address both by mixing the exploitative response with equilibrium play in proportion to the confidence of the estimate. A small confident deviation collects most of the available value while limiting the exposure created.

Exposure created by deviating. When a strategy leaves equilibrium to attack an opponent’s error, its own exploitability grows. A pot-sized bluff at 100 percent frequency in a spot loses roughly a pot per hand to any caller who defends correctly, which is orders of magnitude more than the exploit collected against an opponent folding too often. The exposure is only theoretical if the opponent does not adjust, and it becomes real the moment they do.

This asymmetry is why professional play at high stakes stays closer to equilibrium than analytics would suggest: the population of opponents at that level adjusts fast, and the sustainable edge over equilibrium is small even against measurable errors.

Modeling opponents from limited observations. The practical difficulty is a probability-estimation problem rather than a conceptual one. Estimating an opponent’s frequency in a specific situation requires observing that situation repeatedly, and most situations arise rarely. A read formed from a handful of hands is dominated by noise.

Opponent modeling therefore works better against populations than against individuals. Aggregate tendencies across a player pool are estimated from large samples and are stable; individual reads are estimated from small samples and are not. Against an unknown opponent, a strong opponent, or when a sample is too small to support a read, equilibrium play is the model with the smallest worst-case exposure because it cannot be punished.

Bayesian updating against an adaptive range. A player does not observe an opponent’s hand. The player maintains a probability distribution over combinations, then updates it after each action. With hidden hand $H$, observed action $A$, public history $I$, and opponent strategy $\sigma_o$, Bayes’ rule is

$$\Pr(H=h \mid A=a, I) = \frac{\sigma_o(a \mid h, I)\Pr(H=h \mid I)}{\sum_{h'} \sigma_o(a \mid h', I)\Pr(H=h' \mid I)}.$$

The denominator is the range-wide probability of seeing the action. A preflop no-limit hold’em deal begins with 1,326 two-card combinations before blockers. If earlier actions leave an opponent with 180 weighted combinations and a turn raise is taken with 18 value combinations at frequency 1, 24 drawing combinations at frequency 0.5, and the remaining 138 combinations at frequency 0.05, the action has total weight $18 + 12 + 6.9 = 36.9$. The posterior probability of value after the raise is $18 / 36.9$, or 48.8 percent, not $18 / 180$, or 10 percent. The action is evidence because its likelihood differs by holding.

Against nature, the likelihood term is fixed by the data-generating process. Against an opponent, $\sigma_o(a \mid h, I)$ is itself a policy chosen by someone who knows the observer is updating. A balanced strategy therefore chooses frequencies that make the posterior hard to exploit, while an unbalanced strategy leaks probability mass through actions that occur too often with one part of the range. Fedczyszyn, Koszalka, and Pozniak-Koszalka’s 2012 opponent-modeling chapter treats this as belief construction over a hostile and uncertain environment rather than as a guess about a single hand.

Pot odds, equity, and the call decision

The comparison between the price a call offers and the probability of winning, which every decision in the game reduces to.

Divide the amount to call by the total pot after calling to get the required equity, then compare against estimated equity.

The arithmetic is trivial and the estimate is not. Bias and errors in poker are almost never arithmetic errors.

The big idea

Every call decision, and as the frame within which range estimation matters.

Pot odds are a decision rule. Let $P$ be the pot before the call and $c$ be the amount to call. A call with equity $e$ has expected value

$$EV(call)=eP-(1-e)c.$$

Break-even occurs at $EV(call)=0$, so $eP=c-ec$ and

$$e^* = \frac{c}{P+c}.$$

Calling 50 when the pot is 150 before the call therefore requires $50/(150+50)=25$ percent equity. If the estimated probability of winning is 26 percent, the call is positive by $0.26\cdot150 - 0.74\cdot50 = 2$ units. At 24 percent, it is negative by 2 units. The threshold is arithmetic; the equity estimate is the hard part.

Equity against a range, not against a hand. Estimated equity must be computed against the distribution of hands the opponent can hold given their actions, not against a single guessed hand. A hand that is 70 percent against one holding may be 30 percent against the range that actually takes this line.

This is where nearly all bias and error enter. A range estimate is a judgment about what an opponent would do with each part of their distribution, and it is the part of the calculation that cannot be reduced to a formula. Solver output helps by pinning down what a competent opponent’s range at a given node should contain, but transferring solver output to a real opponent still depends on how closely their play tracks the solver’s.

Implied and reverse implied odds. Pot odds as stated ignore future betting. Implied odds account for money expected to be won on later streets when a draw completes, which justifies calls that immediate pot odds do not. Reverse implied odds account for money expected to be lost when the hand improves and is still second best.

Both adjustments are estimates rather than calculations, and both are routinely used to rationalize a call that immediate arithmetic rejects. An honest version specifies how much additional money is expected and with what probability, and that specification is usually where the argument collapses.

Where the range estimate breaks down. Range estimates fail most reliably in three places. Multiway pots have too many joint distributions to hold in mind, and even experienced players default to head-to-head reasoning that ignores the third and fourth players. Rare lines have insufficient data to support a stable range, so any estimate is dominated by a prior. And board runouts that carry heavy card removal effects, particularly ones that put a paired board on a wet turn, produce ranges whose composition is genuinely counterintuitive.

Solver work sharpens the estimate in each case, but only for a player willing to memorize the outputs. Without that, a call decision defaults to broad heuristics that the solver has shown to be wrong on specific textures.

Probability audit of a single decision. A single river call shows why poker is a probability instrument rather than a story about nerve. Suppose the pot is 100, the opponent bets 75, and the call costs 75. The break-even probability is $75/(100+75+75)=30$ percent if measured against the final pot, equivalently $75/(175+75)=30$ percent using the pot-after-bet convention. The action is correct only if the conditional probability that the bettor is bluffing exceeds 30 percent.

Information asymmetry supplies the prior. The caller sees public cards and their own blockers, but the bettor’s hole cards remain private. If the bettor reaches the river with 50 value combinations and 30 missed draws, then the prior bluff probability inside that betting range is $30/(50+30)=37.5$ percent. If the observed bet is used by all 50 value combinations but by only 10 of the missed draws, Bayes’ rule changes the probability to $10/(50+10)=16.7$ percent. The same private state and the same board produce opposite decisions because the action changed the conditional probability.

The record of the decision is pot 100, bet 75, call threshold 30 percent, and posterior bluff probability 16.7 percent or 37.5 percent depending on the modeled betting frequency. Every term is observable or explicitly estimated. That is the difference between analyzing information asymmetry and describing a dramatic card hand.

Ranges, board texture, and what a bet represents

Reasoning about distributions of hands rather than single holdings, which is the structural change that separates modern play from earlier styles.

Assign each player a set of hands consistent with their actions, then evaluate decisions against those sets.

A bet is informative about a distribution, not about a hand, and the correct response depends on the whole distribution.

The big idea

Every postflop decision, and any attempt to interpret an opponent's action.

A range is the set of hands a player can hold given the actions taken so far, weighted by how often each is played that way. Reasoning in ranges rather than in specific hands is the single largest change in how the game has been played over the past two decades, and it follows directly from the structure of imperfect information.

Range advantage describes which player’s distribution performs better on a given board. A board of high cards favors the range that contains more high cards, which is typically the preflop aggressor, and that advantage determines who can profitably bet and at what frequency.

Board texture, and how it interacts with a range. A dry board with few draws favors a player with the stronger range, because the opponent has little incentive to continue without a made hand. A wet board with many draws compresses the advantage, because the weaker range contains more hands with meaningful equity against the stronger one.

The equilibrium pattern is about distribution rather than hand strength: strategies bet more often and smaller on boards that favor their range as a whole, and less often but larger where advantage is concentrated in a few hands. Solver output on unpaired dry boards is characteristically a small bet at high frequency; on wet paired boards it is a mix of large bets, small bets, and checks that a hand-strength heuristic never predicts.

Polarized against condensed ranges. A polarized range consists of strong hands and bluffs with little in between, and it is bet large because the middle is absent. A condensed range consists mostly of medium-strength hands and is bet small or checked, because large bets are called only by better.

This distinction determines what a large bet on a late street should be read as. If a player’s range at that node is polarized, a large bet is the correct sizing for the whole range and does not narrow it further. If the range is condensed, a large bet represents a small slice of it and is far more informative. Reading a bet without knowing which type of range the bettor’s line has produced is where most misreads originate.

Blockers, and card removal effects. Blockers are cards in one’s own hand that remove combinations from the opponent’s range. Holding an ace reduces the number of ace-high hands the opponent can have, which changes the frequencies on both sides. Card removal is a small effect per hand and a large one in aggregate, and it is one of the clearest places where solver output disagrees with intuition.

Blocker effects also determine which hands a solver selects as bluffs. Given a choice between two busted draws, a solver prefers to bluff with the one that removes the opponent’s calling hands and unblocks their folding hands, since that combination extracts the maximum fold equity. Recovering this intuition from first principles is possible; matching solver bluff selection without memorizing it is not.

Bet sizing and making an opponent indifferent

Choosing a bet size so that the opponent is indifferent between calling and folding, which fixes the bluffing frequency exactly.

Set the ratio of bluffs to value bets equal to the pot odds the bet lays.

If an opponent is indifferent, they cannot exploit either response, which is what makes the strategy unexploitable at that node.

The big idea

Constructing a betting strategy, and understanding why solver frequencies take the values they do.

At equilibrium, a bet size determines a bluffing frequency and therefore a bluff probability. Let $P$ be the pot before the bet and $B$ the bet. A bluff-catcher calling $B$ wins $P+B$ when the bettor is bluffing and loses $B$ when the bettor has value. If the bettor’s betting range contains value weight $V$ and bluff weight $L$, the call is indifferent when

$$\frac{L}{V+L}(P+B)-\frac{V}{V+L}B=0.$$

Solving gives

$$\frac{L}{V}=\frac{B}{P+B}.$$

A pot-sized river bet has $B=P$, so the solved ratio is one bluff for every two value bets, and bluffs are one third of the betting range. A half-pot bet has ratio $0.5P/(1.5P)=1/3$, so bluffs are one quarter of the betting range. Bluffing frequency is not a personality trait; it is the indifference condition that prevents the caller from profiting by always calling or always folding.

Minimum defense frequency. The symmetric quantity on the calling side is the minimum defense frequency, or MDF, the proportion of the time a player must continue against a bet to prevent the bettor from profiting with any two cards. Against a pot-sized bet it is 50 percent; against a half-pot bet it is roughly 67 percent.

MDF is often misapplied as a target probability. It describes the frequency required to make an indiscriminate bluffing strategy unprofitable, not a frequency that maximizes expectation against a real opponent whose bluffs are concentrated in particular holdings. A defender who folds more than MDF against a bettor who never bluffs is playing correctly; a defender who folds less than MDF against a bettor who over-bluffs is also playing correctly. The number matters as a floor rather than as a target.

Multiple sizings and mixed strategies. Solver output typically mixes several bet sizes at a single node rather than selecting one. Different parts of a range benefit from different sizes, and using multiple sizes lets a strategy extract more from strong hands without making its small bets transparent.

The gain from mixing sizes is real but modest, and it is far smaller than the gain from having correct frequencies at one size. A player implementing a single well-chosen size correctly gives up little against one implementing three sizes badly, which is why solver simplifications that collapse mixed sizings into a single one lose only a small fraction of the value.

Sizing against stack depth. Bet sizing at equilibrium also tracks the stack-to-pot ratio. Deep stacks support multi-street plans that resolve on the river; shallow stacks compress the plan into fewer decisions and shift equilibrium bets toward all-in.

A short-stacked heads-up game plays almost entirely as a preflop shove-or-fold game once effective stacks fall below about ten big blinds, and equilibrium in that regime is a table indexed by hand and position. Deep-stacked no-limit is a very different game and its solver output looks very different: small bets, mixed sizings, and river decisions that turn on card-removal effects a shove-fold table cannot express. Confusing regimes is a specific error class and it is common enough at intermediate skill that it accounts for a measurable fraction of losing decisions in tournament middle stages.

Counterfactual regret minimization

The algorithm behind every superhuman poker result, which converges to equilibrium through repeated self-play.

Accumulate regret at each information set for each action not taken, and play proportionally to positive accumulated regret.

Minimizing regret at every information set independently bounds the exploitability of the average strategy over time.

The big idea

Any large imperfect-information game, and as the standard against which later methods are compared.

Counterfactual regret minimization, or CFR, works by playing a game against itself many times and, at each information set, tracking how much better each alternative action would have done. Actions with high accumulated regret are played more often in subsequent iterations, and the time-averaged strategy converges to equilibrium.

Zinkevich, Johanson, Bowling, and Piccione introduced the method at the Neural Information Processing Systems conference in 2007. That paper has no Crossref DOI, so it is named here rather than cited; the results built on it are cited below and are all in Science.

Regret matching, and the update rule. At each information set, a player maintains a running sum of counterfactual regret for each action, defined as the value they would have collected by taking that action multiplied by the reach probability of the information set, minus the value they actually collected. Actions with positive accumulated regret are played in the next iteration with probability proportional to that regret; actions with non-positive accumulated regret are not played.

This update is called regret matching, and its guarantee is local. It bounds the regret at each information set as the iteration count grows, and the bound is uniform across information sets. Nothing about the update requires reasoning about the whole game at once, which is the property that makes CFR scale to games where a single explicit game tree is out of reach.

Convergence of the time-averaged strategy. A key theoretical property is that the current strategy does not converge to anything useful, but the average over iterations does. This is counterintuitive and is a source of most implementation errors, since the natural thing to store is the latest strategy rather than the running average.

The bound is on average regret, which decreases as iterations accumulate, and exploitability of the average strategy is bounded by that quantity. Convergence is therefore guaranteed but slow, with error falling roughly as the inverse square root of iterations in the original formulation. Practical solves rely on variants that improve this rate substantially.

CFR+ and variants used in practice. CFR+ modifies the update so that accumulated regrets are floored at zero rather than allowed to go negative, and it uses a weighted rather than uniform average. In practice it converges substantially faster than the original algorithm, and it was the variant that made solving heads-up limit hold’em feasible rather than merely theoretically possible.

Later variants used at scale include Monte Carlo CFR, which samples chance and opponent actions rather than enumerating them, and Discounted CFR, which discounts early iterations that reflect an untrained strategy. Each variant trades a different property for wall-clock speed, and each represents a specific engineering commitment about which parts of the game tree require exact treatment.

Solving heads-up limit hold’em

The first solution of an imperfect-information game played competitively by humans for money.

CFR+ run to convergence over the full game, with compressed storage of the accumulated regrets.

Limit poker has a finite discrete action space, so the game could be represented exactly without abstraction.

The big idea

As the reference case for what solving means, and for the distinction between limit and no-limit.

The Science paper of January 9, 2015 reported the result and stated its guarantee carefully. The strategy’s exploitability is below the threshold at which a human opponent could establish an edge over a lifetime of play, which the authors formalized as essentially weakly solved.

The engineering behind the solve was substantial. Storing regrets for a game of this size required compression, and the computation ran across a cluster for months. A point that generalizes is that the game was solved exactly rather than approximately in structure: no abstraction was applied to the game tree itself.

Precision, and what essentially solved covers. The authors reported an average exploitability below 1 mbb/g. To put that number in scale, professional heads-up limit hold’em win rates at the highest levels rarely exceed a few tens of mbb/g against strong opposition. An exploitability under 1 mbb/g is inside the noise of any humanly gatherable sample, and that is the criterion the authors argued makes the game solved for practical purposes.

The strategy is not the only equilibrium, and not exactly at equilibrium, and the paper is careful to say so. It is a strategy no human strategy can beat over any length of play a person can complete.

Compute and storage the solve required. The compute was substantial but not extraordinary by 2015 cluster standards. The storage was the tighter constraint, since a naive representation of the accumulated regrets for the whole game exceeds any single machine’s memory. The authors used a compression scheme that exploited the sparsity and structure of the regret table, storing entries in a way that traded a small amount of computation at lookup for a large reduction in memory footprint.

That engineering trade is worth naming because it is what turned a formally feasible solve into an actually feasible one. Reports that credit hardware for the result understate the role of the algorithmic and representational work that made the hardware usable.

Why limit was tractable and no-limit was not. Limit hold’em has a fixed bet size at every decision, so the action tree is discrete and modest at each node. No-limit permits any bet between a defined minimum and the effective stack, so the action tree at every decision is finite only because chips are discrete, and it is large enough that solvers abstract the bet sizes down to a handful before they can search it.

Any no-limit solver must therefore choose a finite set of bet sizes to represent, and the strategy it computes is an equilibrium of the abstracted game rather than of the real one. Bounding the gap between the two is a research problem in its own right, and the fact that no equivalent to the 2015 result exists for no-limit is a consequence of that gap rather than of insufficient compute.

DeepStack, Libratus, and no-limit

Two independent programs that beat professional players at heads-up no-limit in 2017, using different approaches to the same intractability.

DeepStack resolved subgames continually using learned value functions; Libratus solved a blueprint and refined endgames in real time.

Neither could solve the full game in advance, so both moved computation from offline preparation to the moment of decision.

The big idea

Understanding the shift from precomputed strategies to real-time reasoning in imperfect-information games.

DeepStack, reported by Moravcik and colleagues in Science on May 5, 2017, applied continual resolving: rather than computing a complete strategy in advance, it solved a local subgame at each decision, using deep networks trained to estimate the values of the resulting situations. It defeated professional players over a large number of hands.

Libratus, reported by Brown and Sandholm in Science on January 26, 2018, took a different route. It computed a coarse blueprint strategy for the whole game in advance, then solved the endgame exactly in real time whenever play reached a subgame small enough to handle, and it patched the holes opponents found in its abstraction between days of play.

DeepStack and continual resolving. DeepStack decomposed the game not spatially but temporally. Instead of computing a full strategy offline, it treated each decision as the root of a small subgame extending a few streets forward, solved that subgame with CFR, and used a neural network to estimate the counterfactual values of the resulting leaf states. The network was trained on a large synthetic corpus of solved random poker situations, so the estimator generalized to positions the program had never seen.

The key theoretical point is that DeepStack maintained a valid subgame decomposition despite imperfect information, using the counterfactual value machinery that CFR provides. That machinery is what makes continual resolving safe: a subgame solved in isolation would in general not be part of any coherent full-game strategy, and DeepStack’s construction was designed to avoid exactly that failure.

Libratus and endgame refinement. Libratus took the opposite architectural bet. It precomputed a blueprint strategy for the entire game using a coarse abstraction of bet sizes, and on the first two betting rounds it rounded an opponent’s off-tree bet to the nearest size in that abstraction. From the third betting round, the turn, it solved the rest of the game in real time with no card abstraction and a dense set of bet sizes, building a new subgame each time an opponent bet, which took the observed betting into account with far greater fidelity than the blueprint’s abstraction supported.

Between days of the Brains vs AI match at Rivers Casino in January 2017, the team analyzed the professionals’ play and added strategy patches where opponents had exploited a specific abstraction hole. That between-session update is part of the result: Libratus was not a static strategy that beat professionals; it was a system that adapted at daylong resolution to attempts at exploitation.

Results against professional opposition. Both programs beat strong human professionals in heads-up no-limit at statistically meaningful margins. Libratus, at the 2017 Brains vs AI event, played 120,000 hands against four professionals and won at 147 mbb/g, a margin the authors reported at 99.98 percent statistical significance. DeepStack played 44,852 hands against 33 professionals recruited by the International Federation of Poker, each asked to play 3,000 hands (11 completed them), and won 492 mbb/g overall, more than four standard deviations above zero.

The guarantee is still the two-player zero-sum one. Neither program was shown to be at equilibrium in the full game, but in a two-player zero-sum setting an empirical result against strong opposition is a meaningful measurement rather than an anecdote. Sample sizes here were designed to reach significance, and reaching it is the fact those papers established.

Six players, and a game with no equilibrium guarantee

A program that beat professionals at six-player no-limit poker despite the absence of any theoretical guarantee that it could not be exploited.

A blueprint strategy computed in self-play, refined with limited real-time search during play.

With more than two players, an equilibrium strategy no longer guarantees the value of the game, so the theory that justified the heads-up results does not apply.

The big idea

Judging what a multiplayer game result does and does not establish.

Pluribus, reported by Brown and Sandholm in Science on August 30, 2019, defeated elite professionals in six-player no-limit hold’em. The engineering result is striking on its own terms: the blueprint was computed in a matter of days on a modest amount of hardware, orders of magnitude less than the earlier heads-up systems consumed.

The theoretical position is materially weaker, and the authors said so. In games with more than two players, an equilibrium strategy carries no guarantee: multiple equilibria exist with different values, and playing your part of one equilibrium against opponents playing another can lose. Pluribus was not claimed to be unexploitable.

Why multiplayer breaks the zero-sum guarantee. Nash’s guarantees in two-player zero-sum games rely on the minimax theorem, which does not extend to games with more than two players. In a three-player game, playing your component of an equilibrium against opponents who are not playing theirs can put you in a worse position than a random strategy would. Coalitions, implicit or explicit, further complicate the picture: two players who play to break even against each other can jointly extract from the third, and no equilibrium refinement rules that out.

None of these problems has a clean fix. Multiplayer poker admits algorithms that produce strong empirical play, but the theoretical object those algorithms are approaching is not the same object that CFR converges to in the two-player case, and the difference is not cosmetic.

Pluribus, and the compute it used. Pluribus computed its blueprint strategy in about 8 days on a 64-core server, with less than 512 GB of RAM, at a cloud-computing cost the authors put at under $150. During play it used two CPUs, no GPU, and took on the order of seconds per decision. Those figures were the surprising part: none of the earlier headline results had used so little.

The efficiency comes from a specific engineering choice. Pluribus computed a coarse blueprint offline and then did limited real-time search over the current subgame using a modified CFR variant, so the offline compute did not need to be extreme. That inversion of the compute budget is what made the whole system affordable, and it has since propagated to other domains where a similar decomposition applies.

What the result does and does not establish. The result establishes an empirical fact rather than a theoretical one. Against that set of opponents, over that sample, the program won at a statistically significant rate. That is a real measurement and it was designed carefully, with variance reduction applied to reach significance at a feasible sample size.

It does not establish that the strategy is unbeatable, that it approximates any equilibrium, or that it would survive opponents who studied it and adapted. Those are open, and in a multiplayer setting some of them may not have clean answers at all. Headlines that report Pluribus as having solved six-player poker are wrong in a specific way, and knowing why is the value of the qualifier the paper carefully attached.

Variance, and the sample a poker result needs

The gap between a poker win rate and its standard deviation, which is roughly an order of magnitude and dictates enormous sample requirements.

Compare a win rate in big blinds per hundred hands against a standard deviation measured on the same scale.

When noise exceeds signal by a factor of ten, the sample needed to distinguish them grows with the square of that ratio.

The big idea

Interpreting any claimed win rate, and judging whether a downswing carries information.

Win rates in online no-limit hold’em are conventionally quoted in big blinds per hundred hands, or bb/100, and a strong winning rate at moderate stakes is a few big blinds per hundred. Standard deviation over the same hundred hands is typically in the range of 80 to 100 big blinds, depending on the game and playing style.

The ratio between the two is the problem. With signal of about 3 bb/100 and noise of about 90 bb/100, distinguishing a winning player from a breakeven one to reasonable confidence requires a sample in the hundreds of thousands of hands. Most players never reach it in a career.

Sample size and its confidence bounds. The 95 percent confidence interval around an observed win rate follows the standard formula for a mean:

$$\text{CI}_{95} = \bar{w} \pm 1.96 \cdot \frac{\sigma}{\sqrt{n / 100}}$$

Here $\bar{w}$ is the observed win rate in bb/100, $\sigma$ is the per-hundred standard deviation, and n is the number of hands. Plugging in $\sigma = 90$ and requiring the half-width to be less than the observed win rate demands $n$ on the order of $(1.96 \cdot 90 / 3)^2 \cdot 100$, which is about 350,000 hands. That is the sample at which a player observing 3 bb/100 can reject the null hypothesis of breaking even, and no self-assessment made before that point is inference.

Downswings consistent with a winning strategy. A genuinely winning player has a substantial probability of losing stretches of tens of thousands of hands. Such a stretch is not evidence that anything has changed, and treating it as evidence is a common error in bankroll management.

The same arithmetic runs the other way. A winning month is not evidence of skill, and sample sizes at which people form confident conclusions about their own play are far below the threshold at which conclusions are supportable. Standard bankroll tables reflect this by requiring buy-ins that would be excessive if variance were smaller.

Why results are not evidence at typical samples. The practical consequence is that self-assessment from results is unreliable over any horizon a person actually observes. The alternatives are to evaluate decisions rather than outcomes, or to gather samples large enough that noise averages down, and only the first is available to most players.

Solver-driven decision review is the dominant modern discipline for exactly this reason. It sidesteps the variance problem by asking whether a specific decision at a specific node was correct given the range and the pot odds, and it gives an answer whose signal is not diluted by everything else that happened in the session.

Cognitive biases priced by the pot. Poker punishes bias because each probability bias has a denominator. The pot fixes the price, the estimated equity fixes the probability threshold, and the result supplies noisy feedback that is often uninformative at human sample sizes. That makes the game a laboratory for the failures cataloged in cognitive biases rather than a morality tale about risk preference.

Outcome bias is the poker form of results-oriented thinking. A call with 26 percent equity at a 25 percent threshold is correct even when the next card loses, and a call with 24 percent equity is wrong even when it wins. The sign of the expected value is known before the outcome; the outcome adds variance, not retroactive evidence about the decision. Aiyer and colleagues’ 2023 replication of Baron and Hershey’s outcome-bias experiment found that outcome information changed decision-quality judgments even when the underlying choice was held fixed.

The sunk-cost effect appears as pot commitment. Money already placed in the pot is no longer owned by the decision maker, so it cannot justify a new call. The current decision sees only the pot now offered, the cost to continue, and the conditional probability of winning. Arkes and Blumer’s 1985 sunk-cost experiments give the general bias; poker prices the same mistake as a call whose equity falls below $c/(P+c)$ but feels mandatory because earlier bets were large.

Illusion of control appears when private agency is confused with card generation. Choosing a line changes the opponent’s range, fold probability, and future information asymmetry, but it does not change the distribution of unseen cards. Langer’s 1975 experiments named the bias. In poker notation, strategic control enters through $\sigma_o(a \mid h, I)$ and future action sets, not through the chance distribution over the deck.

Rake, and the economics of a poker game

The fee the house extracts from each pot, which is large relative to the win rates players compete for.

A percentage of the pot up to a cap, charged per hand, or a time charge in live games.

Players compete against each other for a pool that the house has already reduced, so the average player must lose.

The big idea

Evaluating whether a game is beatable, and understanding why player pools deteriorate over time.

Poker is not a negative-expectation game against the house in the way that casino games are; it is a zero-sum game between players from which the house removes a fee. That fee is nonetheless decisive, because it is large relative to the edges players are competing for.

Online rake is typically around 5 percent of the pot subject to a cap, and in small-stakes games the effective rate measured against amounts won can consume the entire edge of a moderately skilled player. Live games charge either a drop per pot or a time charge, with the latter usually cheaper for a winning player.

Rake as a percentage of a typical win rate. The comparison that matters is between rake paid per hundred hands and win rate per hundred hands. In many small-stakes online games, rake paid exceeds the win rate a good player achieves, meaning the same player would be winning comfortably in a game with no fee and is losing or breaking even in the real one.

This is the clearest statement of the game’s economics. The question is not whether a player beats their opponents but whether they beat their opponents by more than the house charges to let them try, and at every stake below high the answer for the median player is no.

Online against live economics. Online and live cardrooms operate on different fee structures for reasons that have as much to do with staffing as with theory. Live rooms carry high fixed costs per table and rake accordingly; online rooms carry marginal costs near zero per hand and use rake volume rather than rake per hand to fund operations. That difference produces the recurring result that a hand of live poker is far more expensive to play than a hand of online poker even at nominally similar stakes.

Hand rates run the other way. Online tables run 60 to 100 hands per hour per table, and a multi-tabling player can put in several thousand hands per session; live tables run about 25 to 35 hands per hour. Rake per hour, which is what a professional cares about, therefore favors live games in some segments and online in others, and the honest comparison depends on stakes, format, and playing style rather than on a fixed preference.

Ecosystem dynamics, and why games die. A poker game is sustained by losing players, and rake accelerates their departure. Every hand transfers money from the player pool to the house, so in a closed pool with no new entrants the pool shrinks until it cannot support the games. That dynamic explains a great deal about how online games have evolved: pressure toward recreational-friendly rules, restrictions on tools and tracking software, and seat-matching limits are responses to it.

Cardrooms and sites that permitted unrestricted use of tracking software, hand-history sharing, and seat selection watched their recreational player populations depart faster than their reg populations could adjust. Restrictions that look anti-competitive from a professional’s perspective are commercially rational from an operator’s, since a game with no losers has no house fee to collect. That tension is the ongoing shape of the industry, and it is why the strongest online games available at any given moment are usually the ones that just changed rules.