54 minute read

Background – History of Solving Poker

There has been a rich literature of research and algorithms involving games and game theory, with poker specific research starting to grow in the late 1990s and early 2000s, especially after the creation of the Computer Poker Research Group (CPRG) at the University of Alberta in the early 1990s. Real life decision making settings almost always involve imperfect information and uncertainty, so algorithmic advances in poker are exciting with regards to the possible applications in the real world.

Research accelerated, partially thanks to the founding of the Annual Computer Poker Competition (ACPC) in 2006, which has led to people from around the world competing to build the best poker agents, which has in turn led to significant research in algorithm development, abstraction, and game solving techniques. In recent years, we have also seen multiple “man vs. machine” contests to test the latest poker agents against some of the best poker players in the world.

After some major results from around 2015-2018, some researchers have moved on to even more complex imperfect information games like Hanabi.

Note that it may make sense to read highlights from this section and come back to learn more about the history of certain algorithms when they are mentioned later in the tutorial.

Some of the most important highlights from poker research have been:

  1. 1998: Opponent Modelling: University of Alberta Computer Poker Research Group creates agent Poki – output (fold, call, raise) based on effective hand strength (Billings et al)
  2. 2000: Abstraction Methods: Bucketing hands together (also bet sizes) (Shi, Littman)
  3. 2003: Approximating Game Theoretic Optimal Strategies for full scale poker: Using abstraction on HULHE (Billings et al)
  4. 2005: Optimal Rhode Island Poker (1 private card, 2 public cards, 3 betting rounds): Sequence Form Linear Programming to produce equilibrium player with only lossless card abstraction (Gilpin, Sandholm)
  5. 2006: Beginning of the Annual Computer Poker Competition, which became a yearly part of the AAAI conference
  6. 2007: The iterative Counterfactual Regret Minimization (CFR) algorithm was developed at the University of Alberta which allowed for solving games up to size 10^12 and solving for inexact solutions (Zinkevich et al)
  7. 2015: Heads-up Limit Hold’em Poker is Solved, the first major poker game solved without abstractions (Bowling et al)
  8. 2015: Brains vs. AI competition, the first major Heads Up No Limit poker competition where top humans defeated CMU’s poker agent
  9. 2017: DeepStack Expert-Level Artificial Intelligence in No-Limit Poker, likely the first agent that was superior to humans in No Limit Hold’em (DeepMind)
  10. 2017: Superhuman AI for heads-up no-limit poker with the Libratus agent that overwhelmingly defeated top humans in a large scale competition (CMU)
  11. 2019: Deep CFR, the first algorithm to use deep neural networks to implement CFR without the need for abstractions (Facebook)
  12. 2019: Superhuman AI for multiplayer poker, the first algorithm to defeat top humans in multiplayer poker (Facebook)

Early Poker Research and Theories

The earliest signs of poker research began over 70 years ago. In the Theory of Games and Economic Behavior from 1944, John von Neumann and Oskar Morgenstern used mathematical models to analyze simplified games of poker and showed that bluffing is an essential component of a sound poker strategy, but the games analyzed were very basic.

Harold W. Kuhn solved 1-card poker with 3 cards, also known as Kuhn poker, by hand, analytically, in 1950.

The first major effort to build a poker program occurred in the 1970s by Nicholas Findler. He designed a 5-card draw poker program as a method to study computer models of human cognitive processes in poker. The program was able to learn, but was not considered a strong player, despite 5-card draw being less complex than Texas Hold’em, in part because all cards in 5-card draw are private.

In 1984, Mike Caro, now perhaps most famous for his “Book of Poker Tells”, a book that examines poker player behaviors that can give away, or tell, information about their hands, created a one on one no limit hold’em program called Orac. It was faced off against strong human opponents, but the matches were not statistically significant and were not well documented.

Darse Billings, in his master’s thesis from 1995, adeptly noted that despite poker generally being considered very dependent on “human elements” like psychology and bluffing, perfect poker is based on probabilities and strategic principles, so Billings suggested focusing on the mathematics of poker and game theory before considering the human or rule-type systems.

Given this insight, how have “feel” players who rarely consider advanced mathematical principles thrived in poker for decades? Doyle Brunson’s famous book, Super/System: A Course in Power Poker from 1978 mainly prescribes a style of aggression, notably in cases where most players would be more passive. Billings notes that this advice was based off of experience and not mathematical soundness, but is valid from a mathematical perspective, because opponents are far more likely to make mistakes when faced with a bet or raise than when facing a check or call.

The implication here is that despite many players not considering sophisticated mathematical and game theoretical principles, they can develop to use plays that make sense mathematically learned over time from trial and error and recognizing patterns that work. However, nowadays, many of the world’s best players do indeed use sophisticated statistical software to analyze their hands and their opponent’s hands, as well as simulation software to run mathematical simulations on common hand situations.

History of Poker Agents

The main types of poker agents have been knowledge based (using rules or formulas), simulation based, and game theoretical.

Research in recent years has mainly focused on Texas Hold’em using game theoretical strategies and building agents that play as close as possible to the Nash equilibrium of the full game, while generally using either abstractions or deep learning or other methods for games beyond the size of Heads Up Limit Hold’em.

An excellent (though now outdated) paper explaining the state of computer poker, called “Computer poker: A review” was published in 2010 by Jonathan Rubin and Ian Watson of the University of Auckland. Parts of this section follow their review closely for the time period leading up to 2010.

Knowledge Based Agents

Domain knowledge systems can be rule-based, which means that they use if-then logic, incorporating variables such as number of opponents, hand strength, position, and bets to call. Or they can be formula-based, which requires a numeric representation of hand strength and pot odds. The domain knowledge system then ranks the agent’s hand over all possible hands (ignoring the opponent’s hand likelihoods) and can be modified to also include hand potential (i.e., not only the current hand value, but the probabilistic hand value by the end of the hand).

These exploitative agents focus on isolating weaknesses in the agent’s opponents. They must have an opponent model that is generally set up as (fold, call, raise) probability triples and generally set up histograms of hands into buckets based on actions.

David Sklansky and Mason Malmuth’s well regarded 1994 book Hold’em Poker for Advanced Players recommended a basic rule-based system for Limit Texas Hold’em, with variations according to betting position and game conditions. Most human players probably operate under some sort of rule-based system whether they are thinking about it or not. This is more obvious late in poker tournaments when most players are either going allin or folding and so players generally have a rule-based system for which hands they’d go allin with in which position and with knowledge of opponent actions. For example, with 10 big blinds or lower, a good strategy might be to go allin with over 50% of hands from the small blind as a rule, but the rule could also include other factors like the opponent in the big blind.

In David Sklansky’s 1999 text, The Theory of Poker, Sklansky states his Fundamental Theorem of Poker: “Every time you play a hand differently from the way you would have played it if you could see all your opponents’ cards, they gain; and every time you play your hand the same way you would have played it if you would call all their cards, they lose. Conversely, every time opponents play their hands differently from the way they would have if they could see all your cards, you gain; and every time they play their hands the same way they would have played if they could see all your cards, you lose.”

The individual hand vs. individual hand concept was then expanded by Phil Galfond in a 2007 article in Bluff Magazine that pioneered the idea of thinking of opponents’ hands in terms of a range of possible hands, not a particular hand.

Quantitative Poker, a web blog, expanded on this and added the levels of range vs. range (level 2), strategy vs. strategy (level 3), and finally, strategy vs. inferred strategy distribution, which allows for the idea that opponent strategies will be employing a strategy from a range of strategies.

In 2001, Furnkranz identified opponent modeling as a neglected area of research within game playing domains and mentions its importance in a game like poker. The risk with opponent modeling techniques that seek to exploit opponents (exploitative strategies) is that the agent then subjects itself to exploitation, especially if the model of the opponent is invalid or incorrect. Such a model is required for exploitative agents, which can be either static, based on predefined opponent characteristics, or adaptable, based additionally on opponent’s actions during the match. While opponent modeling has shown some promise in poker research, the most successful solutions now tend to use a more game theoretic approach, which limit their exploitability at the expense of maximizing value against weaker opponents. Although perhaps after the game theoretic approach is exhausted, improving opponent modelling will be the next frontier.

Rule-based systems generally show up preflop, where fixed rules make more sense given the limited possibilities of hand combinations. The University of Alberta Computer Poker Research Group created one of the first poker agents in 1999, called Poki, which took an effective hand strength as input and outputted a (fold, call, raise) probability triple, using a form of basic opponent modelling.

Various programs were created in the late 1990s and early 2000s including r00lbot, which, based on rules in the Advanced Hold’em Poker book, did 3000 trials simulating hand to showdown against n random opponent holdings, and then made decisions. Poki (and its earlier iteration, Loki) used opponent modelling and a formula-based betting strategy and was successful when tested in Internet Relay Chat (IRC). A later version of Poki took first place in the 2008 6-player Limit Hold’em tournament in the ACPC.

The problem with such systems is that they require expert rules and there are too many situations possible to have rules for all of them, which requires merging situations that are not necessarily similar enough. Also the systems are generally not very scalable, because they cannot maintain enough rules for the game and therefore become exploitable. If you tried writing down rules for how a bot should play poker, it would quickly become overwhelming! That said, this was an exciting first approach and may work well in smaller game settings.

Simulation Based Agents

Simulation agents can use a simulation like Monte Carlo Tree Search (MCTS – see Section 2.2 for more details) to predict action and hand rank probabilities for specific opponents. By predicting opponent strategies, the agent can develop its own optimal counter strategy.

After the Alberta Computer Poker Research Group’s formula based Loki-1 agent, they created Loki-2, which added a simulation function to “better determine the expected value of check call and bet/raise decisions”. They called this selective-sampling simulation since there is a selective bias when choosing a random sample at a choice node. Opponent hands were based on a weight table, instead of randomly. This means that each possible hand the opponent could have would be assigned a weight (e.g. 55 0.03, 88 0.05, JT 0.08, etc.), which would allow for the Loki-2 agent to make decisions accordingly. In the late 1990s, Alberta recommended this as a general framework for solving games with imperfect information. These changes made the second iteration of this agent earn 0.05 small blinds per hand more on average than the first version.

The next version of this agent was called Poki, which was split into a formula based model and a simulation based model. The simulation based model was the more advanced one, using more recent research techniques, but did not really outperform the formula based model. The simulation model strangely seemed to play overly tight in one on one situations and overly loose in full table situations (strong poker players are usually exceptionally loose in one on one situations, playing almost all hands, and quite tight at full tables). The Alberta team realized that decision making based on the expected holdings of the opponent may not be the best technique.

In 2008, a simulation-based agent called AKI Realbot participated in the Annual Computer Poker Competition (ACPC) and took second place. It used an exploitative strategy to play more against weak players, which it defined as players who had lost money overall in the prior 500 hands. Most of its gains came from exploiting these weak opponents, which is also a strategy commonly seen in real life poker where expert players target weak opponents and try to play as many hands as possible with them.

In 2009, Van den Broeck created an agent using MCTS for a NLHE game, using “an offline non-player specific opponent model by learning a regression tree from online poker data, which were used to predict actions and hand rank probabilities for opponents”. The agent worked in no limit poker games and seemed to perform well against basic rule-based opponents.

Game Theoretic Equilibrium Solutions

We have continued to see interesting applications with games and AI succeeding in the real world, including Deep Blue famously beating Kasporov in chess in 1997, Watson defeating Jennings in Jeopardy in 2011, and most recently AlphaGo defeating Lee Se-dol in Go in 2016. These major results, however, differ from the trends in poker solving, since most work in poker is now involving equilibrium strategies that are unexploitable, rather than building AI whose goal is to defeat or exploit individual opponents (although AlphaGo was a robust agent that did not target a specific opponent) and also because poker has imperfect information. In recent years we have seen major results using game theoretic approaches, which are now almost exclusively used in poker solving.

In this section, we look at the history of linear programming game theoretic solution techniques, while saving the more recent CFR game theoretic techniques for the next section.

In 1999, Selby produced equilibrium strategies for the preflop betting round of Texas Hold’em. This was done by effectively assuming that the hand ended after the preflop betting round and that the five community cards would simply be revealed with no further betting. Even though there are 52c2 = 1326 possible starting hands in poker, this can be reduced to 169 since many are equivalent (since only both cards being the same suit or different suits is relevant in the preflop stage).

In 2000, Shi and Littman produced a game theoretic agent for a one on one game called Rhode Island Hold’em by using abstractions to bucket together hands and bet sizes. This was a major result because even though the game is a toy game with three betting rounds and one private card per player, this was effectively the start of the game theoretic trend and the trend of using abstractions to solve large games. This agent was able to beat a range of rule-based opponents.

In 2003, the research community began to focus on Heads Up Limit Hold’em (HULHE) as a challenge problem, which is when the CPRG at Alberta first attempted to create an approximate solution for the game using abstraction, rather than sequence form linear programming (SFLP). Billings et al. created PsOpti, which focused on near-equilibrium solutions for full two player HULHE, a game with 10^18 game states.

They knew that the sequence form linear programming techniques of Koller were not sufficient to solve such a large game, so abstractions were needed, which reduced the game to about 10^7 game states by using bucketing abstractions and eliminating betting rounds. They also built a preflop model and multiple postflop models, since they couldn’t create one postflop model given the size of the game. In practice, the same preflop model would always be used and then it would choose a postflop model according to what happened preflop.

A combination of PsOpti agents was named Hyperborean and won the first AAAI Computer Poker Competition in 2006, although the PsOpti agent was found later to not be a great equilibrium approximation since it was defeated by a wide margin by later Nash approximation strategies.

The combination of preflop and postflop models was later identified as problematic since this does not guarantee a coherent equilibrium strategy (since there may be multiple equilibriums). However, in 2009, Waugh showed that separate subgames can be pieced together as long as they are “tied together by a common base strategy”.

The Abstraction-Solving-Translation approach requires that a full-scale poker game be modeled by a smaller, solvable game. The abstracted game is then solved to find an approximate Nash equilibrium, which is then translated to choose actions in the full game. This technique is now very common in game theoretic solution techniques in the poker space. In general, SFLP can solve games up to about 10^8 info sets, so most games require more abstraction than possible through lossless (i.e. card isomorphism in poker – when different sequences of cards are actually the same strategically) abstraction alone.

In 2005, Gilpin and Sandholm, after applying abstractions, used SFLP to solve Rhode Island Hold’em completely, which has 3.94x10^6 info sets after symmetries are removed.

In 2006, Gilpin and Sandholm created the GS1 agent to find game theoretic solutions to HULHE. They used a new technique called GameShrink to abstract the game tree automatically and then used offline solving for the preflop and flop betting rounds and real time solving for the turn and river betting rounds. They later improved this with a GS2 iteration that improved the abstraction technique with k-means clustering.

The game tree in NLHE can have up to around 10^161 nodes (depending on the stack sizes), significantly larger than in LHE, where it has only about 10^18 nodes. Aggrobot was created in 2006 and is similar to PsOpti, but was made to solve NLHE and adds the use of betting abstractions, which meant that the agent could bet only half pot, full pot, double pot, and all-in. While the abstraction itself is relatively straightforward, the issue of translating opponent bets to match the best estimate within the abstract game was a challenge because it can lead to easy exploitation (e.g. by betting between the abstracted sizes).

Miltersen and Sorenson, in 2007, solved for Nash equilibrium solutions for two players at the end of a tournament with 8000 chips each and 300-600 blinds where actions were restricted to be allin or fold only. They showed that this simple allin or fold strategy approximated Nash equilibrium for the full tournament setting (probably because with only about 13 big blinds, going allin or folding is generally the best play). Because of special considerations with tournaments that require hands to be looked at in the context of the stage of the tournament in addition to the expected value of winning chips on a particular hand, most poker research has avoided these complicating factors by focusing on cash games, in which each hand is independent from the next, and generally simplify things even further by starting stack sizes at the same amount for each hand rather than proceeding with having the effective stack size be smaller than the default size (which would entail using a more complex agent).

In 2006, Gilpin, Hoda, Pena, and Sandholm developed Scalable EGT (Excessive Gap Technique), which was improved in 2007, by Gilpin, Sandholm, and Sorensen. EGT works to solve optimization problems like linear programming, but allows for suboptimality and therefore faster solving. This was used to solve Rhode Island Hold’em with 3x10^9 histories.

Also in 2007, Counterfactual Regret was developed by Zinkevich et al., which revolutionized poker research and will be explored in detail in this tutorial.

The linear programming (LP) algorithms in 1995 were capable of solving games with around 10^5 nodes in the game tree. This grew in 2003 to 10^6, in 2005 to 10^7, and then in 2006 to 10^10 with the development of EGT. Since 2007, both EGT and CFR have been able to solve games in full with about 10^12 nodes.

Iterative Algorithms

Ficticious play involves two players who “continually play against each other and adapt and improve over time”. Each player starts with an arbitrary strategy and trains with random situations, while knowing the correct moves since he knows the opponent strategy, and then keeps updating the strategy, which results in approaching a Nash equilibrium as iterations increase.

In 2006, Dudziak used fictitious play to develop an agent called Adam for HULHE with abstractions. In 2010, Fellows developed other fictitious play agents called INOT and Fell Omen 2, which respectively took 2nd of 15 and 2nd of 9 in the HULHE instant run off competition in 2007 and 2008 in the ACPC.

In 2007, the Range of Skill algorithm was developed at Alberta by Zinkevich. This is an iterative procedure that creates a “sequence of agents where each next agent employs a strategy that can beat the strategy of the prior agent by at least some epsilon amount”. The agents would then eventually approach a Nash equilibrium. This technique led to the creation of SmallBot (uses three bets per round) and BigBot (uses four bets per round), which performed well compared to prior agents in HULHE.

Other Algorithms

Case based reasoning effectively stores “cases” of knowledge of previous hands and then when similar situations come up again, uses those to determine how to play. Agents are trained on hand histories of other agents or players and use that data which gives the output of a betting decision from the most similar case encountered before. It would seem difficult to outperform agents that are almost being copied using this technique, but the method shows promise when using data from multiple top agents, since this could make its play more unpredictable and perhaps more robust since if one agent had a particular leak, it would be countered by cases from other agents.

Bayesian agents were attempted, but achieved poor results in the ACPC in 2006 and 2007.

Major challenges with agents in general are the creation of “accurate opponent models in real time with limited data, dealing with multiple opponents, and dealing with tournaments that have a variable structure”, including changing blinds as the tournament progresses. Only very recently in around 2019 have we started to see solutions for multiple opponents, while tournaments and variable structure has not yet been studied in academia. Accurate opponent models may be an interesting future research domain, but most recently research has been very focused on game theoretic solutions.

Exploitative Strategies

As we have discussed, a Nash equilibrium strategy is “static, robust, and limits its own exploitability by maximizing a minimum outcome against a perfect opponent”. A true Nash equilibrium cannot be outplayed and also will not exploit weaknesses in opponents. To exploit weaknesses requires opponent modelling, which means deciding actions based on some opponent characteristic, which opens up the agent itself to exploitability as a consequence.

In 2004 the CPRG created HULHE agents Vexbot and BRPlayer, which use game trees as a way to create opponent models through hands that are observed and generalizing these observations to other similar situations, where there are various levels of similarity, and if the closest similarity level is not possible then the agent attempts to find a similarity at the next level. The difference between the two agents is simply that BRPlayer uses five similarity levels and Vexbot uses only three.

Robust counter-strategies offer a compromise between exploiting an opponent and minimizing one’s own exploitability. For example, an \(\epsilon\)-safe best response is a strategy from the set of strategies exploitable for no more than \(\epsilon\), that maximizes utility against a particular opponent. The Restricted Nash Response (RNR) algorithm takes a target opponent’s strategy and a parameter \(p \in [0, 1]\) which trades off between minimizing exploitability and exploiting the opponent. CFR is then used to solve the game, fixing the player’s strategy at the input strategy for \(p\) and for \((1 − p)\), and allowing the player to choose their own actions. When p = 0, the agent uses equilibrium strategies and when p = 1, it attempts to use a fully exploitative strategy.

Since Restricted Nash Response can only be implemented if given the opponent’s strategy, Data Biased Response (DBR) is an option to create robust counter-strategies, which works by choosing an abstraction for an opponent model, mapping the real game observations of the opponent into the abstract game, and determining frequency counts of the observed actions. Then, like with the RNR algorithm, a modified game is solved in which one player is playing based on the opponent model with probability determined at each information set (not at the root), and the other player converges to a robust counter-strategy. The Hyperborean agent got 2nd in the two player limit bankroll competition in 2009 using DBR.

Another strategy is using Teams of Agents, in which the agent selects which strategy to use from a group of exploitative strategies, and picks one depending on the opponent, using the UCB1 algorithm for the policy selection procedure, described as like having a “coach” that determines which “player” to put in against certain opponents to maximize the chance of winning. Johanson showed that this results in an agent that is able to achieve a greater profit against a set of opponents, compared to the use of a single \(\epsilon\)-Nash equilibrium strategy against the same opponents.

Finally, there is frequentist best response, which takes an offline model for an opponent by observing training games and assuming that his strategy is fixed, which requires having a default policy for gaps in the observed data, and then plays based on a best response to this strategy.

Ganzfried and Sun explored the idea of opponent exploitation in an early 2017 paper. They note that opponent exploitation is crucial in imperfect information games and has much wider use than Nash equilibrium, which is only valid in two-player zero sum games.

They propose opponent modelling based on a Bayesian model. The model involves starting with an equilibrium approach and then adjusting to the opponent’s weaknesses (deviations from equilibrium) as more information about his play is revealed by modelling the opponent assuming that he begins play with an equilibrium strategy and then playing an approximate best response to that strategy, which is updated as more information about the opponent is learned.

They had strong results with this method against weak ACPC agents and trivial opponents that always call or play randomly.

One major problem with opponent modelling in poker is that most hand data does not include private hole cards unless they are shown at the end of a hand, which means hands that are seen tend to be stronger, since, for example, bluffs would not be seen as frequently (players only have to reveal their hole cards if they bet first or called a bet and are winning, otherwise they can throw them into the “muck”). Counter-strategies such as building up a best response model offline against a particular opponent are possible, but may require a very large hand sample on the opponent before having enough data to effectively learn their strategy. This strategy is also risky because it could lead to significant exploitation against other opponents or an opponent who changes his strategy, as most do. Finally, on the fly opponent modelling is difficult because sample sizes tend to be quite low, so only very few information sets would be seen even over thousands of hands.

The advantage with opponent modelling is that an agent could rotate between playing a game theory optimal strategy and using opponent exploitation, where the exploitation would be used only against perceived weaker opponents with a sufficient amount of data on their play. In fact this is how many professional human players attempt to play currently.

Counterfactual Regret Minimization (CFR)

CFR grew out of the Annual Computer Poker Competition that began in 2006 as a way to build a more competitive agent. Counterfactual regret was introduced in the 2007 paper by Zinkevich et al. called Regret Minimization in Games with Incomplete Information and has continued to this day to be the primary computer poker research focus and the basis of the majority of the best poker agents and algorithms that exist.

The CPRG developed two successful agents based on CFR: Hyperborean to compete in the ACPC and Polaris to compete in Man vs. Machine poker events. The Man vs. Machine events are covered in the AI vs. Human Competitions Section 5.2.

In CFR, memory needed is linear in the number of info sets (whereas linear program memory needed is quadratic). An approximate Nash equilibrium is reached by averaging player strategies over all iterations, which improves as the number of iterations increases. Solving games is usually bounded by memory constraints, so CFR has resulted in a similar improvement as compared to the sequence form linear programming as sequence form did with normal form linear programming. CFR is guaranteed to converge and has experimentally worked well even with non two player non zero sum games. It has been used to solve games with as many as 3.8x10^10 information sets in 2012.

Evolution in Measuring Agent Quality

The main ways to measure the quality of a poker agent are (1) against other poker agents, (2) against humans, and (3) against a “best response” agent that always plays the best response to see how well the agent does in the worst case.

Measuring Best Response

2-player Limit Texas Hold’em Poker has about 10^18 game states, so in 2011, computing a best response was thought to be intractable because a full game tree traversal would take 10 years to compute at 3 billion states per second. Strategies could still be computed by using abstract games, but a true best response was out of reach.

In 2011, Michael Johansen et al. developed an accelerated best response computation that made it possible to measure a Heads Up Limit Hold’em strategy’s approximation quality by efficiently computing its exploitability. This was able to show that exploitability has dropped drastically from 2006, when it stood at about 330 micro big blinds per hand using a Range of Skill program, to the high 235 in 2008 to 135 in 2010, and continued to drop until achieving CFR+ in 2014, which solves HULHE directly without abstraction, with an exploitability of less than 1 mbb/h.

The paper was important because it was able to evaluate agents that had been used in competitions in terms of their exploitability. They showed that the 2010 competition winner was significantly more exploitable than agents that it defeated slightly.

Another important insight is that while increasing abstraction size in toy domains, improvements are not guaranteed, but in large games like Texas Hold’em, it seems that increasing abstractions continues to provide a consistent, although decreasing improvement.

Finally, an experiment was done to show that the technique of minimizing exploitability in an abstract game in order to use that strategy in the main game can have some interesting overfitting consequences. The exploitability in the main game tended to decrease (a good thing) initially, but then as iterations continued to increase, the exploitability would increase in the main game. That is, the strategy improves in the abstracted game while worsening in the main game.

The paper concludes that finer abstractions do produce better equilibrium approximations, but better worst-case performance does not always result in better performance in a tournament.

In 2013, Johanson et al. created CFR-BR, which computes the best Nash approximation strategy that can be represented in a given abstraction. This additional evaluation method compares the representation power of an abstraction by how well it can approximate an unabstracted Nash equilibrium. Using CFR-BR, Johanson showed that imperfect recall abstractions tend to perform better than their perfect recall abstractions, despite the theoretical issues.

In 2014, Sandholm and Kroer introduced a framework to give bounds on solution quality for any perfect-recall extensive-form game. It uses a newly created method for mapping abstract strategies to the original game and uses new equilibrium refinement for analysis. This resulted in developing the first general lossy extensive-form game abstraction with bounds. It finds a lossless abstraction when one is available and a lossy abstraction when smaller abstractions are desired, and has now been extended to imperfect-recall.

In 2017, Lisy and Bowling showed a new method called Local Best Response to compute approximate lower bounds on the best response strategy. These can be used even in large NLHE games like those played in the ACPC. The algorithm computes the approximation by looking one action ahead and assuming that players will check until the end of the game after this action. The tests against recent successful ACPC agents were quite surprising, because even though when they played against themselves, the results were quite close, it is shown that they are extremely exploitable, even more so than if they had simply folded every hand!

Poker Abstractions

The three main ways to create a smaller game from a larger one are to merge information sets together (card abstraction) to restrict the actions a player can take (action abstraction), to use imperfect recall, or a combination of the three. Other possibilities are reducing the number of betting rounds or bets allowed per round or modifying the game itself, for example to use a smaller deck size or smaller starting chip size. In 2012, Lanctot developed theoretical bounds on bucketing abstraction that shows that CFR on abstractions leads to bounded regret in the full game. In 2014, this was expanded by Kroer and Sandholm to give bounds on solution quality for any perfect recall extensive form game. The framework can find abstractions, both lossy and lossless, given a specific bound requirement.

In the early 2000s, lossy abstractions were created by hand based on game-specific knowledge, but since then experimental advances in automation have advanced. The main ideas are:

1) Using integer programming to optimize the abstraction, usually within one level of the game at a time

2) Potential-aware abstraction, where information sets of a player at a given level are based on a probability vector of transition to state buckets at the next level

3) Imperfect recall abstraction, such that a player forgets some details that he knew earlier in the game

Most abstractions now use numbers 2 and 3 above and divide the game across a supercomputer for equilibrium finding computation.

Sandholm states that abstractions tend to be most successful when starting with a game theoretical strategy and modifying it based on the opponent.

Abstraction Translations

Abstraction translations are mappings from an opponent action to an action within the abstracted framework of the game. The goal of these reverse mappings, or translations, is to minimize the player’s exploitability, while also wanting one’s own abstraction to exploit other player abstraction choices.

Early abstraction translation methods included:

  • Deterministic arithmetic: Map to A or B that is closest to x. This is highly exploitable, for example, by strong hands that bet slightly closer to A and therefore can be responded to as if they were much weaker. This was used in the 2007 ACPC by Tartanian1 and lost to an agent that didn’t even look at its cards!
  • Randomized arithmetic: f_{A,B}(x) = (B-x)/(B-A). This is exploitable when facing bets close to 1/2 of the pot. This was used by AggroBot in 2006 in the ACPC.
  • Deterministic geometric: Map to A if A/x > x/B and B otherwise. This was used by Tartanian2 in the 2008 ACPC.
  • Randomized geometric 1: f_{A,B}(x) = A(B-x)/(A(B-x) + x(x-A)). This was used by Sartre and Hyperborean in the ACPC.
  • Randomized geometric 2: f_{A,B}(x) = A(B+x)(B-x)/((B-A)(x^2+AB)) was used by Tartanian4 in the 2010 ACPC.

In 2013, Ganzfried and Sandholm developed the following translation, as explained in the section on Abstractions.

f_{A,B}(x) = (B-x)(1+A)/((B-A)(1+x))

The new mapping was tested by “rematching” the agents from the 2012 ACPC and keeping the Tartanian5 agent the same, except for revising its mapping. The mapping performed well, but interestingly performed worse than the simple deterministic arithmetic mapping that simply maps the bet to the closest abstraction. However, this leaves agents open to exploitation that perhaps was not acted upon in prior years, but could be in the future.

Asymmetric Abstractions

The standard approach when abstracting is to use symmetric abstraction, assuming that all agents distinguish states in the same way. Agents in Texas Hold’em have been shown to perform better in head-to-head competitions and to be less exploitable when using finer-grained abstractions, although there are no theoretical guarantees of this. Until Bard et al’s 2014 paper, all research was done by examining only symmetric abstractions.

The choice of using asymmetric abstractions could affect both “one on one performance against other agents and exploitability in the unabstracted game”. By examining a number of different abstraction combinations in Texas Hold’em, a few main conclusions were drawn:

  • “With symmetric abstractions, increasing the abstraction size results in improved utility against other agents and improved exploitability”

  • “Smaller abstractions for ourselves, while our opponent uses larger abstractions, tended to result in improved exploitability, but decreased one on one mean utility”

  • “Larger abstractions for ourselves, while our opponent uses smaller abstractions, tends towards our exploitability worsening and our one on one utility improving, leading to the conclusion to want to increase abstractions when creating an agent for a one on one competition”

We see that the goals of minimizing exploitability and increasing one on one utility can be at odds with each other, so the agent designer must make abstraction decisions based on his goals and beliefs about other agents. A non poker example given is that if worst case outcomes resulted in people being injured or killed, the only goal may be to increase worst case performance.

Abstractions are Not Necessarily Monotonically Improving

Although logic would suggest that finer abstractions result in superior agents, this was shown to not be true necessarily by Waugh et al. in 2009. However, they did show that if one player is using abstraction while the other is playing in the unabstracted game, then the abstracted player’s strategies do monotonically improve as the abstractions get finer. As the annual poker competitions have advanced, empirically the winning strategies have generally been the teams that have solved the largest abstracted games.

An example with Leduc Hold’em is shown, in which a finer card abstraction can result in a more exploitable strategy. They also tested betting abstraction and again found instances in which exploitability increased as the abstraction became finer. One theory presented is that providing additional strategies to a player can encourage the player to “exploit the limitations of the opponent’s abstraction”, resulting in a strategy that is more exploitable by actions that become available to the opponent in the full game.

Earth Mover’s Distance Metric in Abstraction

In 2014, Ganzfried and Sandholm developed the leading abstraction algorithm for imperfect information games using k-means with the earth mover’s distance metric to cluster similar states together in Potential-Aware Imperfect-Recall Abstraction with Earth Mover’s Distance in Imperfect-Information Games.

Standard clustering techniques merge information sets together in the abstracted game according to hand strength. A common metric is equity against a random uniform hand, but this is problematic when the uniform hands have very different distributions against those given hands. The example given is that KQ suited is very strong against some hands and weak against some hands, while a pair of 66 has most of its weight in the middle, meaning it’s roughly 50% against most hands.

Distribution aware abstractions group states together at a given round if their full distributions over future strength are similar, instead of just basing this on the expectation of their strength. The goal of the earth mover’s algorithm is to build on this by not only looking at the future strength of the final round, but instead at trajectories of strength over all future rounds.

The recommended method for computing distances between histograms of hand strength distributions is the earth mover’s distance, which is the “minimum cost of turning one pile into the other where cost is assumed to be the amount of dirt moved times the distance by which it moved”. This metric accounts for both the amount and distance moved, not just the amount.

The algorithm incorporates imperfect recall and potential aware abstractions (where the maximum number of buckets in each round is determined based on the size of the final linear program to solve) using earth mover’s distance. We can find a case where two very different hands have similar equity distributions on the river, but are extremely different on the turn. The potential aware abstraction will observe this difference based on the earth mover’s distance and place them into different histograms on the turn. This is especially valuable when hand strengths can change significantly between rounds – the paper notes that Omaha Hold’em could be a good fit.

Simultaneous Abstraction and Equilibrium Finding

In a 2015 paper, Brown and Sandholm show a method to combine action abstraction and equilibrium finding together. An agent can start learning with a coarse abstraction and then can add abstracted actions that seem like they would be useful, which is determined by trying to minimize average overall regret. They showed that it converges to improved equilibrium solutions with no computational time loss.

CFR Extensions

In recent years, nearly all poker research has involved optimizations and extensions on counterfactual regret minimization. Here we highlight five interesting developments from the past few years.

Monte Carlo CFR

Marc Lanctot et al. of Alberta expanded the standard vanilla CFR method to include Monte Carlo sampling in 2009. The Monte Carlo methods are important because they converge faster than vanilla CFR, since vanilla CFR iterations require an entire tree traversal. Monte Carlo CFR sampling (MCCFR) requires more iterations, but each iteration tends to be much faster, resulting in a significantly faster convergence.

In Lanctot’s thesis, he goes beyond only poker and explores other interesting imperfect information games such as Latent Tic Tac Toe, a form of Tic Tac Toe in which players cannot see their opponents moves until after each round, and Princess and Monster, a pursuit-evasion game where two players are in a “dark” grid where they can’t see each other and can only move to points adjacent to their current locations. Scoring is based on catching the princess as fast as possible.

The primary game that Lanctot focuses on is called Bluff, where each player has private dice to roll and look at without showing to opponents. Each round, players alternate bidding on the outcome of all dice in play until one calls “bluff”, claiming that the bid is invalid. If caught, the losing player has to remove a number of dice between the amount he bid and the actual amount in the game. If a player loses all of his dice, he loses the game.

He shows that in these games and in poker, vanilla CFR is essentially always less efficient than Monte Carlo CFR methods. A method called external sampling also tends to dominate chance sampling in his experiments.

CFR+

In early 2015, the Computer Poker Research Group from the University of Alberta announced that “Heads-up limit hold’em poker is solved”, a huge achievement in the poker research community and poker community at large since this is the first significant imperfect-information game played competitively by humans that has been solved. Their article was published in the January 9, 2015 issue of Science and is discussed in detail in Section 4.4 CFR Advances. The team used a new version of CFR called CFR+, which can solve games much larger than with prior CFR algorithms and is also capable of converging faster than CFR in both poker games and matrix games.

The algorithm was originally developed by Oskar Tammelin, an independent researcher from Finland. This major achievement was noticed in the research community, the poker community, and was featured in many mainstream publications. Despite LHE’s popularity being in decline recently in favor of NLHE games, this was a very exciting announcement.

The main change in CFR+ from CFR and prior research is that the regrets are calculated differently such that all regrets are constrained to be non-negative, so that actions that have looked bad (i.e. those regrets with value <0) are chosen again immediately after proving useful instead of waiting many iterations to become positive. Also, the final strategy used is the current strategy at that time, not the average strategy (as in vanilla CFR), and no sampling is used. Additionally, advances were made enabling further use of compression to store the average strategy and regrets.

Previously only perfect information games of this size, like checkers, have been solved and despite this game being smaller than checkers, it was much more challenging due to the incomplete information. HULHE has 3.16x10^17 game states, which places it as larger than Connect 4 and smaller than checkers. There are 3.19x10^14 decision points, or information sets, where the state is indistinguishable based on the player’s information in the hand (this is reduced to 1.38x10^13 after removing game symmetries). The main considerations when solving a game of this size come in the form of memory and computation power.

CFR+ was implemented on 200 computation nodes, each with 24 2.1-GHz AMD cores, 32 GB of RAM, and a 1TB local hard disk. The solution came after 1,579 iterations in 69 days, using 900 core years of computation and 10.9 TB of disk space (the game without compression would have required 262 TB of space!), reaching an exploitability of .986 mbb/g, which required full traversal of the game tree to determine.

Exploitability is defined as the amount less than the game value that the strategy achieves against the worst-case opponent strategy in expectation and it was determined that a 1 mbb/g was a good threshold for the game being considered essentially solved. This is based on assuming a player playing a worst case strategy for a lifetime and that he would be playing 200 games/hour, 12 hours/day, for 70 years, with a standard deviation of 5 bb/g, and with a 95% confidence interval (1.64 standard deviations). This results in the following threshold computation: \(1.64 * 5 * 200 * 12 * 365 * 70 = .00105\)

Bowling et al. describe their solution as weakly solved, which means that the strategy finds a game theoretic value in reasonable time for initial conditions (whereas ultraweakly is finding a game theoretic value for initial positions and strongly is strategy determined for all positions to find the game theoretic value).

The main findings from the solution are that, as expected, the dealer has a substantial advantage because he acts last, giving him more information to work with while making decisions. He is expected to win +87.7-89.7 mbb/g.

Other interesting characteristics of the strategy are: that it is considered rarely (.06%) good to only call preflop, meaning that it is almost always best to raise or fold. The program also rarely folds preflop as the non dealer who has already put in the big blind. Finally, as dealer, who is acting last in rounds after preflop, the program rarely puts in the capped (4th) bet preflop, perhaps because it will have an advantage from acting last in later rounds and so keeping the pot smaller while less information is known may make sense. While other Nash equilibria could play differently, they would always achieve the same game value.

Despite the significant growth in capabilities of solving computer poker games, NLHE games are generally still much too large to be solved unabstracted. NLHE usually has about 10^71 states, depending on the rules. The Royal NLHE simplified game has about 10^9 game states, solvable by most standard computers.

Compact CFR and Pure CFR

Although most CFR research had been tilted towards sampling optimization after 2009, Oskari Tammelin also developed the Pure CFR algorithm that uses pure strategy profiles on the vanilla version of CFR.

Pure CFR was described in Richard Gibson’s 2014 thesis. Since all of the utilities of in poker are integers, all computations in Pure CFR can be done with integer arithmetic, which is both faster than floating-point arithmetic and allows for the cumulative regret and cumulative strategy profile as integers, which reduces memory costs by 50%.

Another CFR version was published in early 2016 by Eric Jackson in a paper called Compact CFR. This version uses follow-the-leader instead of regret matching, which assigns the entire strategy probability at a node to the action with highest regret. Compact CFR is not a no-regret algorithm (i.e. it loses theoretical guarantees), but the average strategy does converge to Nash equilibrium.

Since all we need to know in Compact CFR is the action with the highest regret at each information set, the regrets can be represented with offsets from 0, where the highest regret is 0 and others are positive values. These can be represented by bucketed unsigned data types, a form of compression that is only slightly worse than uncompressed regret storage. By also only taking the final strategy rather than the average strategy, every action at every information set can be represented by only one byte, whereas vanilla CFR requires 16 bytes and pure external CFR requires 8 bytes, a significant reduction in memory requirements.

Strategy Purification and Thresholding

In 2012, Ganzfried et al proposed that instead of solving abstract games for an equilibrium’s strategy and using this strategy translated into the full game, we can first modify the abstract game equilibrium using procedures called purification and thresholding. The main idea is that these approaches provide a “robustness to the solutions against overfitting one’s strategy to one’s lossy abstraction”, and the results “do not always come at the expense of worst-case exploitability”.

Purification means that the player will always play his best strategy at each information set with probability 1 (rather than play an action based on his behavioral strategy distribution). This has similarities to Compact CFR from above. In the case of ties, both strategies are played with uniform probability. This is useful because it compensates for the failure of equilibrium finding algorithms to fully converge in the abstract game.

Thresholding is a more relaxed approach, which simply eliminates certain actions with probabilities below a threshold value \(\epsilon\), and then renormalizes the action probabilities. Logic given is that these strategies may either be due to noise or are played primarily to protect a player from being exploited, which may be an overstated issue about realistic opponents. Humans in practice generally use thresholding when implementing a poker strategy because it isn’t practical for humans to properly randomize with so many probabilities, and so it makes sense to remove the very lowest ones to simplify.

A study was done using the game Leduc Hold’em and for almost all abstractions, purification was shown to bring a significant improvement in exploitability. Thresholding was also beneficial, but purification always performed better in the cases that it improved exploitability.

Identical agents except for using either thresholding or purification were submitted to the ACPC and the purification agent performed better against all opponents, including the thresholding agent. In terms of worst case exploitability, the least exploitable was the one that used a thresholding level of 0.15 (meaning all actions below 15% were removed), perhaps because too much thresholding results in too little randomness and no thresholding at all results in overfitting to the abstraction.

Decomposition

Decomposition, analyzing different subgames independently, has been a well known principle in perfect information games, but has been problematic in imperfect information games and when used has abandoned theoretical guarantees. In 2014, Burch, Johanson, and Bowling proposed a technique that does “retain optimality guarantees on the full game”. In perfect information games, “subgames can be solved independently and the strategy fragments created can be combined to form an optimal strategy for the entire game”.

Decomposition can allow large savings in the memory required to solve a game and also allows for not storing the complete strategy, which may be too large, but rather to store and recomputed subgame strategies as needed.

The authors show a method of using summary information about a subgame strategy to generate a new strategy that is no more exploitable than the original strategy. An algorithm called CFR-D, for decomposition, is shown to achieve “sub-linear space costs at the cost of increased computation time”.

Endgame Solving in Large Imperfect-Information Games

While the naive approach simply abstracts a game and uses translation to go back to the full game, more recent approaches tended to divide games into sequential phases, and now Ganzfried and Sandholm show how to solve the endgame specifically with a finer abstraction.

In 2015, Ganzfried and Sandholm modified the standard CFR abstraction solution method by keeping the initial portion of the game tree and discarding the strategies for the final portion, the endgames. Then in real time, they solve the relevant endgame that has been reached using a linear program, with a greater degree of accuracy than the initial abstract strategy. Bayes’ rule is used to compute the distribution of player private information leading into the endgames from the precomputed strategies from the initial part of the game.

Another benefit of this method is that “off-tree” problems are solved – that is, cases in which the opponent’s action is not allowed in the abstraction will actually be solved exactly in the endgame.

The problem with endgame solving is that the Nash equilibrium guarantees are no longer valid. Combining an equilibrium from the early part of the game and end of the game could lead to mismatched equilibrium since games tend to have multiple different ones.

Although endgame solving can lead to highly exploitable strategies in some games, it’s shown to have significant benefits in large imperfect information games, especially games where the endgame has is important, which is the case in all poker games, and especially no limit hold’em.

This technique showed improved performance against the strongest agents from the 2013 ACPC. Having better abstractions at the endgame seems intuitively very valuable since this tends to be where it’s possible to narrow down opponent hand ranges and where the largest bets tend to be made, since the pot gradually builds up over the betting rounds.

In 2017, Brown and Sandholm advanced the previous methods by using nested endgame solving in their Libratus agent in place of action translation in response to off-tree opponent actions. This may have made the difference in defeating world class human opponents.

Warm Starting and Regret Pruning

Also in 2015, Brown and Sandholm found in Strategy-Based Warm Starting for Regret Minimization Games that it is possible to warm start in CFR by using a predetermined strategy and that with a single full traversal of the game tree, CFR is effectively warm started to as many iterations as it would have taken to reach a strategy profile of the same quality as the input strategies, and the convergence bounds are unchanged. By warm starting, CFR can bypass spending time analyzing early and expensive iterations that contain all nodes, even never-used ones that would be pruned.

Brown and Sandholm developed a regret-based pruning (RBP) method in 2015 to prune actions with negative regret temporarily (for the minimum number of iterations that it would take for the regret to become positive in CFR) in the paper Regret-Based Pruning in Extensive-Form Games.

This process was shown to speed up CFR and then in 2016, they improved this with a new RBP version that can even reduce the space requirements of CFR over time by completely discarding pruned branches.

Deep Learning

Beginning in 2017 with the University of Alberta’s DeepStack, poker algorithms have been relying on deep neural networks as an alternative or complement to game abstractions. Noam Brown et al published Deep Counterfactual Regret Minimization in 2018 that was was the first algorithm to successfully implement CFR with deep neural networks rather than the standard tabular format. It takes as input the exact cards and betting sequences to a neural network and outputs values in proportion to the regrets that would be found in tabular CFR. The technique effectively leaves the neural network to do the abstracting rather than requiring fixed abstractions built into the algorithm. We discuss this result in detail in the CFR Advances section. As computer processing capabilities have improved, using large neural networks becomes an increasingly valuable method for approximately solving poker games.

Superhuman AI for multiplayer poker

In 2019, Noam Brown and Tuomas Sandholm of Carnegie Mellon University released this paper with an agent called Pluribus that beat strong human players in six-handed poker. The agent was able to outperform the humans in two settings: (1) five agents at the table with one human and (2) 5 humans at the table with one agent.

Because there isn’t a clear game-theoretic solution to a multiplayer game, the goal was to create an agent that could empirically outperform top human players. This is detailed in the Top Agents and Research section.

Computer Poker Competitions

The Annual Computer Poker Competition

In 2006, the University of Alberta and Carnegie Mellon University jointly founded the Annual Computer Poker Competition (ACPC), which is now held at the Poker Workshop of the annual Advancement of Artificial Intelligence (AAAI) conference during the Poker Workshop part of the conference, which began in 2012.

The competition has attracted both hobbyists and academics from around the world each year, although a complaint is that supercomputer access that is possible for academics may not be feasible for hobbyists. Each match is played in either HUNL or HULHE consists of 3000 hands where each player starts with 200 big blinds, using blind sizes of 50 and 100 and a minimum bet of 1 chip. Matches are played in duplicate so that the same cards are given to each agent, then memories are cleared, and the same match is played again with cards being dealt to the opposite player.

There are two primary competition types. The first is instant-run off, where agents earn or lose 1 point for each match’s win and loss, which favors small wins and equilibrium solutions. The total bankroll competition type counts the total winnings of each agent over all of its contests. This favors agents that are more exploitative.

The general strategy in recent years has been for teams to develop algorithms that allow for larger and larger games to be solved, meaning that finer grained abstractions can be used, which have generally correlated to stronger performance, though such a result is not theoretically guaranteed, and counterexamples do exist.

MIT now even holds a mini-course called MIT Pokerbots in which students are given one month to program autonomous pokerbots to compete against other teams in a tournament. The contest receives heavy sponsorship from trading companies due to the similarities in dealing with imperfect information and decisions based on probabilities and statistics.

CFR and its variants are have been the most common approach used in the ACPC recently. It was first seen in 2007 with Zinkevich and the University of Alberta CPRG, using imperfect recall abstraction. CFR was used in 2/11 agents in 2012, 5/12 in 2013, 10/12 in 2014 (there was no competition in 2015 and the details of the competitors for more recent competition have not been released, except for the winners). In 2013, 2014, and 2016, the top three agents in the bankroll and instant run-off competitions all used some form of CFR.

Alberta’s Hyperborean won every limit hold’em run-off competition from 2006 to 2008 using an imperfect recall abstraction in CFR. In 2009 it was defeated by GGValuta from the University of Bucharest, which also used CFR with a k-means clustering algorithm to do card abstraction.

Hyperborean again won in 2009 by fusing independent separate solutions from independent pieces of the game tree. Hyperborean also won the 2010 no limit run-off event.

Hyperborean, Slumbot by Eric Jackson, and Tartanian from Carnegie Mellon have consistently had excellent results since this time.

We will now briefly go over some of the winners and their techniques from the most recent competitions.

The 2014 winner was the Tartanian7 team from Carnegie Mellon. The program plays approximately a Nash equilibrium strategy that was computed on a supercomputer. They developed a new abstraction algorithm that clusters public flop boards based on how often their previous program grouped private hands together on the flop with different sets of public cards. Within each of the public flop clusters, the algorithm then buckets the flop, turn, and river hands that are possible given one of the public flops in the cluster, using imperfect-recall abstraction. They did not do any abstraction for the preflop round.

They based their equilibrium finding algorithm on external sampling MCCFR by sampling one pair of preflop hands per iteration. Postflop, they sample community cards from their public clusters and MCCFR in parallel, and add weights to the samples to remove bias. They also used thresholding and purification and made the interesting observation that it is valuable to bias towards conservative actions to reduce variance, since higher variance means that the inferior opponent is more likely to win.

Slumbot, by Eric Jackson, uses Pure External CFR for equilibrium computation. He breaks the game tree into pieces to be solved separately and uses differing abstraction levels depending on how often the game states are reached. Those more common ones are given more granularity in both bucketing and bet sizes possible.

Hyperborean (for the auto run-off competition), made by the Computer Poker Games Research Group from the University of Alberta, also uses Pure CFR, imperfect recall, and k-means card bucketing abstraction. They interestingly use the final strategy of the algorithm rather than the average strategy, which is the one proven to converge to equilibrium. They use an asymmetric betting system in which the opponent can have more options than the agent, including such actions as minimum-betting.

For the total bankroll competition, the program uses three distinct strategies and chooses one according to an algorithm. Two of the strategies are data-biased responses to aggregate data of competitors from the years 2011/12 and 2013, and the other strategy is similar to the auto run-off competition strategy, but also separates betting sequences into an “important” and “unimportant” part and creates more buckets for the important sequences. Importance is based on how often these sequences are seen in self-play.

In 2016, only the no-limit hold’em total-bankroll and instant run-off competitions took place. The same three teams took the top three places in both competitions.

Unfold Poker was trained by a distributed implementation of the Pure CFR algorithm and uses a heuristic to sometimes avoid certain game tree paths. Certain bet sizes were omitted and a distance metric that considers features from all postflop betting streets was used to construct the card abstraction on the river. Unfold Poker took 2nd place in the instant run-off event and 3rd in the total bankroll event.

Slumbot took 1st in the instant run-off and 2nd in the total bankroll by using a new memory efficient CFR technique called Compact CFR, which was detailed in the section on CFR above.

Finally, Carnegie Mellon University’s Baby Tartanian 8 won the total bankroll competition and took 3rd place in the instant run-off event. Baby Tartanian 8’s main new feature was to add pruning to cut down actions worth considering. They also took feedback from the 2015 man vs. machine competition in which Tartanian 7 suffered a loss to improve translations and do better endgame solving.

Going forward, the event will feature six player games, which will accelerate research towards games that are commonly played by humans and present a new set of complexities, including whether the optimal approach is to aim for opponent exploitation or to continue on the unexploitable path, despite multiplayer games invalidating theoretical results that would be valid in zero-sum games. Based on the recent results from Pluribus, it would seem that the former works well.

The last workshop with the ACPC competition took place in 2018 with heads up no limit Texas Hold’em and six player no limit Texas Hold’em. Teams now have a maximum submission size, which is a way to even the field between teams that might have more resources. Although there have been breakthroughs in recent years with Pluribus and other top agents, it could be interesting to see even better multiplayer agents. However, it does seem that many academic groups have moved on from poker and hobbyists may prefer to use their algorithms commercially or privately, so the future of the ACPC remains to be seen.

Updated: