Computer Poker: a Review ∗ Jonathan Rubin, Ian Watson
Total Page:16
File Type:pdf, Size:1020Kb
JID:ARTINT AID:2578 /REV [m3G; v 1.51; Prn:28/01/2011; 9:50] P.1 (1-30) Artificial Intelligence ••• (••••) •••–••• Contents lists available at ScienceDirect Artificial Intelligence www.elsevier.com/locate/artint Computer poker: A review ∗ Jonathan Rubin, Ian Watson Department of Computer Science, University of Auckland, New Zealand article info abstract Article history: The game of poker has been identified as a beneficial domain for current AI research Received 19 February 2010 because of the properties it possesses such as the need to deal with hidden information Accepted 21 December 2010 and stochasticity. The identification of poker as a useful research domain has inevitably Available online xxxx resulted in increased attention from academic researchers who have pursued many separate avenues of research in the area of computer poker. The poker domain has Keywords: Computer poker often featured in previous review papers that focus on games in general, however a Imperfect information games comprehensive review paper with a specific focus on computer poker has so far been Nash equilibrium lacking in the literature. In this paper, we present a review of recent algorithms and Computational game theory approaches in the area of computer poker, along with a survey of the autonomous Opponent modelling poker agents that have resulted from this research. We begin with the first serious attempts to create strong computerised poker players by constructing knowledge-based andsimulation-basedsystems.Thisisfollowedbytheuseofcomputationalgametheory to construct robust poker agents and the advances that have been made in this area. Approaches to constructing exploitive agents are reviewed and the challenging problems of creating accurate and dynamic opponent models are addressed. Finally, we conclude with a selection of alternative approaches that have received attention in previously published material and the interesting problems that they pose. © 2011 Elsevier B.V. All rights reserved. 1. Introduction Many games have successfully (and enjoyably) been used as domains for Artificial Intelligence (AI) research over the years. Early efforts for games such as chess, checkers, backgammon, Connect-4, Othello, Lines of Action and Scrabble have produced strong computerised players and notable success stories. More recently, games such as bridge, Hex, Go and poker have increasingly been the focus of AI research, so too have modern interactive computer (video) games such as first- person shooters, role-playing games and real-time strategy games. Jonathan Schaeffer states in his 2001 article, A Gamut of Games [81]: Successfully achieving high computer performance in a non-trivial game can be a stepping stone toward solving more challenging real-world problems. The use of games as research domains has also played a significant role in the proliferation of competitions that guide modern AI related research. Beginning with the highly publicised human vs. machine chess matches between Kasparov and Deep Blue in the 1990s, these days all sorts of competitions exist for many different games. The International Computer * Corresponding author. E-mail addresses: [email protected] (J. Rubin), [email protected] (I. Watson). URL: http://www.cs.auckland.ac.nz/research/gameai (I. Watson). 0004-3702/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.artint.2010.12.005 JID:ARTINT AID:2578 /REV [m3G; v 1.51; Prn:28/01/2011; 9:50] P.2 (1-30) 2 J. Rubin, I. Watson / Artificial Intelligence ••• (••••) •••–••• Games Association [39] presently organises regular Computer Olympiads, where competitions are held for board and card games such as chess, checkers, bridge and Go. RoboCup [73] is an annual competition for both simulated and robotic soccer matches. There are also annual competitions for poker [1], real-time strategy games [19] and general game playing [29]. For the above domains, competitions are successfully used to drive research and bring about improvements to the state-of-the- art. Each of the non-trivial games mentioned above share common beneficial properties such as well defined rules over possi- ble actions and behaviours, the existence of clear goals and objectives, as well as embedded performance metrics. Each par- ticular game also has its own unique, beneficial properties that make them good research areas. The game of poker has been identified as a beneficial domain for current AI research [16] due to properties of the game such as imperfect information – the inability to observe opponents’ cards and chance events – the random private and public cards that are dealt each hand. The popularity of poker experienced a massive boom in the early 2000s due to a combination of factors such as the emergence of online poker, the introduction of hole card cameras to televised events, as well as an amateur’s unlikely win at the main event of the 2003 World Series of Poker. The use of poker as an academic research domain has undergone a similar increase in popularity. The identification of poker as a useful research domain has inevitably resulted in increased attention from academic researchers and has led to the investigation of a wide range of new algorithms and approaches. In particular, work done by the University of Alberta’s Computer Poker Research Group1 (CPRG) and the establishment of annual computer poker competitions at conferences such as AAAI and IJCAI since 2006, has resulted in the development of a large number of computerised poker agents. However, as yet a comprehensive review of the algorithms and approaches used to construct these numerous poker agents has been lacking. Many past reviews have focused on games in general. In The Games Computers (and People) Play, Schaeffer [80] provides a history of program development and case-studies for classic board and card games including backgammon, bridge, checkers, chess, Othello, poker and Scrabble. Schaeffer identifies poker as a game where human supremacy may soon be challenged and highlights the use of Monte-Carlo simulations within the poker domain due to an inability to directly apply alpha-beta search. Fürnkranz [27] adopts a problem-oriented approach in his survey of machine learning approaches in game playing. The problem of opponent modelling is identified as a somewhat neglected area of research within game playing domains and its importance in a game such as poker is highlighted. More recently (and more specific to the poker domain), Darse Billings’ PhD thesis [8] provides a comprehensive descrip- tion and analysis regarding the evolution of computer poker algorithms from 1997 to 2005 that he and his colleagues at the University of Alberta CPRG have investigated. While many of the approaches discussed in [8] are also addressed in this work, Billings’ obvious focus is on the personal contribution made by himself and his colleagues, whereas this document incorporates previously published work from other researchers that has further extended and evaluated the original ideas presented in [8], along with further advances that have occurred over recent years. As the vast majority of published work focuses on one particular variant of poker – Texas Hold’em, this will be the sole focus of this review paper. Furthermore, previous computer poker research can roughly be split into two broad categories [16]. The first attempts to analyse smaller isolated subsets of the problem domain which typically leads to valuable insights about specialised aspects of the game. The second category, on the other hand, attempts to tackle the domain as a whole, resulting in the creation of an agent that is able to play the game autonomously. In this paper we will mostly restrict our attention to the second category, i.e. published work that attempts to tackle the full game of Texas Hold’em by producing autonomous poker playing agents that have been evaluated by challenging human or computerised opposition. Each approach presented within this document will roughly follow the same pattern. First, the approach being reviewed is introduced, typically followed by an example. Next, a survey of the autonomous agents that have employed the above approach to play Texas Hold’em is presented, along with a discussion of the agent’s performance. Section 2 begins with a review of early work on knowledge-based poker agents, followed by Section 3 which introduces Monte-Carlo simulation. Section 4 focuses on -Nash equilibrium based agents produced via both linear programming and state-of-the-art iterative algorithms. Section 5 reviews exploitive agents that attempt to adapt to their opponents’ playing style by constructing accurate opponent models and Section 6 briefly reviews some alternative approaches that have received attention in the literature, such as the use of case-based approaches, Bayesian poker and evolutionary algorithms. Finally, Section 7 concludes this review. Before reviewing the avenues of research that have been investigated in the area of computer poker, the remainder of this section presents a description of the rules of Texas Hold’em, followed by a brief overview of the various performance metrics and evaluators mentioned throughout the document. 1.1. Texas Hold’em Here we briefly describe the game of Texas Hold’em, highlighting some of the common terms which are used throughout this work. For more detailed information on Texas Hold’em consult [90], or for further information on poker in general see [89]. 1 http://poker.cs.ualberta.ca/. JID:ARTINT AID:2578 /REV [m3G; v 1.51; Prn:28/01/2011; 9:50] P.3 (1-30) J. Rubin, I. Watson / Artificial Intelligence ••• (••••) •••–••• 3 The game of Texas Hold’em is played in 4 stages – preflop, flop, turn and river. During the preflop all players at the table are dealt two hole cards, which only they can see. Before any betting takes place, two forced bets are contributed to the pot, i.e. the small blind and the big blind. The big blind is typically double that of the small blind. The player to the left of the big blind, known as under the gun, then begins the betting by either folding, calling or raising.