Complex games have long been an important benchmark for testing the progress of artificial intelligence algorithms. AlphaGo, AlphaZero, and MuZero have defeated top human players in Go and Chess, garnering widespread societal attention towards artificial intelligence. Concurrently, large language models (LLMs) have exhibited remarkable capabilities across various tasks, raising the question of whether LLMs can achieve similar success in complex games. In this paper, we explore the potential of LLMs in mastering complex card games. We systematically assess the learning capabilities of LLMs across eight diverse card games, evaluating the impact of fine-tuning on high-quality gameplay data, and examining the models' ability to retain general capabilities while mastering these games. Our findings indicate that: (1) LLMs can approach the performance of strong game AIs through supervised fine-tuning on high-quality data, (2) LLMs can achieve a certain level of proficiency in multiple complex card games simultaneously, with performance augmentation for games with similar rules and conflicts for dissimilar ones, and (3) LLMs experience a decline in general capabilities when mastering complex games, but this decline can be mitigated by integrating a certain amount of general instruction data. The evaluation results demonstrate strong learning ability and versatility of LLMs. The code is available at https://github.com/THUDM/

Complex games have long been an important benchmark for testing the progress of artificial intelligence algorithms. AlphaGo, AlphaZero, and MuZero have defeated top human players in Go and Chess, garnering widespread societal attention towards artificial intelligence. Concurrently, large language models (LLMs) have exhibited remarkable capabilities across various tasks, raising the question of whether LLMs can achieve similar success in complex games. In this paper, we explore the potential of LLMs in mastering complex card games. We systematically assess the learning capabilities of LLMs across eight diverse card games, evaluating the impact of fine-tuning on high-quality gameplay data, and examining the models' ability to retain general capabilities while mastering these games. Our findings indicate that: (1) LLMs can approach the performance of strong game AIs through supervised fine-tuning on high-quality data, (2) LLMs can achieve a certain level of proficiency in multiple complex card games simultaneously, with performance augmentation for games with similar rules and conflicts for dissimilar ones, and (3) LLMs experience a decline in general capabilities when mastering complex games, but this decline can be mitigated by integrating a certain amount of general instruction data. The evaluation results demonstrate strong learning ability and versatility of LLMs. The code is available at https://github.com/THUDM/LLM4CardGame

I recently started a new newsletter focus on AI education. TheSequence is a no-BS( meaning no hype, no news etc) AI-focused newsletter that takes 5 minutes to read. The goal is to keep you up to date with machine learning projects, research papers and concepts. Poker has been considered by many the core inspiration for the formalization of game theory. John von Neuman was reportedly an avid poker fan and use many analogies of the card game while creating the foundation of game-theory.

Regret minimization has played a key role in online learning, equilibrium computation in games, and reinforcement learning (RL). In this paper, we describe a general model-free RL method for no-regret learning based on repeated reconsideration of past behavior. We propose a model-free RL algorithm, the AdvantageRegret-Matching Actor-Critic (ARMAC): rather than saving past state-action data, ARMAC saves a buffer of past policies, replaying through them to reconstruct hindsight assessments of past behavior. These retrospective value estimates are used to predict conditional advantages which, combined with regret matching, produces a new policy. In particular, ARMAC learns from sampled trajectories in a centralized training setting, without requiring the application of importance sampling commonly used in Monte Carlo counterfactual regret (CFR) minimization; hence, it does not suffer from excessive variance in large environments. In the single-agent setting, ARMAC shows an interesting form of exploration by keeping past policies intact. In the multiagent setting, ARMAC in self-play approaches Nash equilibria on some partially-observable zero-sum benchmarks. We provide exploitability estimates in the significantly larger game of betting-abstracted no-limit Texas Hold'em.

Facebook researchers have developed a general AI framework called Recursive Belief-based Learning (ReBeL) that they say achieves better-than-human performance in heads-up, no-limit Texas hold'em poker while using less domain knowledge than any prior poker AI. They assert that ReBeL is a step toward developing universal techniques for multi-agent interactions -- in other words, general algorithms that can be deployed in large-scale, multi-agent settings. Potential applications run the gamut from auctions, negotiations, and cybersecurity to self-driving cars and trucks. Combining reinforcement learning with search at AI model training and test time has led to a number of advances. Reinforcement learning is where agents learn to achieve goals by maximizing rewards, while search is the process of navigating from a start to a goal state.

Most chess computers play a purely mathematical strategy in a game yet to be solved. They are raw calculators and look like it too. AlphaZero, at least in style, appears to play every bit like a human. It makes long-term positional plays as if it can visualize the board; spectacular piece sacrifices that no computer could ever possibly pull off, and exploitative exchanges that would make a computer, if it were able, cringe with complexity. In short, AlphaZero is a genuine intelligence.

Computer programs have shown superiority over humans in two-player games such as chess, Go, and heads-up, no-limit Texas hold'em poker. However, poker games usually include six players--a much trickier challenge for artificial intelligence than the two-player variant. Brown and Sandholm developed a program, dubbed Pluribus, that learned how to play six-player no-limit Texas hold'em by playing against five copies of itself (see the Perspective by Blair and Saffidine). When pitted against five elite professional poker players, or with five copies of Pluribus playing against one professional, the computer performed significantly better than humans over the course of 10,000 hands of poker. Science, this issue p. 885; see also p. 864

Pluribus is the first AI bot capable of beating human experts in six-player no-limit Hold'em, the most widely played poker format in the world. This is the first time an AI bot has beaten top human players in a complex game with more than two players or two teams. We tested Pluribus against professional poker players, including two winners of the World Series of Poker Main Event. Pluribus succeeds because it can very efficiently handle the challenges of a game with both hidden information and more than two players. It uses self-play to teach itself how to win, with no examples or guidance on strategy. Pluribus uses far fewer computing resources than the bots that have defeated humans in other games. The bot's success will advance AI research, because many important AI challenges involve many players and hidden information. For decades, poker has been a difficult and important grand challenge problem for the field of AI. Because poker involves hidden information -- you don't know your opponents' cards -- success requires bluffing and other strategies that do not apply to chess, Go, and other games.

In recent years there have been great strides in artificial intelligence (AI), with games often serving as challenge problems, benchmarks, and milestones for progress. Poker has served for decades as such a challenge problem. Past successes in such benchmarks, including poker, have been limited to two-player games. However, poker in particular is traditionally played with more than two players. Multiplayer games present fundamental additional issues beyond those in two-player games, and multiplayer poker is a recognized AI milestone. In this paper we present Pluribus, an AI that we show is stronger than top human professionals in six-player no-limit Texas hold'em poker, the most popular form of poker played by humans. Poker has served as a challenge problem for the fields of artificial intelligence (AI) and game theory for decades (1). In fact, the foundational papers on game theory used poker to illustrate their concepts (2, 3). The reason for this choice is simple: no other popular recreational game captures the challenges of hidden information as effectively and as elegantly as poker. Although poker has been useful as a benchmark for new AI and game-theoretic techniques, the challenge of hidden information in strategic settings is not limited to recreational games.

Computer scientists have developed a card-playing bot, called Pluribus, capable of defeating some of the world's best players at six-person no-limit Texas hold'em poker, in what's considered an important breakthrough in artificial intelligence. Two years ago, a research team from Carnegie Mellon University developed a similar poker-playing system, called Libratus, which consistently defeated the world's best players at one-on-one Heads-Up, No-Limit Texas Hold'em poker. The creators of Libratus, Tuomas Sandholm and Noam Brown, have now upped the stakes, unveiling a new system capable of playing six-player no-limit Texas hold'em poker, a wildly popular version of the game. In a series of contests, Pluribus handedly defeated its professional human opponents, at a level the researchers described as "superhuman." When pitted against professional human opponents with real money involved, Pluribus managed to collect winnings at an astounding rate of $1,000 per hour.