General game playing

General game playing algorithms are typically restrained to games with finite possible moves that take place synchronously between players. These games are described in terms of a start state, possible intermediary states, and a goal.

The first commercial usage of general game playing technology was Zillions of Games in 1998. General game playing was also proposed for trading agents in supply chain management there under price negotiation in online auctions from 2003 onwards.

History

Origins (1958–1990s)

Researchers at the RAND Corporation first described a General Problem Solver, which they called "General Problem Solving Program I" (GPS), in 1958. The goal of their project was to solve problems, which they defined broadly as "whenever a problem solver desires some outcome or state of affairs that he does not immediately know how to attain." The RAND group emphasized the importance of heuristic methods in problem solving, and described a theoretically ideal problem-solving process as a cycle of discovery, information, and conclusion through educated guesswork.

RAND's GPS algorithm grew out of the more specialized "Logic Theorist" algorithm, which produced proofs in propositional logic. Like the Logic Theorist, GPS relied on the creation of intermediary goals through means–ends analysis. In particular, the algorithm was capable of solving Symbolic Logic problems by using heuristic methods which the researchers derived from the problem solving strategies of actual human test subjects. Central to the model's mechanisms for simplification was the method of abstraction, in which the logical elements were replaced by more basic elements which functioned similarly in order to simplify problems. Although the algorithm is theoretically capable of solving any sufficiently formalized problem in the form of well-formed formulae and directed graphs, in practice the amount of computation required for the algorithm to adequately explore the graph is infeasible to implement due to the combinatorial explosion.

In 1968, Jacques Pitrat founded the field of General Game Playing in a paper in the proceedings of the International Federation for Information Processing. Pitrat later worked on the more specialized problem of chess-playing algorithms, but his earliest contribution to game playing algorithm development was to separate the subclass of general game players from the broader space of general problem solvers. Pitrat's game playing algorithm was theoretically capable of playing chess-like games, which could be described using a simple set of rule variations.

Formalization and competition (1992–2016)

In 1992, Barney Pell defined the concept of Meta-Game Playing and developed the "MetaGame" system. This was the first program to automatically generate chess-like game rules, and one of the earliest programs to use automated game generation. Pell then developed the system Metagamer. This system was able to play a number of chess-like games without any human interaction once the games were generated.

In 1998, the commercial system Zillions of Games was developed by Jeff Mallett and Mark Lefler. The system used a LISP-like language to define the game rules. Zillions of Games derived the evaluation function automatically from the game rules based on piece mobility, board structure and game goals. It also employed usual algorithms as found in computer chess systems: alpha–beta pruning with move ordering, transposition tables, etc. The package was extended in 2007 by the addition of the Axiom plug-in, an alternate metagame engine that incorporates a complete Forth-based programming language.

In 1998, z-Tree was developed by Urs Fischbacher. z-Tree is the first and the most cited software tool for experimental economics. z-Tree allows the definition of game rules in z-Tree-language for game-theoretic experiments with human subjects. It also allows definition of computer players, which participate in a play with human subjects.

In 2005, the Stanford Project General Game Playing was established. This included the General Game Playing Competition, which was launched at the associated AAAI conference. The competition ran annually and further establish GGP as a research field. It was suspended after 2016, by which point it was considered to have largely served its original purpose of stimulating increased research in the area. The Stanford project also produced Game Description Language (GDL), which allowed for the production and description of game variations and was used extensively in the General Game Playing Competition.

In 2012, the development of PyVGDL (Python Video Game Description Language) started, and has now migrated to PyVGDL 2.0. PyVGDL is a library made for the programming language Python that functions as a game engine, enabling the integration of various AI applications in tandem with simple game rules to create complex games. The library is frequently used in artificial intelligence research centered around General Game Playing problems due to its dynamic capabilities.

Deep reinforcement learning

In 2018, Google DeepMind released a paper describing their new AlphaZero algorithm. The algorithm built on their earlier game-specific work with AlphaGo, and used a Monte Carlo tree search algorithm. AlphaZero met and even surpassed the performance of specialized game playing algorithms, and reviewers characterized it as a notable step toward general game-playing systems.

In 2020, the DeepMind group that had produced AlphaZero published their work on the MuZero algorithm.While AlphaZero required knowledge of the rules of games before playing, MuZero was able to match its performance without being given any prior information on the underlying structure of the game it was playing. The authors of the original MuZero paper noted that the lack of a need for pre-set rules potentially made the algorithm more applicable to use in real-world situations, in which there are no known governing principles or underlying rules.

New description languages (2019-2020)

Later research efforts produced new game-description languages which addressed the limitations of GDL. In 2019, researchers at the University of Wrocław created Regular Boardgames (RBG), a computer language based on the theory of regular languages. RBG encodes board game rules as regular expressions, and was designed for expressiveness, naturalness, and efficiency. Although built for usage in board games specifically, it is universally capable of describing any finite deterministic turn-based game with perfect information. The Wrocław group's paper reported that it was the first GGP language able to efficiently encode and play games with complex rules and large numbers of legal moves in any given turn, such as Arimaa. In 2020, Éric Piette, Cameron Browne and colleagues introduced Ludii, a general game system built on "ludemes" — units of game-related information — as part of the Digital Ludeme Project. Ludii was designed to model a wide range of games, including ones which are no longer actively played, and to study their structure and history. Ludii matched the performance of previous descriptive schemes such as RBG and, because of its high level structure, produced game descriptions that were generally more human-readable compared to other systems.

Large Language Models (2020s)

In the early 2020s, researchers started evaluating large language models (LLMs) as general game players, testing their ability to reason strategically and make choices in games they were not specifically trained on. In 2024, Oguzhan Topsakal and colleagues introduced an open-source benchmark which could be used to compare the performance of LLMs in grid-based games such as Tic-Tac-Toe, Connect Four, and Gomoku. This benchmark made use of several different methods of describing game states to the models, comparing performance across methodologies. In 2025, a team from the University of California, San Diego, Mohamed bin Zayed University of Artificial Intelligence, and the University of California, Berkeley introduced lmgame-Bench, which evaluated models' performance across a wide range of video game types. This work emphasized that directly testing LLMs on games is difficult, identifying problems such as brittle visual perception, sensitivity to prompt wording, and the presence of preexisting gameplay strategy in training data.