What is General Game Playing (GGP)?
General Game Playing (GGP) is a subfield of AI focused on building agents that can play a wide range of games competently without being designed or trained for any single one in advance. A general game player receives a formal description of a new game's rules — often written in the Game Description Language (GDL) — only at the start of a match, and must parse those rules, reason about legal moves and win conditions, and play skillfully with no prior game-specific knowledge or human tuning.
This distinguishes GGP from systems built to master one game: a chess engine tuned for chess, or a Go-playing system trained on Go, does not qualify as a general game player even if it plays that game at a superhuman level. The field grew out of the AAAI General Game Playing Competition, first held in 2005, which pits GDL-based agents against each other on games unknown to their developers beforehand.
What are the benefits of GGP?
- Tests genuine general-purpose reasoning rather than performance memorized or tuned for one game
- Provides a controlled benchmark for comparing search and reasoning strategies across many game types
- Offers insight into transferable decision-making and planning techniques
What are the challenges of GGP?
- Interpreting an arbitrary GDL rule set and deriving legal moves, states, and goals on the fly
- Building general-purpose search and evaluation methods that work without game-specific heuristics
- Balancing limited time per move against the cost of reasoning from first principles about an unfamiliar game
What are some common GGP systems and techniques?
- General game players such as CadiaPlayer and Fluxplayer, built specifically to compete in the AAAI GGP Competition
- Monte Carlo Tree Search (MCTS), used to search game trees without requiring a hand-crafted evaluation function
- Automatically derived heuristics extracted from the GDL rules of the current game, rather than heuristics written in advance for a specific game
What are some applications of GGP?
- Benchmarking general reasoning and search algorithms in AI research
- Informing techniques for automated planning in domains where rules are only known at runtime
- Providing a research foundation for later general game-playing efforts that combine GDL-style rule input with learned components