General Game Playing

General Game Playing (GGP) is the problem of building agents that can play arbitrary games given only their rules, without domain-specific engineering. Unlike specialized systems like AlphaZero — which are optimized for a fixed game — a general game player receives a formal description of an unfamiliar game and must devise a competent strategy in real time. The field originated in the GGP competition established at Stanford in 2005, and its formal language (Game Description Language, GDL) encodes game rules as logical propositions that a general reasoner can interpret.

The challenge is not merely computational but epistemological. A specialized player knows what matters in its domain: territory in Go, material in chess, liberty in checkers. A general player has no such priors. It must infer what is important from the rules themselves — identifying symmetries, detecting terminal conditions, estimating heuristic value — all without the benefit of human strategic wisdom. GGP is therefore a test of synthetic competence: can an agent construct a useful theory of a game from its formal structure alone?

Modern approaches combine logical reasoning with learned heuristics, treating GDL descriptions as programs to be compiled into efficient search representations. The connection to automated reasoning and symbolic AI is direct: GGP agents must perform logical inference at scale, often under severe time constraints. The field remains stubbornly difficult. While specialized agents dominate their target games, general agents perform at amateur levels across the board. The gap between specific and general competence is one of the sharpest measures of how far AI remains from genuine generality.

The GGP enterprise exposes a hidden assumption in most AI research: that competence in a specific domain is evidence of general capability. It is not. A system that masters chess is no closer to mastering Go than a fish is to climbing a tree — unless the system was designed for generality from the start. AlphaZero's generality across three games is impressive, but it is still generality within a narrow class: perfect-information, deterministic, zero-sum board games. True general game playing remains the horizon, not the shore.