Introduction
In the early 1990s, artificial intelligence research in games was characterized by "game-specific experts." Systems like Deep Blue for chess or Chinook for checkers achieved high performance by leveraging the deep, pre-existing knowledge of their human developers. Barney Pell identified this as the fixed representation trick: human programmers essentially pre-digest game strategy into the software's architecture, limiting the system's true generality.
To address this, Pell introduced METAGAMER, the first AI designed to play an entire class of "symmetric chess-like games" (SCL-games) without prior specific tuning. METAGAMER represents a shift toward general intelligence by requiring the program to autonomously develop strategies from raw rules provided at runtime.
The Metagame Paradigm
The core of this research is the "Metagame" paradigm. Unlike traditional game players that are hard-coded for a single set of rules, METAGAMER operates within a framework that includes a game generator and a formal grammar for defining rules. This allows the system to take raw rules as input and autonomously develop a strategy without human mediation.
METAGAMER is built on three main pillars:
- SCL-Metagame Grammar: A formal language used to define "Symmetric Chess-Like" games, including Chess, Checkers, Shogi, and infinite novel variations.
- Knowledge Compilation: The system uses partial evaluation to compile abstract rules into optimized, game-specific move generators and goal detectors at runtime.
- The Advisor Architecture: Instead of a hard-coded evaluation function, METAGAMER uses a library of "Advisors"—general heuristics like mobility, threats, and center-control.
Material Analysis and First-Principles Valuation
One of METAGAMER's most significant achievements is its ability to perform "static analysis" on game rules to assign values to pieces before playing a single move. By evaluating a piece's potential for mobility and its capacity to capture "victims" or reach goals, the system derives valuations from first principles.
As shown in the analysis for Chess, METAGAMER's derived values (e.g., Queen at 103, Rook at 75.5, Knight at 46.9) align remarkably well with human expert valuations. This emergent understanding allows the program to recognize, for instance, that a Queen is significantly more valuable than a Rook based solely on its movement and capture potential defined in the rules.
Key Results and Insights
METAGAMER demonstrated that general heuristics could successfully transfer to unknown environments:
- Checkers: It performed competitively against Chinook (the world-class engine) when given a one-man handicap, significantly outperforming simpler "greedy" programs.
- Chess: It played reasonably against GnuChess (at its easiest level) with a knight handicap.
- Emergent Strategy: In checkers, the system "re-discovered" the importance of keeping back-row men in place. It realized that moving them increased the opponent's material value by making promotion squares accessible.
Limitations and Future Work
Despite its strategic depth, METAGAMER faces challenges in tactical precision. Its general-purpose nature results in a shallower search depth (often 1 or 2-ply) compared to hyper-optimized engines. The lack of a quiescence search makes it vulnerable to "horizon effects" where it misses tactical captures just beyond its view.
Future work suggests using Temporal-Difference (TD) learning or self-play to autonomously tune the weights of different advisors, further reducing the need for human-set parameters and moving closer to fully autonomous general game playing.
Summary updated