Abstract
This paper explores a "situation control" technique for chess endgames, moving away from traditional brute-force search toward a semiotic model of piece relations. By formalizing human perception of "structural generality," the authors developed a system using generative grammars to calculate solutions through logical relations. Experimental results with human players and computer simulations on the Minsk-32 demonstrate the effectiveness of this relational approach in replicating the "mental look" of expert players.
Introduction
In the early days of artificial intelligence, the dominant paradigm for chess was the "labyrinth" search—brute-forcing a tree of possible moves. While effective for the tactical skirmishes of the mid-game, this approach famously struggled with the endgame. This 1970s-era research from the U.S.S.R. proposes a different path: instead of searching deeper, we should teach computers to "see" like humans. The core hypothesis is that human experts do not simply calculate more moves; they perceive the board as a graph of relations between pieces rather than a sequence of board states.
The Problem of Search Depth
Traditional chess programs of the era were severely limited by "calculation depth"—the number of moves they could look ahead. In the endgame, an optimum solution often requires looking many moves ahead, far beyond the computational limits of 1970s hardware. The author argues that the gap between human and machine performance is a lack of a mathematical language capable of describing the "process of comprehending a position" and the "mental look" that precedes a move.
The Situation Control Method
The authors employ a Situation Control method (based on work by Pospelov and Klykov) which uses a semiotic system composed of three parts:
- Base Notions ($I$): Natural language units like "white king" or "field."
- Base Relations ($E$): Binary relations such as "is an element of," "is on," or "has a name."
- Formation Rules ($G$): Generative grammars used to formalize how situations are structured and generalized.
Formalizing Base Notions
To bridge the gap between human perception and machine logic, the system first defines base notions and relations. This allows the computer to represent the board not as a grid of squares, but as a set of logical entities.
Generative Grammars and Logical Chains
The system features a "determination of potential relations" block—an analog to the human "mental look." This block represents a position as a set of ordered triplets $(x_i, R_m, x_j)$, where $x$ are pieces/fields and $R$ is a "beam" of potential relations. Crucially, the model uses four types of generative grammars (correlation, categorial, transformation, and commanding) to enumerate micro-situations and form generalized structures, such as recognizing a "fork" or "pin."
These grammars allow the AI to recognize "structural generality"—the determination of identical relations between pieces in different specific conditions. This allows the system to understand that a "Knight fork" is the same logical structure regardless of its specific coordinates on the board.
Experimental Results
The study utilized psychological experiments observing human players of varying skill levels and computer simulations on a Minsk-32 electronic computer. Key results include:
- Library of Rules: The researchers developed 175 correlation grammar rules for a 7-piece endgame.
- Syntagmatic Chains: The system generated logical sentences describing the board, such as: "Knight attacks black King and black-field Bishop in three half-moves."
- Generalization: The model successfully identified approximately 50 generalization rules for complex chess concepts.
- Dynamic Planning: Unlike traditional programs that look a fixed $N$ moves ahead, this model's solution length changes dynamically based on the logical structure of the position.
Conclusion
The paper posits that the basic difficulty of programming isn't a lack of speed, but a lack of a language to describe the process of establishing relations. By focusing on "relational depth" rather than move depth, the situation control technique provides a framework for more human-like decision-making in complex environments. While hardware constraints of the era restricted the complexity of the tests, the model demonstrated that a semiotic approach can successfully replicate the structural thinking of high-level chess players.
Summary updated