Abstract
UMRAO is an intelligent tutoring system (ITS) designed to teach chess endgames by focusing on strategic concepts rather than brute-force search. By representing chess knowledge through a "strategy graph" that includes both expert plans and common student misconceptions, the system provides conceptual feedback that helps players understand the logic behind winning (and losing) lines of play.
Introduction
In the late 1980s, computer chess was dominated by search-based engines—programs that could beat humans by calculating millions of moves ahead but could not explain why a move was strategically sound in human terms. For a student, a numerical evaluation like "+1.5" offers little pedagogical value. UMRAO addresses this gap by moving away from "minimax search" toward "knowledge-based chess," modeling how humans think in terms of goals, plans, and constraints.
Methodology
The UMRAO prototype focuses specifically on bishop-pawn endgames. Its architecture consists of two primary modules:
- The EXPERT Module: Operating "off-line," this module uses an extension of Donald Michie’s "advice language" to represent chess plans as objects. These plans are compiled into a Strategy Graph, a comprehensive map of plausible student moves justified by specific strategic plans.
- The TUTOR Module: This real-time interface uses the pre-compiled strategy graph to perform "model tracing." When a student makes a move, the tutor identifies the likely plan (expert or flawed) the student is following and provides context-aware feedback.
A key technical innovation is the dual-search approach: ExSearch identifies the optimal expert response, while StSearch exhaustively analyzes all plausible student responses, including suboptimal ones, to ensure the tutor can respond to any mistake.
Modeling Strategic Knowledge
A core strength of UMRAO is its explicit modeling of "flawed" student logic. By "perturbing" expert plans—for instance, by removing a safety check—the system can recognize specific, predictable strategic errors.
The system encodes procedural knowledge (how to play) as declarative descriptions of board states. By defining plans through predicates like black-king-threatens-P2, the system generates human-readable explanations. These plans are organized into the hierarchical graph structure mentioned previously, tracking the progression of a game across various strategic paths.
Results and Discussion
UMRAO serves as a successful proof-of-concept for conceptual chess tutoring. Unlike standard engines, it provides high-level advice, such as: "If you have double pawns and the first is blocked, exchange it for the bishop to queen the second."
The system supports four distinct tutoring styles, ranging from "strict tutor control" (immediate correction) to "student initiative" (where the student must request a hint). This flexibility allows for "delayed feedback," letting students pursue a suboptimal strategy until the consequences become apparent, at which point the tutor assists in backtracking.
Limitations and Future Work
While effective, UMRAO is currently limited to bishop-pawn endgames. Scaling to the middle game would require a significantly larger plan library. Additionally, student misconceptions were manually identified through "think-aloud" protocols with human novices; future iterations might automate the discovery of these flawed plans. Finally, while the architecture is robust, large-scale longitudinal studies are needed to quantify its impact on student rating improvement compared to traditional study methods.
Summary updated