AN EXPLAINABLE AND REPRODUCIBLE ARTIFICIAL INTELLIGENCE BASED DECISION-SUPPORT FRAMEWORK FOR CHESS MOVE EVALUATION

Abstract

The present work provides an AI-driven move-level chess analysis decision support framework which is transparent and repeatable. The framework incorporates the Stockfish chess engine as a quantitative evaluation component and adds a structured analytical layer which transforms the engine evaluations into interpretable move classification and human readable explanations. During each move, the framework is able to compute the centipawns changes of evaluation by normalizing the scores with respect to the mover and to categorize the moves within the pre-defined classes, with the use of fixed and well-documented thresholds, such as good, neutral, inaccuracy, mistake, and blunder. The evaluation changes are connected to observed properties of the board, e.g. material changes, captures, checks, and the size of the evaluation, by using a deterministic rule-based explanation layer. A deterministic material-based fallback mechanism is also provided which ensures analytical operations in the absence of engine evaluation. The modular design allows the separation of position parsing, engine evaluation, position classification, explanations, and visualization, enabling transparency and reproducibility. The operation of the framework is demonstrated through a representative example of playing chess and repeated run consistency checks in an experimental analysis. The results show that output of evaluation, classification, and explanation are consistent when the input and configuration are identical. However, all the proposed explanations are limited to the specific characteristics of the board that can be seen and these explanations are not explanations of the kind of thought that would happen in the mind of the chess engine itself. Similarly, leveraging the hard-coded cut-off for the delta for the centipawns per move was not reflective of context-dependent and/or nonlinear changes in position, and the material-based fallback may not have been an adequate measure of tactical sacrifices. The framework was then put in place as a “rule-based” decision support system instead of a model-level Explainable AI system.