MSc Thesis: Explainable Chess Engine
Neuro-symbolic integration of Scenario-Based Programming (SBP) with a dynamic-querying LLM layer, so a chess engine can explain its decisions in natural language.
Problem
A chess engine can choose a strong move without being able to say why. For someone learning the game, the reasoning is the useful part, and complex tactical positions are where explanations are most likely to fail as teaching.
Approach
- A symbolic-to-natural-language bridge. The symbolic side is built with Scenario-Based Programming. When its rules conflict, the bridge translates those conflicts into natural-language explanations.
- A dynamic query layer. For complex tactical positions, an LLM layer probes board-state traces, so the instruction it gives is grounded in the specific position.
Evaluation
I designed an evaluation framework that measures two things about the generated explanations, faithfulness and pedagogical quality, across a standardised suite of tactical positions.