Learning Machine Learning With a Game

This paper by Christoph Lürig from the University of Applied Sciences Trier presents an explainable AI system for Connect-Four, built on an AlphaZero-style agent that combines a deep residual neural network with Monte Carlo Tree Search. The central aim is to make the decisions of a reinforcement learning game agent understandable to a human player, so that the player can learn both about the AI's inner workings and about the game itself. The work positions Connect-Four as a suitable testbed because its rules are simple, it is theoretically solved, and it requires no domain expert, in contrast to high-stakes fields such as medicine.

The system has two complementary components. The dynamic analysis renders a pruned, interactive radial visualization of the search tree, where nodes and edges encode visit-based probabilities and players' turns, and transpositions are handled through Zobrist hashing that turns the tree into a directed acyclic graph. The static analysis is the paper's main contribution: it adapts the LIME method to AlphaZero by perturbing board positions, weighting the resulting situations by their occurrence probability, and fitting a weighted Lasso model. The authors report that this avoids the more than 100 ms per-evaluation cost that a naive LIME application would impose, and that batch evaluation on a GPU makes the approach practical for interactive use. The resulting coefficients are visualized directly on the board as contributions favoring or disfavoring each player.

The paper is a system-description and methods contribution demonstrated through qualitative examples on selected game positions; it reports no user study or quantitative evaluation of explanation quality. The authors note that a planned extension to Othello is intended to support future user testing. The implementation uses Python, PyGame, PyTorch, and Scikit-Learn.