Neural Networks in solving Minesweeper

Abstract

The purpose of this documentation is to present the operation of certain neural networks in solving the Minesweeper
game and to assess whether it is possible to represent the decisions made by these neural networks in an understandable
way using logical rules. Existing solutions such as CSP (Constraint Satisfaction Problem) were utilized to design an
algorithm that analytically solves the Minesweeper game. The results obtained were then used to train Multi-Layer
Perceptron (MLP), Encoding Neural Network (ENN), and Convolutional Neural Network (CNN) models. The CNN
emerged as the best-performing network. Based on the tests conducted by this network, a decision tree was constructed
that represents the network’s logic for these specific tests with approximately 90% accuracy. Ultimately, none of the
tested neural networks were able to match the analytical approach. However, based on the decision trees obtained for
the functioning networks (mainly CNN), it was inferred that, in theory, with a sufficiently large number of tests, it
should be possible to closely replicate the network’s operation using logical rules (nested conditional statements).