Neural Network Learner for Minesweeper

TL;DR

This paper introduces a neural network-based learner for Minesweeper, demonstrating competitive performance with traditional solvers in beginner and intermediate modes, while highlighting speed limitations.

Contribution

It presents the first neural network architecture tailored for Minesweeper, trained on extensive game data, and compares its effectiveness to deterministic solvers.

Findings

Neural network learner performs well in beginner and intermediate modes.

The neural approach is slower than deterministic solvers.

High success rates achieved despite speed limitations.

Abstract

Minesweeper is an interesting single player game based on logic, memory and guessing. Solving Minesweeper has been shown to be an NP-hard task. Deterministic solvers are the best known approach for solving Minesweeper. This project proposes a neural network based learner for solving Minesweeper. To choose the best learner, different architectures and configurations of neural networks were trained on hundreds of thousands of games. Surprisingly, the proposed neural network based learner has shown to be a very good approximation function for solving Minesweeper. The neural network learner competes well with the CSP solvers, especially in Beginner and Intermediate modes of the game. It was also observed that despite having high success rates, the best neural learner was considerably slower than the best deterministic solver. This report also discusses the overheads and limitations faced while creating highly successful neural networks for Minesweeper.