Summary

• Proof of concept for using TensorFlow to improve AI at this Minesweeper game
• Mission : Improve accuracy(=predictive power) when AI should sweep or flag random cell with given condition

Procedure

• Create map randomly
• Sweep empty block for the first time
• Sweep or flag cells if AI could sweep or flag certain cell with given condition
• Collect data when AI should sweep or flag random cell with given condition
  • Target
    • Unrevealed cell which has at least one revealed number cell(=not empty or mine) around
    • X data
      • Sum of probabilities
      • Number of revealed mines around
      • Number of revealed cells around
      • Number of unrevealed cell around
    • Y data
      • 0 or 1 (Mine exists or not)
    • Repeat above procedure many times to get enough data
• Training sets
  • Method
    • Supervised Learning
      • Logistic classification
  • Optimization - Data initialization - Learning rate = 0.1
• Test sets : Measure accuracy of model

Prerequisite

• Python 3
• Numpy
• TensorFlow

Command

$ python3 ai.py

Result

• Collect data with 1000 games
• 2000 training sets
• (With test sets) Accuracy ~= 87(%)