https://github.com/ukjin1192/tensorflow-minesweeper

Proof of concept for using TensorFlow to improve AI at this Minesweeper game
https://github.com/ukjin1192/tensorflow-minesweeper

machine-learning minesweeper-game tensorflow tensorflow-experiments

Last synced: 4 months ago
JSON representation

Proof of concept for using TensorFlow to improve AI at this Minesweeper game

• Host: GitHub
• URL: https://github.com/ukjin1192/tensorflow-minesweeper
• Owner: ukjin1192
• Created: 2017-03-02T13:43:15.000Z (almost 8 years ago)
• Default Branch: master
• Last Pushed: 2018-08-26T17:15:59.000Z (over 6 years ago)
• Last Synced: 2024-10-10T08:42:34.536Z (4 months ago)
• Topics: machine-learning, minesweeper-game, tensorflow, tensorflow-experiments
• Language: Python
• Homepage: https://minemine.io/
• Size: 21.5 KB
• Stars: 10
• Watchers: 2
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
## 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(%)