MineSweeper Agent PPO

Features

1. PPO implementation based on Ray

2. Medium level: 16 x 16, 40 mines, Premium level version is on expert branch.

3. ~62% winning rate (inference with max probability policy), with stochastic policy the rate is only ~53%.

Pretrained Checkpoint

1. with clicking reward (every valid click earns a constant reward of +0.1)
2. medium-checkpoint: policy_value_1250000.pt

Training Stats

Inference

1. cd inference
2. python inference.py

If you want to estimate the winning rate, run python stat_win_rate.py.

Requirements

• einops==0.8.1
• gym==0.26.2
• gymnasium==1.0.0
• numpy==2.1.2
• ray==2.42.1
• six==1.17.0
• torch==2.6.0+cu118

Train your own model

1. ensure python version == 3.10
2. pip install -r requirements
3. mkdir ckpt
4. python main.py

Result

Acknowlegements

1. The minesweeper env is borrowed from aylint/gym-minesweeper.
2. The parallel PPO implementation is based on Ray framework.