shsjxzh / gvin

Generalized Value Iteration Network

Generalized Value Iteration Networks: Life Beyond Lattices

Sufeng Niu, Siheng Chen, Hanyu Guo, Colin Targonski, Melissa C. Smith, Jelena Kovačević Generalized Value Iteration Networks.

Dependencies

• Python >= 2.7.6
• TensorFlow >= 1.0
• SciPy >= 0.18.1 (to load the data)

Training

data directory already contains 16x16 grid world and 10 nodes irregular graph. To play with irregular graph which contains 100 nodes, Download the datasets from here. Then, put it in data directory.

# Runs 10 nodes graph with default parameter using imitation learning
python \irregular\IL\train.py

# Runs 10 nodes graph with default parameter using reinforcement learning
python \irregular\RL\train.py

# Runs 16x16 grid world with default parameter using imitation learning
python \regular\IL\train.py

Acknowledgement:

This code is partially based on the TensorFlow implementation of Abhishek Kumar here.

Resources

• Generalized Value Iteration Networks: Life Beyond Lattices on arXiv
• Value Iteration Networks on arXiv
• Value Iteration Networks in TensorFlow