What Planning Problems Can A Relational Neural Network Solve?

What Planning Problems Can A Relational Neural Network Solve?
Jiayuan Mao, Tomás Lozano-Pérez, Joshua B. Tenenbaum, and Leslie Pack Kaelbling
In Conference on Neural Information Processing Systems (NeurIPS) 2023
[Paper] [Project Page] [BibTex]

@inproceedings{Mao2023RegressionWidth,
  title={{What Planning Problems Can A Relational Neural Network Solve?}},
  author={Mao, Jiayuan and Lozano-Perez, Tomas and Tenenbaum, Joshua B. and Leslie Pack Kaelbing},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2023}
}

Setup

This repo is based on the released code of "Neural Logic Machines": GitHub Repo.

Please follow the setup instructions in the original repo. To replicate the results:

Assembly3

# Depth = 1
jac-run learn_policy_reinforce_find3.py --task find3 --blocks 8 --epochs 100 --depth 1
# Depth = 2
jac-run learn_policy_reinforce_find3.py --task find3 --blocks 8 --epochs 100 --depth 2

BlocksWorld-Clear

# No recursion
jac-run learn_policy_reinforce_blocks.py --task single-clear --curriculum --blocks 4 --depth 3

# With recursion
jac-run learn_policy_reinforce_blocks.py --task single-clear --curriculum --blocks 4 --depth 6 --recursion=True --io-residual=True

Logistics

# No recursion
jac-run learn_policy_reinforce_logistics.py --task directed-pathfinding --curriculum --blocks 10 --depth 3

# With recursion
jac-run learn_policy_reinforce_logistics.py --task directed-pathfinding --curriculum --blocks 10 --depth 10 --recursion=True --io-residual=True