README.md

---
language: en
tags:
- machine-learning
- reinforcement-learning
- sokoban
- planning
license: apache-2.0
---

# Trained learned planners

This repository contains the trained networks from the paper ["Planning behavior in a recurrent neural network that
plays Sokoban"](https://openreview.net/forum?id=T9sB3S2hok), presented at the ICML 2024 Mechanistic Interpretability
Workshop.

To load and use the NNs, please refer to the [learned-planner
repository](http://github.com/alignmentresearch/learned-planner), and possibly to the [training code
](https://github.com/AlignmentResearch/train-learned-planner).

# Model details

## Hyperparameters:

See `model/*/cp_*/cfg.json` for the hyperparameters that were used to train a particular run.

## Best Models:

The best models for each of the model type are stored in the following directory:
| Model | Directory | Parameter Count |
|:-------|:-----------|:-----------------|
| DRC(3, 3)  | `drc33/bkynosqi/cp_2002944000` | 1,285,125 (1.29M) |
| DRC(1, 1)  | `drc11/eue6pax7/cp_2002944000` | 987,525 (0.99M)  |
| ResNet     | `resnet/syb50iz7/cp_2002944000` | 3,068,421 (3.07M) |

## Probes & SAEs:

The trained probes and SAEs are stored in the `probes` and `saes` directories, respectively.

## Training dataset:

The [Boxoban set of levels by DeepMind](https://github.com/google-deepmind/boxoban-levels).

# Citation

If you use any of these artifacts, please cite our work:

```bibtex
@inproceedings{garriga-alonso2024planning,
    title={Planning behavior in a recurrent neural network that plays Sokoban},
    author={Adri{\`a} Garriga-Alonso and Mohammad Taufeeque and Adam Gleave},
    booktitle={ICML 2024 Workshop on Mechanistic Interpretability},
    year={2024},
    url={https://openreview.net/forum?id=T9sB3S2hok}
}
```