README.md

---
license: apache-2.0
---
# Dataset Card for Pong-v4-expert-MCTS
## Table of Contents

- [Supported Tasks and Baseline](#support-tasks-and-baseline)

- [Data Usage](#data-usage)

  - [Data Discription](##data-description)
  - [Data Fields](##data-fields)
  - [Data Splits](##data-splits)
  - [Initial Data Collection and Normalization](##Initial-Data-Collection-and-Normalization)

- [Additional Information](#Additional-Information)

  - [Who are the source data producers?](## Who-are-the-source-data-producers?)
  - [Social Impact of Dataset](##Social-Impact-of-Dataset)
  - [Known Limitations](##Known-Limitations)

  - [Licensing Information](##Licensing-Information)
  - [Citation Information](##Citation-Information)
  - [Contributions](##Contributions)

## Supported Tasks and Baseline

- This dataset supports the training for [Procedure Cloning (PC )](https://arxiv.org/abs/2205.10816) algorithm.
- Baselines when sequence length for decision is 0:

| Train loss                                         | Test Acc | Reward |
| -------------------------------------------------- | -------- | ------ |
| <img src="./img/sup_loss.png" style="zoom:50%;" /> | 0.90     | 20     |

- Baselines when sequence length for decision is 4:

| Train action loss                                     | Train hidden state loss                           | Train acc (auto-regressive mode)                    | Reward |
| ----------------------------------------------------- | ------------------------------------------------- | --------------------------------------------------- | ------ |
| <img src="./img/action_loss.png" style="zoom:50%;" /> | <img src="./img/hs_loss.png" style="zoom:50%;" /> | <img src="./img/train_acc.png" style="zoom:50%;" /> | -21    |

## Data Usage

### Data description

This dataset includes 8 episodes of pong-v4 environment. The expert policy is [EfficientZero]([[2111.00210\] Mastering Atari Games with Limited Data (arxiv.org)](https://arxiv.org/abs/2111.00210)), which is able to generate MCTS hidden states. Because of the contained hidden states for each observation, this dataset is suitable for Imitation Learning methods that learn from a sequence like PC.

### Data Fields

- `obs`: An Array3D containing observations from 8 trajectories of an evaluated agent. The data type is uint8 and each value is in 0 to 255. The shape of this tensor is [96, 96, 3], that is, the channel dimension in the last dimension. 
- `actions`: An integer containing actions from 8 trajectories of an evaluated agent. This value is from 0 to 5. Details about this environment can be viewed at [Pong - Gym Documentation](https://www.gymlibrary.dev/environments/atari/pong/).
- `hidden_state`: An Array3D containing corresponding hidden states generated by EfficientZero, from 8 trajectories of an evaluated agent. The data type is float32.

This is an example for loading the data using iterator:

```python
from safetensors import saveopen

def generate_examples(self, filepath):
    data = {}
    with safe_open(filepath, framework="pt", device="cpu") as f:
        for key in f.keys():
            data[key] = f.get_tensor(key)
        
    for idx in range(len(data['obs'])):
        yield idx, {
            'observation': data['obs'][idx],
            'action': data['actions'][idx],
            'hidden_state': data['hidden_state'][idx],
        }
```

### Data Splits
There is only a training set for this dataset, as evaluation is undertaken by interacting with a simulator.

### Initial Data Collection and Normalization

- This dataset is collected by EfficientZero policy.
- The standard for expert data is that each return of 8 episodes is over 20.
- No normalization is previously applied ( i.e. each value of observation is a uint8 scalar in [0, 255] )

## Additional Information

### Who are the source language producers?

[@kxzxvbk](https://huggingface.co/kxzxvbk)

### Social Impact of Dataset

- This dataset can be used for Imitation Learning, especially for algorithms that learn from a sequence.
- Very few dataset is open-sourced currently for MCTS based policy.
- This dataset can potentially promote the research for sequence based imitation learning algorithms.

### Known Limitations

- This dataset is only used for academic research.
- For any commercial use or other cooperation, please contact: opendilab@pjlab.org.cn

### License
This dataset is under [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).

### Citation Information

```
@misc{Pong-v4-expert-MCTS,
    title={{Pong-v4-expert-MCTS: OpenDILab} A dataset for Procedure Cloning algorithm using Pong-v4.},
    author={Pong-v4-expert-MCTS Contributors},
    publisher = {huggingface},
    howpublished = {\url{https://huggingface.co/datasets/OpenDILabCommunity/Pong-v4-expert-MCTS}},
    year={2023},
}
```

### Contributions
This data is partially based on the following repo, many thanks to their pioneering work:

- https://github.com/opendilab/DI-engine
- https://github.com/opendilab/LightZero

Please view the [doc](https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cardsHow) for anyone who want to contribute to this dataset.