Active Offline Policy Selection

This is supporting example code for NeurIPS 2021 paper Active Offline Policy Selection by Ksenia Konyushkova, Yutian Chen, Tom Le Paine, Caglar Gulcehre, Cosmin Paduraru, Daniel J Mankowitz, Misha Denil, Nando de Freitas.

To simulate the active offline policy selection for a set of policies, one needs to provide a number of files. We provide the files for 76 policies on cartpole_swingup environemnt.

1. Sampled episodic returns for all policies on a number of evalauation episodes (full-reward-samples-dict.pkl), or a way of sampling a new episode of evaluation upon request for any policy. The file full-reward-samples-dict.pkl contains a dictionary that maps a policy by its string representation to a numpy.ndarray of of shape (5000,) (number of reward samples).

2. Off-policy evaluation score, such as fitted Q-evaluation (FQE) for all policies (ope_values.pkl). The file ope_values.pkl contains dictionary that maps policy info into OPE estimates. We provide FQE scores for the policies.

3. Actions that policies take on 1000 randomly sampled states from the offline dataset (actions.pkl). The file actions.pkl contains a dictionary with keys actions and policy_keys. actions is a list of 1000 ( number of states used to compute the kernel) elements of numpy.ndarray type of dimensionality 76x1 (number of policies by the dimensionality of the actions). policy_keys contains a dictionary mapping from string representation of a policy to the index of that policy in actions.

Installation

To set up the virtual environment, run the following commands. From within the active_ops directory:

``` python3 -m venv activeopsenv source activeopsenv/bin/activate

pip install --upgrade pip pip install -r requirements.txt ```

To run the demo with colab, enable the jupyter_http_over_ws extension:

jupyter serverextension enable --py jupyter_http_over_ws

Finally, start a server:

jupyter notebook \ --NotebookApp.allow_origin='https://colab.research.google.com' \ --port=8888 \ --NotebookApp.port_retries=0

Usage

To run the code refer to Active_ops_experiment.ipynb colab notebook. Execute blocks of code one by one to reproduce the final plot. You can modify various parameters maked by @param to test various baselines in modified settings. This code loads the example of data for cartpole_environment provided in the data folder. Using this data, we reproduce the results of Figure 14 of the paper.

Citing this work

@inproceedings{konyushkovachen2021aops, title = "Active Offline Policy Selection", author = "Ksenia Konyushkova, Yutian Chen, Tom Le Paine, Caglar Gulcehre, Cosmin Paduraru, Daniel J Mankowitz, Misha Denil, Nando de Freitas", booktitle = NeurIPS, year = 2021 }

Disclaimer

This is not an official Google product.

The datasets in this work are licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.