HNPG

This repository contains our hybrid NPG-style algorithm for the comblock environment.

Prerequisites

Please see requirements.txt for required packages.

We use wandb to perform result collection, please setup wandb before running the code or add os.environ['WANDB_MODE'] = 'offline' in main.py.

For Cifar100 Comblock, please download cifar-100 from here, and set the path in main.py to the path of your cifar-100-python folder.

The offline dataset is collected by following $\pi^\star$ with $\epsilon$-greedy exploration with $\epsilon=1/H$.

Please refer to our paper for more details.

To reproduce our result in continuous comblock:

python main.py --seed 12345 --env-name lock --horizon 50

To reproduce our result in cifar100 continuous comblock:

python main.py --seed 12345 --env-name cifarlock --horizon 30