/successor_uncertainties_atari

Code for paper "Successor Uncertainties: Exploration and Uncertainty in Temporal Difference Learning" by David Janz*, Jiri Hron*, Przemysław Mazur, Katja Hofmann, José Miguel Hernández-Lobato, Sebastian Tschiatschek. NeurIPS 2019. *Equal contribution

Primary LanguagePython

Successor Uncertainties, Atari Experiments

This code allows for reproduction of the Atari experiments in https://arxiv.org/abs/1810.06530. Click here for code to reproduce the tabular experiments.

To reproduce results, clone && pip install the requirements, then run

python3 run_atari.py --game Enduro

to train a Successor Uncertainties model with parameters as per the paper. This will output training information in tensorboard format to a subdirectory called logs. To obtain test scores, run

python3 run_test.py /path/to/log_folder output_file.txt

The final score will be output to output_file.txt and progress of testing will be reported to stdout.