AGH University of Krakow
Center of Excellence in Artificial Intelligence
Maciej Aleksandrowicz
Joanna Jaworek-Korjakowska,
Due to a co-author’s preference, the preprint version of this work is not publicly available before formal peer review.
A public version (with the implementation code) will be made available as soon as possible.
This study approaches the deep reinforcement learning issue of sample inefficiency with a paradigm shift for observation encoding. The agent’s trained perception module is moved to the environment and kept frozen during training. This approach is evaluated on visual observations, utilizing pre-trained visual models. While results are not conclusive, they yield preliminary insights for future research directions.
Fig.1 - The overview of the proposed Frozen Feature Extractor architecture for an arbitrary DeepRL agent.