| Paper: | PS-2A.26 | ||
| Session: | Poster Session 2A | ||
| Location: | H Lichthof | ||
| Session Time: | Sunday, September 15, 17:15 - 20:15 | ||
| Presentation Time: | Sunday, September 15, 17:15 - 20:15 | ||
| Presentation: | Poster | ||
| Publication: | 2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany | ||
| Paper Title: | Modular RL for Real-Time Learning in Physical Environments | ||
| Manuscript: | Click here to view manuscript | ||
| License: |  This work is licensed under a Creative Commons Attribution 3.0 Unported License. |
||
| DOI: | https://doi.org/10.32470/CCN.2019.1270-0 | ||
| Authors: | Per R. Leikanger, UiT - The Arctic University of Norway, Norway | ||
| Abstract: | Reinforcement Learning is a powerful approach to a machine learning from interaction with the environment. Recent impressive achievements show the potency of combining Deep Learning with RL for board games and simple visually oriented tasks, but similar feats seem to wait for physical systems. We ask ourself for the reason(s) for this apparent hole in the science, on our path toward adaptive automata for physical domains. Inspired by computational systems in biology, revisiting conceptual fundamentals of Reinforcement Learning, and dividing complex tasks into numerous parallel learners, we aim to lessen this effect for physical interaction. A simple demonstration of an algorithmic framework is implemented, creating an agent that learns complex reactive behavior in a continuous parameter space by tabular RL methods. We conclude with a discussion on possible implications from this work for adaptive agents in the physical world, and plausible further directions toward life-long learning. | ||