- Abstract:
-
Research into using reinforcement learning to find optimal solutions to tasks where only partial information is available, i.e. partially observable Markovian decision processes (POMDPs), has traditionally focused on augmenting learning algorithms with memory or the ability to build internal models of the world. Our approach differs in that we consider agents with active perception, i.e. agents who exercise control over the sensory input they obtain from the world. Our conjecture is that agents should be able to learn to use active perception to find optimal solutions to what are otherwise POMDPs; and further, that this is possible using basic reinforcement learning techniques that do not employ memory or internal models.
This two page paper presents our preliminary empirical investigation into this conjecture. We present some suggestive results, though at this stage our work is a limited proof of concept, focusing on a single grid world problem with no comparison against the efficacy of reinforcement learning techniques enhanced with memory or modelling abilities.
- Copyright:
- 2003 by The University of Edinburgh. All Rights Reserved
- Links To Paper
- No links available
- Bibtex format
- @InProceedings{EDI-INF-RR-0181,
- author = {
Paul Crook
and Gillian Hayes
},
- title = {Active Perception in Navigation of Partially Observable Grid Worlds},
- book title = {proceeding of the Sixth European Workshop on Reinforcement Learning (EWRL-6,2003)},
- year = 2003,
- month = {Sep},
- }
|