Information-theoretic approach to embodied category learning

We address the issue of how statistical and information-theoric measures can be employed to quantify the categorization process of a simulated robotic agent interacting with its local environment. We show how correlation, entropy, and mutual information can help identify distinct informational structure which can be used for object classification. Further, by means of the isometric feature mapping algorithm, we analyze the weights of a neural network designed to find clusters based on these distinct information theoretic characteristics of the object’s shape, size and color. We conclude that an understanding of the information-theoretic implications of categorization could help design robots with improved catego rization and better exploration strategies.


Published in:
The Tenth International Conference on Artificial Life and Robotics, 10, 332-337
Presented at:
The Tenth International Conference on Artificial Life and Robotics, Oita,Japan, Feb 4 - Feb 6, 2005
Year:
2005
Keywords:
Laboratories:




 Record created 2006-04-19, last modified 2018-03-17

n/a:
Download fulltext
PDF

Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)