TR: Self-organizing levels of articulation in sensory-motor systems.

[Jun.Tani (SONY CSL)]

Dear Connectionists,

The following technical paper is avairable in
     http://www.csl.sony.co.jp/person/tani.html
or directly to:
     ftp://ftp.csl.sony.co.jp/CSL/CSL-Papers/97/SCSL-TR-97-008.ps.Z

---------------------------------------------------------------------------
Self-Organization of Modules and Their Hierarchy in Robot Learning
Problems: A Dynamical Systems Approach

Jun Tani and Stefano Nolfi
(Sony CSL Technical Report: SCSL-TR-97-008)

ABSTRACT:
This paper discusses how modular and hierarchical structures can be
self-organized dynamically in a robot learning paradigm.
We develop an on-line learning scheme -- the so-called mixture of
recurrent neural net (RNN) experts -- in which a set of RNN modules becomes
self-organized as experts in order to account for the different categories of
sensory-motor flow which the robot experiences.
Autonomous switching between winning expert modules, responding to
structural changes in the sensory-motor flow, actually corresponds to the temporal segmentation of behavior.
In the meanwhile, another mixture of RNNs at a higher level learns
the sequences of module switching occurring in the lower level, by
which articulation at a further more abstract level is achieved.
The proposed scheme was examined through simulation experiments involving
the navigation learning problem.
The simulated robot equipped with range sensors traveled around rooms
of different shape.
It was shown that modules corresponding to concepts such as turning
right and left at corners, going straight along corridors and encountering
junctions are self-organized in the lower level network.
The modules corresponding to traveling in different rooms are
self-organized in the higher level network.
The robot succeeded in learning to perceive the world as articulated at
multiple levels through its recursive interactions.
--------------------------------------------------------------------------

Jun TANI, Ph.D
Senior Researcher
Sony Computer Science Laboratory Inc.
Takanawa Muse Building, 3-14-13 Higashi-gotanda,
Shinagawa-ku, Tokyo, 141 JAPAN
email: tani at csl.sony.co.jp
http://www.csl.sony.co.jp/person/tani.html
Fax +81-3-5448-4273
Tel +81-3-5448-4380