new IDSIA papers

[Juergen Schmidhuber]

3 new IDSIA publications available.  

Click at http://www.idsia.ch or use ftp:
FTP-host: fava.idsia.ch (192.132.252.1)
FTP-filenames: /pub/papers/ml95.kolmogorov.ps.gz     9 pages
               /pub/papers/ml95.antq.ps.gz           9 pages 
               /pub/papers/iwann95.invertible.ps.gz  8 pages
                                  (use gunzip to uncompress)
 
___________________________________________________________________


            DISCOVERING SOLUTIONS WITH LOW KOLMOGOROV 
          COMPLEXITY AND HIGH GENERALIZATION CAPABILITY
                  Juergen Schmidhuber, IDSIA
  To appear in Machine Learning: Proc. 12th int. conf., 1995.

This paper reviews basic concepts of Kolmogorov complexity 
theory relevant to machine learning. It shows how a derivate
of Levin's universal search algorithm can be used to discover 
neural nets with low Levin complexity, low Kolmogorov complexity, 
and high generalization capability.  At least with certain toy 
problems where it is computationally feasible, the method can 
lead to generalization results unmatchable by previous neural net 
algorithms. The final section addresses problems with incremental 
learning situations.

		         ANT-Q
                 Luca Gambardella, IDSIA
                   Marco Dorigo, IDSIA
  To appear in Machine Learning: Proc. 12th int. conf., 1995.

We introduce Ant-Q, a family of algorithms which share many 
similarities with Q-learning (Watkins, 1989). Ant-Q is a
generalization of the ``ant system'' (AS --- Dorigo, 1992; 
Dorigo, Maniezzo and Colorni, 1996), a distributed algorithm 
for combinatorial optimization based on the ant colony metaphor. 
In applications to symmetric traveling salesman problems (TSPs), 
we demonstrate (1) that some Ant-Q instances outperform AS, 
and (2) that Ant-Q compares favorably with other heuristic 
approaches based on neural nets or local search. Finally, we 
apply Ant-Q to some difficult asymmetric TSP's and obtain 
excellent results: Ant-Q finds solutions of a quality which 
usually can be found only by highly specialized algorithms.


        LEARNING THE VISUOMOTOR COORDINATION OF A MOBILE 
           ROBOT BY USING THE INVERTIBLE KOHONEN MAP
                   Cristina Versino, IDSIA
                   Luca Gambardella, IDSIA
In Proc. International Workshop on Artificial Neural Networks 1995.

This paper is based on the insight that the Extended Kohonen Map 
(EKM) is naturally invertible: given an input pattern, the network 
output is generated by competition among the neuron fan-in weight 
vectors (conventional ``forward mode''). Viceversa, given an output 
value, a corresponding input pattern can be obtained by competition 
among the neuron fan-out weight vectors (unconventional ``backward 
mode''). This invertibility property makes EKM worth considering for 
sensorimotor modeling. We present an experiment concerning visuomotor 
coordination of a simple mobile robot. ``Learning by doing'' creates 
a sensorimotor model: <perception, action> pairs are collected by 
observing the robot's behavior. These pairs are used for estimating 
the model's parameters. Training the network on the robot's direct 
kinematics (forward mode), one simultaneously obtains a solution to 
the inverse kinematics problem (backward mode). The experiment has 
been performed both in a simulation and by using a real robot.

___________________________________________________________________

Related and other papers in  http://www.idsia.ch
Comments welcome.

Juergen Schmidhuber 
Research Director
IDSIA, Corso Elvezia 36 
6900-Lugano, Switzerland
juergen at idsia.ch