sequential learning - lifelong learning

[Ron Sun]

To transfer knowledge from one environment to another, one viable way,
I believe, is to extract generic rules (from a NN)
that are more widely applicable.
This may not solve the interference problem, but surely handles
the transfer problem to a certain extent.
(It  can also deal with the interference problem, if extracted rules
are used to train the NN, interspersed with current data.)

Here is a  TR on extracting rules from a Q-learning network.  
The resulting architecture consists 
of two levels, which  contains  both rules and Q-learning
network so that both rigorous, abstract knowledge (declarative knowledge)
and flexible, embodied knowledge (procedural knowledge)
are maintained. 

The learning and rule extraction
are on-line, while performing the task, and can be continuously adaptive.
Rule extraction is done on top of the connectionist network performing
Q-learning, so the architecture is parsimonious in terms of learning mechanisms.

The TR is available at 
FTP-host: aramis.cs.ua.edu  
FTP-file: pub/tech-reports/sun.clarion.ps


================================================================
Dr. Ron Sun                      
Department of Computer Science            phone: (205) 348-6363
The University of Alabama                 fax:   (205) 348-0219
Tuscaloosa, AL 35487                      rsun at athos.cs.ua.edu
================================================================