ScyberLabPreview
Open in TheBrain

De Jong- A Coevolutionary Approach to Representation Development

    <p class=default>
      
        De Jong, E.D. and Oates, T. (2002). A Coevolutionary Approach to Representation 
        Development . 
      
    <br/>
      
             Proceedings of the ICML-2002 Workshop on Development of Representations.
      
    </p>

<p>

      
             Abstract
      
    <br/>
      
             The representation of a learning or search problem can greatly impact its 
        performance. An alternative to hand-constructing an appropriate representation is to let
      
    <br/>
      
             the learning method adapt its own representation. We investigate this in a setup 
        where building blocks and assemblies thereof are coevolved. Building blocks may
      
    <br/>
      
             employ other building blocks, thus leading to hierarchical constructions. In 
        experiments, this is observed to lead to highly compact representations of sequences
      
    <br/>
      
             that are useful as building blocks. The method is able to solve problems requiring 
        long sequences of primitive operators. Control experiments using a conventional
      
    <br/>
      
             evolutionary method were much less efficient, and did not find solutions to the 
        problems. Limitations of the current method are discussed. . 
      
    </p>

<p>

      
             Keywords: Development of representations, coevolution, bias learning.
      
    </p>

<p>

      
             Download this paper as: Postscript (dejongoates02.ps) Gzipped Postscript 
        (dejongoates02.ps.gz) PDF (dejongoates02.pdf) 
      
    </p>

Attachments