News‎ > ‎

Best paper at EuroGP 2018!

posted Apr 9, 2018, 1:01 AM by Alberto Bartoli   [ updated Apr 9, 2018, 1:59 AM by Eric Medvet ]
Our work "On the Automatic Design of a Representation for Grammar-based Genetic Programming" has been chosen as best paper at the 21st European Conference on Genetic Programming (part of Evostar 2018).

This work is somewhat theoretical, at least for our standards. The basic idea is this: machine learning methods, including evolutionary algorithms, manipulate candidate solutions; solutions must be represented in a form suitable for the algorithm; the representation of a candidate solution must be defined in advance by a human operator; different representations may lead to different results, even for the same problem and the same evolutionary algorithm; selecting a "good representation" and motivating this choice is very difficult, it is basically an unsolved problem. 

In this work we investigated the possibility of designing a good representation automatically. That is, not only the computer constructs a solution automatically, it also constructs automatically the method for representing  solutions. Full details can be found in the paper, available from our lab page; the slides of the presentation are available from the same page.