GECKIES GROUP SEMINAR SERIES

1
GECKIES GROUP SEMINAR SERIES

• State of Iteration and Recursion in Genetic
  Programming
• Edwin Rodriguez
• edwin_at_cis.ksu.edu
• Genetic and Evolutionary Computation KSU
  Investigation and Experimentation Studio
• KDD Laboratory
• Kansas State University

2
Outline

• Introduction to Genetic Programming
• Drawbacks of GP
• Iteration and Recursion in GP
• Conclusions
• References

3
Introduction to GP

• One of the many variants of EAs
• Based on Darwinian evolution concept of natural
  selection
• A direct descendant of GAs
• Can be seen as a more generalized GA with
  variable length chromosomes
• Automated program induction
• Deemed as one of the few candidate technologies
  that could lead to Strong AI

4
Introduction to GP 2

• Same evolutionary cycle as that of GAs
• Selection -gt Gen Op App -gt Selection
• Traditional Genetic Operators
• Crossover and Mutation
• Individuals are programs
• The individuals representation is usually a parse
  tree, but other representations are used

5
Drawbacks of GP

• Code bloat
• Sections of useless code that gets executed and
  makes evaluation very expensive
• Turing Incompleteness
• Traditional GP is not Turing complete
• There are, in advance, problems that GP cannot
  solve
• Breaks down the promise of totally automated
  program induction

6
Drawbacks of GP 2

• Why is traditional GP Turing incomplete?
• No memory
• No iteration or recursion
• Consider the very simple problem
• Recognize the strings of the form
• 0n1n
• Traditional GP cannot solve this problem

7
Iteration and Recursion in GP

• Part of what is needed to fix GP
• Early work in this area
• Koza's work on restricted iteration
• Teller's mental models with GP
• Little work on recursion
• Iteration is simpler to deal with and gives the
  same power
• Some work on restricted iteration

8
Iteration and Recursion in GP 2

• Restricted iteration
• Impose an external limit to the number of
  iterations (usually given by some domain
  information)
• Disallow nested iterations or restrict the
  nesting depth
• Exploit structure of inductively defined data to
  limit number of iterations (inductive iteration)

9
Iteration and Recursion in GP 3

• Restricted recursion
• Not much studied
• Iteration is easier to control externally than
  recursion
• What has been done is basically the same as for
  iteration, with no additional benefits
• Impose an external limit to the number of
  recursive calls (usually given by some domain
  information)

10
Conclusions

• The use iteration and recursion gives a lot of
  benefits to GP
• More compact code
• Increases level of generalization
• Is this enough?
• After all these techniques are restricted
• Technically speaking, this doesn't make GP Turing
  complete
• It does increase the power of GP, but more needs
  to be done

11
References

• Koza, John R. and Andre, David. 1996b. Evolution
  of iteration in genetic programming. In Fogel,
  Lawrence J., Angeline, Peter J. and Baeck, T.
  Evolutionary Programming V Proceedings of the
  Fifth Annual Conference on Evolutionary
  Programming. Cambridge, MA The MIT Press. Pages
  469 -- 478.
• Evan Kirshenbaum. Iteration Over Vectors in
  Genetic Programming. HP Laboratories Technical
  Report HPL-2001-327, December 17, 2001.
• Scott Brave. Evolving recursive programs for tree
  search. In Peter J. Angeline and K. E. Kinnear,
  Jr., editors, Advances in Genetic Programming 2,
  chapter 10, pages 203220. MIT Press, Cambridge,
  MA, USA, 1996.
• Astro Teller. The evolution of mental models. In
  Jr. Kenneth E. Kinnear, editor, Advances In
  Genetic Programming, pages 199 -- 220. MIT Press,
  1994.