PRIOR WORK IN GP

A New Field:

Principal GP Investigators

• In an early example of GP, W. Daniel Hillis( COE of Thinking Machines) used a Connection Machine to generate code for sorting numbers. The code was only slightly less efficient that the best ever developed by a human.
• John R. Koza (Stanford University) has done the most work in the GP field. Koza [see reference below] has demonstrated that GP can be applied to an impressive variety of problem types including:
  • Evaluating a boolean function
  • Planning
  • Optimal Control
  • Solving a pair of consistent, non-determinant linear equations in 2 unknowns
  • Learning an 89 step path in a 32x32 torroidal grid
  • Sequence induction (viz., identifying a mathematical expression that generates a predetermined sequence of values, e.g., 1,2,5, 10, 17, 26,...)
  • Symbolic integration and differentiation
  • Solving differential equations
  • Empirical discovery (Finding the mathematical relationship underlying empirically observed values from a noisy environment)
  • MinMax strategy for a 2 player, zero-sum game

Conclusions From Prior Work

• GP is a feasible method for generating programs (in a reasonable amount of computational time) to perform specific and meaningful tasks
• GP is not a random search process. Koza and Peter Angeline have compared GP with random searches and shown that GP consistently achieves results virtually unattainable by random search.

Limitations of Prior Work

B. Work to date has required considerable human involvement to tailor a GP process. Thus,the GPer is required to define (and program for)

• the set of terminals
• set of primitive functions
• the fitness measure
• the parameters for controlling the run
• the method for designation a result
• the criterion for terminating a run.

C. Results from one GP process, as defined in step B., do not readily form building blocks for other GP processes, especially those in new problem domains.

D. At completion of GP run, the GPer must frequently manually interpret the results from the form of the resulting program.

E. Considerable work remains before GP can be applied to large and multi-faceted problems

Reference: Koza, J. R., "The Genetic Programming Paradigm: Genetically Breeding Populations of Computer Programs to Solve Problems", Dynamic, Genetic and Chaotic Programming, by Branko Soucek and the IRIS Group, John Wiley & Sons, Inc, New York, 1992.

<< Previous Page    Next Page >>

Return to GEMS Home Page

Return to Arktist Home Page

Updated August 27, 2008