VUWLGP

What is VUWLGP?

VUWLGP is a register-machine linear genetic programming (RMLGP or LGP) package developed at the School of Mathematics, Statistics and Computer Science, Victoria University of Wellington. LGP works by creating an initial population of random programs. Of these, some will be better for the target problem than others. These programs are selected and undergo evolutionary operations which can improve the program. As a result of this from generation-to-generation the best program will get better and better.

In LGP, each program consists of a sequence of register machine instructions and perhaps some simple conditionals, e.g., in a C-style representation:

			r[2] = r[1] * 0.2345;
			r[0] = r[1] + r[0];
			if(r[0] < r[2])
			r[1] = 0.6547 + 05624;
			.
			.
			.

At program termination one or more of the register values is interpreted as the output of the program and given semantic meaning. Programs represent directed acyclic graphs whose structure evolves flexibly and mutably during the evolutionary process.

VUWLGP was developed and then extended in 2005 by Christopher Fogelberg as part of his Computer Science Honours programme. This package is released under the GPL (see gpl.txt in the root directory for the license). Christopher Fogelberg's other contact details can be found on his webpage (here).

Directory Structure

This VUWLGP release is structured into 5 directories.

• doc
  Contains this documentation.
• baseSrc
  Contains the base source and abstract classes used by VUWLGP.
• implementations
  Contains some example implementations of the abstract classes.
• exampleProblems/symreg
  Contains an example of a simple symbolic regression for the sine function. This example shows how an implementation of VUWLGP for a particular problem can easily incorporate classes created just for this problem.
• exampleProblems/multiclass
  Contains a very simple multi-class classification example for a three feature problem in which the correct class is whichever feature is largest. This example shows how, if the classes already exist as part of the provided implementations of VUWLGP, a configuration of VUWLGP which addresses a new problem can be trivially created.
• unitTests
  Contains some of the unit tests used when developing VUWLGP. If further modification are being made to VUWLGP these may be useful as examples or base code for further tests.

Documentation Structure

This documentation is organised into two key sections. In the first section, Using VUWLGP, how VUWLGP can be customised for a problem is described, with a particular emphasis given to multi-class classification (although VUWLGP can be used for any problem).
Functions and classes not mentioned in the documentation below should not need to be modified in order to address a new problem. The code itself is commented though - read this if you need to learn more about these functions!

Using VUWLGP

To use VUWLGP, you need to do the following (modifying or extending one of the examples given is probably the easiest approach). The classes which need to be developed or used are first given.

1. Use or implement a new subclass of IPopulation.
2. Define the fitness environment. This usually involves:
   • Creating a new kind of IFitnessCase, or using an existing one.
   • Creating a new subclass of IFitnessMeasure. The fitness measure is a member of and used by the program class.
3. Create a subclass of IProgram which uses the new IFitnessMeasure.

• Defining a new type of instruction operation (e.g. +, -, sine) - here
• Defining a new type of instruction argument (features, registers and ephemeral constants are already provided in implementation) - here.

Christopher Fogelberg
fogelbchri@mcs.vuw.ac.nz or cgf.unimail@syntilect.com