Interactive Evolution for Systematic Exploration of a Parameter Space

Authors: Daryl Hepting - Regina

Complete Citation

• Daryl Hepting. Interactive Evolution for Systematic Exploration of a Parameter Space. Intelligent Engineering Systems through Artificial Neural Networks, 2003.

Abstract

Interactive evaluation of a fitness function for a genetic algorithm through direct manipulation is known as interactive evolution. Because it removes the need to specify a fitness function prior to exploration, the user can change evaluation criteria over time. This is especially important when the fitness function is unknown or not easily specified beforehand. This paper describes a system that allows the user to evaluate and evolve small collections of candidate solutions that represent the range of available solutions in order to explore parameter ranges of interest. Although users may be expert with respect to their particular tasks, some may be novices in relation to their software. For them, interactive evolution removes syntactic barriers to the specification of alternative candidate solutions and navigation amongst them. Expert users benefit from exposure to candidate solutions outside of their prior experience and so may find solutions that better meet their needs.

Annotations

The "curse of dimensionality" adversely affects users trying to match datasets with a fitness function. Genetic algorithms are great for high dimensional problems but the interaction between the human and the computer is not fluid. A new system, cogito, is developed "that employs interactive evolution to allow for an initially incomplete articulation of the problem to be refined as the interaction progresses."

Issues addressed:

• The user may be unaware of better solutions that exist outside of their experience;
• Translation between problem semantics and software syntax can be "onerous";
• Users have difficulty formatting output verbally or programmatically;
• Thus have difficulty working with programmers.

• cogito allows a computational steering approach through the data set by showing various views on the high dimensional data, enabling the creation of a mental model of the solution space;
• The combination of genetic exploration and user interaction provides better access to the solution space "by encouraging deeper understanding of the problem through involvement and exploration, and ultimately providing direct access to a variety of solutions";
• Informal user studies found a preference for the interactive interface.

Related Work

• Hepting cites a variety of interesting research.

-- JustinWozniak - 10 Oct 2007