Friday, November 18, 2005
8 comments and a cloud of dust
The post "Competent GAs revisited" (see here) has gotten a string of 8 (!) comments that I would like to comment on. The debate is over whether genetic and evolutionary computation will continue to be an artform, where the practitioner chooses the representation, "appropriate" crossover, mutation, and other operators, or whether competent GAs obviate the need for such choices.
To frame the question, let's ask what a "typical" GA/EC user does in "choosing" operators and codings. First, he chooses an initial representation and operators, gains "experience" with that choice, and guides the choice of subsequent representations and codings with the "lessons learned." The processes of choice and experimentation in this tableau are less than well specified, but essentially the user "data mines" the stream of examples tested to do better.
What does a "competent GA" do? It "mines" the data stream systematically and uses the probability distribution of adapted operators and the population to sample new points. More recent work, uses these structural models to build a fitness surrogate, cheaply and accurately, and new work is pointing in the direction of automatically adapting mutation operators and other local search techniques in concert with the competent GA.
Seen in this way, competent GAs systematize the informal experimentation of the practitioner and replace it with sound decision making according to what is actually learned about the current landscape. In this view, there is nothing special about the actions of the practitioner; it should be possible to replace the informal and unsystematic experimentation of the practitioner with machine-based techniques that do a better job on many if not most problems.
Thus, as competent GAs and their derivatives take hold, I predict that there will be a decrease in the numbers of bad GA/EC papers advocating weird combinations of operators and weird techniques on strange grounds and an increase in landscape learning and adaptation. I certainly hope so.
To frame the question, let's ask what a "typical" GA/EC user does in "choosing" operators and codings. First, he chooses an initial representation and operators, gains "experience" with that choice, and guides the choice of subsequent representations and codings with the "lessons learned." The processes of choice and experimentation in this tableau are less than well specified, but essentially the user "data mines" the stream of examples tested to do better.
What does a "competent GA" do? It "mines" the data stream systematically and uses the probability distribution of adapted operators and the population to sample new points. More recent work, uses these structural models to build a fitness surrogate, cheaply and accurately, and new work is pointing in the direction of automatically adapting mutation operators and other local search techniques in concert with the competent GA.
Seen in this way, competent GAs systematize the informal experimentation of the practitioner and replace it with sound decision making according to what is actually learned about the current landscape. In this view, there is nothing special about the actions of the practitioner; it should be possible to replace the informal and unsystematic experimentation of the practitioner with machine-based techniques that do a better job on many if not most problems.
Thus, as competent GAs and their derivatives take hold, I predict that there will be a decrease in the numbers of bad GA/EC papers advocating weird combinations of operators and weird techniques on strange grounds and an increase in landscape learning and adaptation. I certainly hope so.
