Friday 25th of April 2003, 19:55
Convergence!
![[It's alive! Aliiiive!]](../images/convergence.jpg)
A couple of hours ago I actually completed my code for Shape Evolution. It works! The genetic algorithm needs some tweaking, but the first results are very encouraging. The graph above shows average score over generation for the evolved designs for one of the first runs of Shape Evolution. There is an obvious improvement over time, and despite some fairly violent fluctuations it seems to be converging.
Aah… This is the culmination of a four year gestation period (i.e. vacation) during which these concepts were mulled over. The venture seems to have yielded fruit. Me satisfied.
Comments
Converging? Gee. The harmonic series is more converging than THAT.
The fluctuations can be (in fact, have been) dealt with by lowering the mutation rate in order to exploit existing traits more than explore the solution space for new ones. I now have a pretty smooth curve that keeps on climbing for several thousand generations. So it takes ages to converge.
Converges to what?
(Let’s not forget Gauss’s zeroth law of convergence: everything converges to itself.)
The initially widely distributed scores for each generation converge to a maximum score value. Ideally, the scores of all individuals at the end should be equal to the maximum score.