|
Kevin[at]thekrf.com |
Optimizing Flocking Controllers using Gradient Descent
I did this project for Aaron Hertzmann's "Machine Learning for Computer Graphics" course in fall 2003. The basic idea behind it is that I'm applying gradient descent learning to the parameters of a Reynolds-esque flocking model, so that an animator can control the flock using more intuitive values. So, instead of manually balancing the cohesion, separation, etc parameters, the animator can just specify things like "mean distance from neighbours" parameter.
The system worked pretty well, barring some numerical stability issues. The files from the project are available below.
Source code (warning- it's messy!)
Example programs (require gsl and glut dlls)