next up previous
Next: Concluding Remarks Up: Constraint satisfaction Previous: Least Action Principles

Multiresolution satisfaction of constraints

Constraints can also be satisfied by gradient descent on a potential surface representing their lack of satisfaction (Sects. 4.2 and 5.2). However, a problem with relaxation techniques is that they may get trapped in local minima. One way to avoid this is to do the relaxation on a multiresolution representation of the potential function. At low resolution, local minima will tend to be averaged away, so relaxation on a low-resolution representation will tend to move toward the global minimum. By gradually increasing the resolution, the system can be allowed to settle into a more accurate representation of the global minimum. This can be accomplished, for example, by gradually activating the higher frequency bands of a Gabor or wavelet representation of the potential surface.


Bruce MacLennan
Wed Oct 2 16:55:07 EDT 1996