Next: The ridge problem
Up: From Genetic Algorithms To
Previous: How to discover
This chapter will explain the application of the third key idea as a means of
improving efficiency:
- Keep track of the directions of recent successes, so as to probe
preferentially in the direction of most rapid ascent.
Last chapter, we introduced the difficulty of traversing narrow ridges in
the search space. In objective functions for which the gradient
information is not available, finding a favorable direction to move is a
hard problem. I propose an adaptive solution, where the algorithm uses its
best guess for the gradient direction, and updates this guess after
successful moves. This approach improves efficiency
substantially---sometimes by an order of magnitude---and is scalable to
high dimensional problems.
Deniz Yuret
Tue Apr 1 21:38:29 EST 1997