The classical design of experiments (DOE) [1] essentially implements the following approach:
- perform a sampling of the experimental space varying each experimental variable individually, holding the other variables constant
- use a model to detect the directions in the experimental space (this may be, for example, an analysis of principle components)
- pick the experimental subspace that shows the most variation
- perform a full factorial exploration (exhaustive search) of this subspace.
For extremely high dimensional experimental spaces and very rugged response functions, this approach typically suffers from reducing dimensionality prematurely, and not taking full advantage of structure that may be discovered gradually by a sequence of carefully planned experiments.
The techniques for design of complex experiments developed in PACE seek to overcome these shortcomings. Typically these techniques use a combination of nonlinear modeling tools to exploit structure along with exploration using random sampling from shaped distributions designed to avoid regions of the experimental space already sampled. The combination of exploitation and exploration is a crucial element in contemporary methods for evolutionary design of experiments.