next up previous contents
Next: IEMAX Up: Decisions with Locally Weighted Previous: EMAX

Using Locally Weighted Learning to Design Experiments

We have seen how Optex can search a model built from existing data to find desirable parameter settings. Another common problem is determining what experiments to run while collecting data to build a model. This is known as ``Experiment Design.'' One way to do this is something created by statisticians known as Response Surface Methodology, or RSM[2]. It involves performing several experiments in a particular region of interest, and then analyzing the resulting data to determine whether to take additional experiments in that region, or to move the region to a more promising area. RSM is widely used and fairly successful. Its drawback is that much of it must be done manually, and often by a person with a strong statistics background. For more examples on using LWL for experiment design see [6].

We can use the information from LWL models to build an automatic experiment design system. The Optex dialogue box allows us to enter new data points as experiments are run. Then we can iterate over the sequence: 1) ask Optex for an experiment to run, 2) run the experiment, 3) enter the results of the experiment. This is a departure from the traditional RSM method which suggests a whole batch of experiments, and then evaluates them to suggest a new batch. Because LWL can perform evaluations quickly and choose new experiments online, it is able to respond to interesting data immediately rather than having to finish an entire batch of experiments first.

Jeff Schneider
Fri Feb 7 18:00:08 EST 1997