Empirical Methods It seems all fields have their theoreticians and experimentalists, art no less than physics. (Perhaps philosophy is pure theory, then again, perhaps we in AI are philosophy's experimentalists.) Progress depends on an interplay between theoretical and empirical work, but in 1990 interplay was hard to find. This was one conclusion of Paul Cohen's survey of all the papers in the 1990 Proceedings of the AAAI. An equally worrying conclusion was that the empirical branch of AI was more or less innocent of elementary empirical methods, such as exploratory data analysis, statistical hypothesis testing, and modeling. As a result, EKSL has devoted some effort to education, developing new empirical methods, providing software for instrumentation and analysis of programs, and related projects. In 1995, Cohen published Empirical Methods for Artificial Intelligence , a textbook treatment of techniques for studying complex computer programs. To support these techniques we developed an interactive environment called CLIP/CLASP . CLIP is a set of methods for instrumenting programs and collecting trial-by-trial or time series data as programs run. CLIP also helps one define experiments and run them automatically. CLASP is a data analysis package with much of the functionality one expects in conventional statistics packages, plus some tools for resampling (Monte Carlo, bootstrap and randomization) methods. CLIP and CLASP are both Common Lisp applications. | |
|
