COMISEF -- T3 Tutorial
We review empirical methods for the analysis of
optimization heuristics with applications in econometrics. In the first
part, we report from the computer science literature general
principles for a correct organization of computational experiments. We
then distinguish different scenarios of analysis and model them in
statistical terms. Simple graphical representations of results for the
comparisons of few well defined algorithms are suggested.
In the central part of the tutorial, we focus on algorithm
configuration and parameter tuning. Several empirical
methods to carry out these tasks have emerged in the past years. We
review the general ideas underlying ANOVA, regression trees, racing
procedures, search methods and response surface methodology.
In the last part, we consider the characterization of performance
distributions. For run time distributions, we summarize three methods of
model fitting, including censored distributions and extreme value
theory, and outline a practical application of the models to the
decision of restart times. For solution quality distributions, we review
an extreme values procedure for the estimation of optimal solutions.
Overall, the emphasis is put on the visualization of results by means of
graphics that facilitate the exploration of data and the communication
of results in an informative and comprehensive way.
The hands on session comprises two exercises. The first exposes the
participants to a practical experience of empirical algorithm
configuration. The second focuses on model fitting with censored data.