A simulated annealing technique for generating artificial data to assess concept learning algorithms

Abstract

Artificial data can be used instead of real world data to derive rigorous results for the average case performance of concept learning algorithms. Empirical testing with artificial data can be used to calculate the average case performance with arbitrary precision. This paper presents an artificial data-generating program, called ADG. The program makes use of simulated annealing to generate its data sets. ADG is distinctive because users can define as few or as many data characteristics as they wish. In most other artificial data generating systems the user must specify a complete distribution of the examples, which can be difficult and tedious.