Empirical Methods for Artificial Intelligence by Paul R. Cohen is an unusual text. Cohen advocates the use of statistical methods for analyzing AI programs and interpreting their performance-a view which is more accepted in principle by the AI community than followed in practice. One might expect such a text to follow a standard form for presenting statistical methods, providing simple rules or decision trees to determine the appropriate test. Alternatively, one might expect a deeper analysis of the statistical methods that included proofs of why various inferences could be made. Cohen follows neither of these traditional paths.
The text is problem-based. There are no proofs in the text proper, although a few simple ones that rely only on algebra are included in appendices. There are no simple recipes for deciding on which statistic to apply. Instead the reader is engaged in the artful and intelligent application of statistics in order to make sense of the data. Through the process of starting with data and then trying to make reasonable inferences from that data, Cohen introduces various statistical concepts. Each statistical concept is grounded in specific problems. Cohen carefully indicates the underlying assumptions and potential misuses of each statistical method. While mathematical equations are necessary for correct calculation and precision, this precision is only achieved when the appropriate implicit context is understood. The text emphasizes the need to determine the context, the assumptions and goals, before applying any statistical method.
Example domains come from a wide range of AI systems. For those with an AI background, this increases the pleasure in reading this text. However those without some AI knowledge will find the examples too difficult to follow. Terms like confidence levels, branching factor, backtracking, A* search, decision trees, plan libraries, frames, forward and backward chaining are used without definition. The spectrum of AI systems discussed includes natural language processing, expert systems, planners, classifiers and theorem