1.1 Introduction

Design of experiments  is the art of organizing an information-gathering campaign where variation is present. All the steps before, during and after the experimental process itself are to be properly chained in order to efficiently separate signal and noise, i.e. separate a pertinent and reproducible part from a non-pertinent and random remainder.

With this goal in mind, Dagnelie (2003) states that "experience design" would be a far better description of what is to be done. The aim is not to juxtapose disparate activities but, on the contrary, to organize various activities in order to construct a proof of something. In the same vein, "a clinical trial" may involve hundreds of patients... but must remain one unique and coherent activity.

On the other hand, expression "experience design" is nowadays broadly used to describe how Big Brother can design the experience and the feelings of the rank and file people. The currently used wrong plural form "design of experiments" has therefore the merit to assert some difference between a kind of lies and a kind of truth-seeking.

In any case, maintaining an overall view on the whole process is even more important nowadays, since computing is increasingly faster and cheaper while the process of experimentation by itself remains slow and expensive. The data mining process can only extract what is contained in the data and from poor data, only poor conclusions can be drawn.

Therefore, when a campaign of experimentation is designed, everything must be done in order to obtain a significant result rather than a disappointing "things should have been conducted otherwise".

Among the questions that must be answered by advance, we have :

1. What is the economic impact of the problem to be solved ?
2. What is the budget of the study ?
3. What is the history of the previous attempts... and why did they fail ?
4. What do we know, how do we know it and what is the quality of all these knowledge and beliefs ?
5. Who else should be consulted ?
6. What will be done with the data after collection ?

It is essential to state explicitly everything. All decisions must be tracked... and moreover all the reasons for those decisions. We call this step "asserting". This part of the job is time consuming... but being forced to replay everything is even more time consuming : without theory, before and after, there is no experiment, only experience, i.e. feeling.

In order to obtain such an efficient experimental design, a not so minimal knowledge is required about what can be extracted from a given experimental series ... and about what cannot be retrieved if the series has been ill-conceived.

Thereafter, the following chapters will discuss how to choose efficiently an experimental series in order to detect the most influential factors (screening) or in order to adjust these factors (improving).

1.1.1 Six Principles

In Natrella (1963), the following six principles are stated for experimental design as applied to process modeling :

1. Capacity for Primary Model
2. Capacity for Alternative Model
3. Minimum Variance of Coefficient Estimators
4. Sample where the Variation Is
5. Replication
6. Randomization