5.
Process Improvement
5.2. Introduction 5.2.1. What's wrong with "traditional" nonstatistical approaches?


By varying the inputs, we see changes in the output
The "One Factor at a Time" Approach to varying
the inputs
The OFAT approach does not take interactions into account

A Process "Black Box" Model
Suppose that we have a process with two inputs and one output (Y), as shown below. We do not know what mechanism is inside the process ‘box’ so we call it a ‘black box.’
Of course we might derive a universal physicochemical model for the process, but this tends to be expensive and best left to Nobelprize winners. We nevertheless have to understand the best setting of the process at least in a local area of operation, so we do empirical experiments. One such experiment could be: Hold all inputs but one fixed, and see the best result when the one free input is varied. Fix that input at that ‘best’ value. Then vary one other input. See where the best output is now. Fix the second input at that ‘best’ value, and so on until we run out of input factors. This is called a ‘One factor at a time’ (OFAT) experiment, and is practiced widely. It used to be thought that this was the only ‘scientific’ approach. OFAT experiments will work if the true model inside the black box looks generally as follows.
This model is ‘flat’ in all dimensions, and is called a ‘main effects’ model. No matter at which point on the surface one begins, increasing an input always has the same effect on the output. There are no ‘interactions’ between inputs. In this model, X_{1} at its highest setting will always
give best Y, and similarly X_{2} at its highest setting
will always give best Y. (We are assuming highest Y is ‘best.’)
OFAT experiments will not work if the true model inside the black box looks something like: An interactions model of this type has a ‘twist’ to the response surface. This means that the response Y_{1} to, say, X_{1} will be different depending on the setting of X_{2} (and viceversa). Interactions can also be plotted in two dimensions as in the following two examples. In this model, if one started experimenting with X_{2} set at its lowest value, X_{1} would have to be moved toward its lowest value to get a high Y. On the other hand, if one started out with X_{2} fixed at its highest value, X_{1}would have to be moved up to get a high Y. We do not know if X_{1} and X_{2} both set at low will give a better Y that X_{1} and X_{2} both set high. In a model with many inputs, the twofactor interactions such as X_{1}*X_{2}are usually of interest, as they might point the way to a better product with minimal additional expense. OFAT experimentation leaves us in the dark about factor interactions. 