Evaluating an adaptive One-Factor-At-a-Time search procedure within the Mahalanobis-Taguchi System

This paper proposes and evaluates an alternative search procedure to be used within the framework of the Mahalanobis-Taguchi System (MTS). An adaptive One-Factor-At-a-Time (aOFAT) search is employed wherein features are individually removed or added to a classification system. Features are retained only if they contribute positively to the signal to noise ratio. This alternative search procedure is compared with orthogonal arrays and forward selection by means of two case studies. aOFAT experimentation provided greater improvements on the median with the same or fewer design alternatives being explored and also exhibited good ability to generalise to new instances after training. Two mechanisms related to interaction size and synergy help to explain the large benefits of aOFAT search observed in these case studies.