Date & Time: Monday, April 07, 2014, 11:00-12:00.
Venue: Room 113
Title: Search Designs for Model Selection
Speaker: Kashinath Chatterjee, Visva Bharati University
Abstract: The major objective of factorial experiments is to provide information on interaction effects, apart from the general mean and all the main effects. When the number of factors is large and little prior knowledge is available on the various factorial effects, conventional fractional factorial experiments that are capable of estimating interactions require too many observations to be economically viable. The effect hierarchy principle is one of the most important principles in experimental design. The principle states that (i) lower-order factorial effects are more likely to be important than higher-order ones, and (ii) effects of the same order are equally likely to be important. To overcome this problem, interaction effects of higher orders are frequently dropped from consideration in an experiment and are assumed to be negligible, often without substantive justification. Also, a fractional factorial design is often run in which only main effects can be investigated (a main effects design). These practices can result in an inadequate understanding of the joint action of the factors on the response and in poor predictive models. In industrial experimentation, it is noted through many research papers that the loss of information on interactions is a serious problem, because a key tool for product improvement is the exploitation of interactions between design (control) factors which can be set in the product specification and noise factors which cannot.
It is to be remarked that before performing an experiment, it is generally very difficult to predict the interaction effects that are possibly present in the model. In such a case the designs capable of just estimating the possibly present interaction effects may not be good enough to serve the objective of the experimenter. Moreover, in such a case if one considers only a main effect model, the outcome of the analysis will mislead the experimenter. To cope with this problem, Srivastava (1975) introduced a design criterion that seeks to maximize the ability to discriminate between models. The main objective of this presentation is to introduce the notion of Search Designs, pioneered by Srivatava (1975), and their applications to model selections in fractional factorial experiments.