Bayesian framework for improved R&D decisions
Anand, Farminder Singh
MetadataShow full item record
This thesis work describes the formulation of a Bayesian approach along with new tools to systematically reduce uncertainty in Research&Development (R&D) alternatives. During the initial stages of R&D many alternatives are considered and high uncertainty exists for all the alternatives. The ideal approach in addressing the many R&D alternatives is to find the one alternative which is stochastically dominant i.e. the alternative which is better in all possible scenarios of uncertainty. Often a stochastically dominant alternative does not exist. This leaves the R&D manager with two alternatives, either to make a selection based on user defined utility function or to gather more information in order to reduce uncertainty in the various alternatives. From the decision makers perspective the second alternative has more intrinsic value, since reduction of uncertainty will improve the confidence in the selection and further reduce the high downside risk involved with the decisions made under high uncertainty. The motivation for this work is derived from our preliminary work on the evaluation of biorefiney alternatives, which brought into limelight the key challenges and opportunities in the evaluation of R&D alternatives. The primary challenge in the evaluation of many R&D alternatives was the presence of uncertainty in the many unit operations within each and every alternative. Additionally, limited or non-existent experimental data made it infeasible to quantify the uncertainty and lead to inability to develop an even simple systematic strategy to reduce it. Moreover, even if the uncertainty could be quantified, the traditional approaches (scenario analysis or stochastic analysis), lacked the ability to evaluate the key group of uncertainty contributors. Lastly, the traditional design of experiment approaches focus towards reduction in uncertainty in the parameter estimates of the model, whereas what is required is a design of experiment approach which focuses on the decision (selection of the key alternative). In order to tackle all the above mentioned challenges a Bayesian framework along with two new tools is proposed. The Bayesian framework consists of three main steps: a. Quantification of uncertainty b. Evaluation of key uncertainty contributors c. Design of experiment strategies, focussed on decision making rather than the traditional parameter uncertainty reduction To quantify technical uncertainty using expert knowledge, existing elicitation methods in the literature (outside chemical engineering domain) are used. To illustrate the importance of quantifying technical uncertainty, a bio-refinery case study is considered. The case study is an alternative for producing ethanol as a value added product in a Kraft mill producing pulp from softwood. To produce ethanol, a hot water pre-extraction of hemi-cellulose is considered, prior to the pulping stage. Using this case study, the methodology to quantify technical uncertainty using experts' knowledge is demonstrated. To limit the cost of R&D investment for selection or rejection of an R&D alternative, it is essential to evaluate the key uncertainty contributors. Global sensitivity analysis (GSA) is a tool which can be used to evaluate the key uncertainties. But quite often global sensitivity analysis fails to differentiate between the uncertainties and assigns them equal global sensitivity index. To counter this failing of GSA, a new method conditional global sensitivity (c-GSA) is presented, which is able to differentiate between the uncertainties even when GSA fails to do so. To demonstrate the value of c-GSA many small examples are presented. The third and the last key method in the Bayesian framework is the decision oriented design of experiment. Traditional 'Design of Experiment' (DOE) approaches focus on minimization of parameter error variance. In this work, a new "decision-oriented" DOE approach is proposed that takes into account how the generated data, and subsequently, the model developed based on them will be used in decision making. By doing so, the parameter variances get distributed in a manner such that its adverse impact on the targeted decision making is minimal. Results show that the new decision-oriented DOE approach significantly outperforms the standard D-optimal design approach. The new design method should be a valuable tool when experiments are conducted for the purpose of making R&D decisions. Finally, to demonstrate the importance of the overall Bayesian framework a bio-refinery case study is considered. The case study consists of the alternative to introduce a hemi-cellulose pre-extraction stage prior to pulping in a thermo-mechanical pulp mill. Application of the Bayesian framework to address this alternative, results in significant improvement in the prediction of the true potential value of the alternative.