|Location||Monash University, Faculty of Information Technology|
|Eligibility||Open to international applicants|
Eliciting and Reasoning with Multi-Criteria Utilities
PhD Project: Eliciting and Reasoning with Multi-Criteria Utilities
Supervisors: Ann Nicholson, Kevin Korb, Steven Mascaro and Thang Cao (DSTG)
Utility is a theoretical entity that helps us explain decision making under uncertainty. In classic decision making (e.g., the idealized agents of much of economic theory) rational decisions are understood as efforts to maximize the expected utility of actions -- i.e., the probability-weighted value of the actions' outcomes. This often works well as a normative standard. At the margins, however, the analysis encounters difficulties. One of them arises when distinctly different kinds of utility are involved in the same decision problem. A classic kind of case is that of human lives versus dollars: how do we weigh lives against, say, property damage in bushfires?
Multi-Criteria Decision Analysis (MCDA) aims to address this kind of problem through one or another kind of formal analytical method, with a view to finding some kind of decision which does well across all the utility dimensions -- i.e., lies on or near the Pareto frontier. For example, Evolutionary multi-objective optimization (EMO) uses evolutionary simulation to find a population of candidate decisions which are near Pareto optimal across all the utility dimensions. In this project we will examine a different set of issues: how to represent and combine the utilities elicited from multiple experts across all the utility dimensions. When experts conflict in their utility estimates one approach is to "force" a consensus between them, for example, using Delphi methods. Another is to somehow weight the experts and provide a weighted average. This project will look at these methods, but will also consider new approaches. Specifically, it is proposed that some utility dimensions are related to others, without being readily interconvertible with each other. For example, the quality of engineering is related to the quality of manufacture of high tech products, but they are not the same, nor are they a simple function of one another. While we may rate these two qualities independently of each other, the results are likely to be (probabilistically) dependent. This project will aim to learn and formally represent those interdependencies using Bayesian network models. It will then incorporate the resultant probabilistic representations of utility in larger scale Bayesian decision networks. This approach will be experimentally compared with the alternatives above (and others).
The methods developed in this project will be applied to complex procurement and selection problems within DST Group to land combat vehicles. To our knowledge, treating uncertainty dependencies has never been reported in conjunction to real-life multi-criteria decision-making problems. A scholarship top-up of $10K plus travel funding is available.
See our full disclaimer