UBC Theses and Dissertations
Value-focused GAI network structure elicitation given a domain Ontology Al Baqui, Abeera Farzana
Making optimal decisions is important yet challenging. A decision maker has to take into account her preferences when she needs to make decisions. Often such preferences are over features in the world and exhibit certain structure. It is possible to exploit this structure by making assumptions of independence, to acquire a decision maker's preferences. However, a compromise must be observed while making these assumptions -too many independence assumptions means the preference model acquired is likely to be inaccurate, however too few assumptions yields an overly complex model. A Generalized Additive Independence (GAI) Network establishes a good compromise between the accuracy and the generality of the model. A GAI network represents a decision maker's preferences in terms of its structure and a set of utilities. Decision theory provides methods of obtaining both the structure and the utilities of a GAI network; however, these methods are too time consuming, error prone and therefore impractical. Several researchers have investigated methods for simplifying the elicitation procedures for GAI network utilities. However, elicitation of a GAI network structure has not received much attention. Value Focused Thinking (VFT) could be a promising solution to this problem, as it can be used to reduce the number of elicitations required to build a DM's GAI network structure as opposed to traditional decision theoretic methods such as standard gambles. VFT proposes that decision-making should start by decomposing a decision maker's values into additively independent objectives (these are the fundamental objectives). VFT shows how to acquire a decision maker's objectives and map attributes of a domain onto these objectives. We assume that the attributes of a domain are represented in an ontology (which specifies the vocabulary to describe the domain). The decision maker can specify how these attributes fulfill their fundamental objectives. It is tempting to build a system where non-expert decision makers, minimally trained in concepts of VFT, Ontologies and GAI networks, may express their preferences by 1) specifying objectives and 2) indicating attributes that fulfill these objectives thus creating a value tree. Once a decision maker's value tree is elicited, it is possible to build a corresponding GAI network. However, it is required that the structure of a decision maker's value tree adhere to the independence assumptions made for GAI networks. We set up an experiment to test whether the structure obtained from eliciting a decision maker's value tree in this manner follows GAI network independence assumptions. We tested this hypothesis in the real-estate domain and found that the resulting structure does not reflect the independence assumptions of a GAI network. We conclude by discussing implications and suggest changes to our original approach.
Item Citations and Data