08.02.2013 Views

Bernal S D_2010.pdf - University of Plymouth

Bernal S D_2010.pdf - University of Plymouth

Bernal S D_2010.pdf - University of Plymouth

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.3. DEFINITION AND MATHEMATICAl. FORMULATION<br />

Therefore the link is defined using ft;;, +ki;^ + ... + kn^ probabilily distributions over X for<br />

Ihe different conipaiible parental confignraiinn. Note Ihese grow linearly with the number <strong>of</strong><br />

parents. Given a set <strong>of</strong> weights W|,...,ii'^', which quantify the relative strength <strong>of</strong> the parent<br />

nodes' inlluence on ihe child node, the entries <strong>of</strong> the CPT can be generated using the following<br />

weighted sum expression,<br />

P{X\UU...,UN)= Y. ^'rP{x\(Comp(Vi = Ui)}) (3.34)<br />

i=\..N<br />

It is important to stress thai {Coinp{U, = «,)} is a parental conliguration in the menial model<br />

<strong>of</strong> the expert where he has chosen to focus on the slate M; <strong>of</strong> parent (/,, while the rest <strong>of</strong> the<br />

states <strong>of</strong> the parents are perceived in his judgement to be in compalible stales with ii,. This<br />

helps ihe expert to simplify his mental model in order lo judge the possible cflect. It does not<br />

mean ihat compalible parental configurations are the only ones lo be found in reality, bul these<br />

are assumed to be more common or normal.<br />

The method described here proposes combining the probability distributions <strong>of</strong> X given com­<br />

patible parental conligurations, to calculate the states <strong>of</strong> X given invompanbU: or less common,<br />

parental configurations, by using the weighted sum expression in Hquaiion (3.34). This can<br />

be understood as a kind <strong>of</strong> interpolation mechanism Ihat exploits the known dala points. Das<br />

(2004) makes use <strong>of</strong> information geometry to demonstrate how these weighted sums capture<br />

Ihe expert.s' judgemental strategy. The method is being employed to design strategic military<br />

applications for the Australian Deparlmenl <strong>of</strong> Defence.<br />

Although the method was derived for populating CFTs using human experts, theoretically il<br />

can be extended to systems that obtain their information using training dala with supervised<br />

learning methods. One such domain is hierarchical object recognition, where, due to the greal<br />

ovedap between receptive fields, parent nodes show contextual inlerdependency and can there­<br />

fore exploit this technique. This is discussed further in Chapter f>. where a toy example is used<br />

to illusiraie the concept.<br />

no

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!