Bernal S D_2010.pdf - University of Plymouth
Bernal S D_2010.pdf - University of Plymouth
Bernal S D_2010.pdf - University of Plymouth
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
4.5. FEEDBACK PROCESSING<br />
George and Hawkins 2009).<br />
4.5.2 Multiple parents<br />
As desciibed in Section 3.3.4 the number <strong>of</strong> elements <strong>of</strong> the CPT P{X\U],-- ,(/^) is exponen<br />
tial lo the number <strong>of</strong> parents, N, as it includes entries for all possible combinaiions <strong>of</strong> the states<br />
in node X and its parent mxles. e.g. given kx — ku — A.N = 8, the number <strong>of</strong> parameters in<br />
the CPT is 4 • 4** - 262,144, Additionally, the number <strong>of</strong> operations to compute the belief is<br />
also exponential lolhe number <strong>of</strong> parents, more precisely it requires ^ sums and N •k'^ product<br />
operations. The exponential growth to the number <strong>of</strong> parameters and operations resulting from<br />
the comhinalion <strong>of</strong> multiple parents is illustrated in Figure 4.14.<br />
4.5.2.1 Weighted .sum or compatible parental configurations<br />
To solve the problem <strong>of</strong> Ihe large number <strong>of</strong> entries in the CPT we implemeni the weighted<br />
sum <strong>of</strong> simpler CPTs based on the concept <strong>of</strong> compalihle parental configurations (Das 2004)<br />
described in Section 3.3,4, This method ohiains a kx x *(/ CPT. P[X\Ui), between node X and<br />
each <strong>of</strong> its N parent nodes, and assumes the rest <strong>of</strong> the parents. Vj, where j / i. are in compatible<br />
states. The hnal CPT F(X|(/i,--.(/AT) is oblained as a weighted sum <strong>of</strong> the A" P{X\JJi) CPTs.<br />
The total number <strong>of</strong> parameters required to be learned is therefore linear with the number <strong>of</strong><br />
parents, more precisely, kx -k-N • N. Using llie values <strong>of</strong> the previous example, the number <strong>of</strong><br />
elements is now 4 • 4 • 8 - 128, several orders <strong>of</strong> magnitude smaller than the previous result.<br />
This is illustrated in Figure 4.15.<br />
The Learning section (4.3) described 1) how to obtain the weight matrices between a parent<br />
node and its children, and 2) how to conven these weight matrices into individual CPTs for<br />
each <strong>of</strong> the child nodes. The resulting CT'T is precisely in the form required to implement ihe<br />
weighted sum method, i.e. for each child node X there are N CPTs <strong>of</strong> the form /'(X It/,), one for<br />
each <strong>of</strong> its parents. These can then be combined to form the [\nii\P{X\U\,--- ,UN).<br />
4.5.2.2 Sampling fnmi parent nodes<br />
To reduce the excessive number <strong>of</strong> operations required lo calculate the belief, only the i^niiK<br />
states, with the highest values, from the NmoK ^ messages, with the higher variance, are used<br />
177