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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

3.3. DEFlNmON AND MATHEMATICAL FORMULATION<br />

Note thai node W also sends a message back to the evidence node S, but in this case the resulting<br />

belief is equal to the initial evidence and is not affeclcd by the message, i.e. Bel{s) — A{s) =<br />

(0,1). This is because the type <strong>of</strong> evidence was hard evidence. If instead, s<strong>of</strong>t evidence was<br />

used (e.g. X(s) = (0.1.0.9)), the top-down message ns{w) would be able lo modify the belief<br />

in S. This can he understood if we consider that s<strong>of</strong>t evidence contains a certain degree <strong>of</strong><br />

uncertainly, and is therefore susceptible lo being modulated (confirmed or contradicted) by<br />

other information in the network, while hard evidence is assumed lo be irrefutable fact.<br />

3.3.3.8 Example <strong>of</strong> belief propagalion with diagnostic evidence and causal evidence (ex­<br />

plaining away)<br />

In this scenario there is both bottom-up and top-down evidence. When evidence is propagated<br />

from a parent lo a child node, from known causes to unknown effects, this is called causal<br />

reasoning or top-down prediction. The main purpose <strong>of</strong> this scenario is lo illustrate the explain­<br />

ing away effect. For clarity, we omit the detailed numerical calculations for the beliefs and<br />

messages, as the reader can easily follow the example using Figure 3.7, which shows all the<br />

relevant resulting values. These have been obtained using the same Equations ((3.27)-(3.31))<br />

and procedure described in Ihc previous example.<br />

In this example, the prior probability <strong>of</strong> Gales is set equal lo hard evidence asserting the true<br />

state <strong>of</strong> the variable. Consequently, top-down evidence from the high-level cause Gales is com­<br />

bined with boiiom-up evidence from the low-level effect Surfing. This is sometimes referred to<br />

as data fusion. The resulting belief in Waves is higher than in the previous scenario (0,87 vs.<br />

0.722). This makes sense, as the variable Waves is now receiving positive evidence not only<br />

from the child node Surfing, but also from the parent node Gales, providing further support for<br />

the belief Waves=true.<br />

This in lum also leads to an update in the belief <strong>of</strong> Fishing, indicating its value is now even<br />

lower than for Uie previous scenario (0.365 vs. 0.439). This is a direct consequence <strong>of</strong> the<br />

probability <strong>of</strong> Waves being higher due to the new evidence introduced in the variable Gales. In<br />

other words, knowing that Ihc moon i.s in a stale which is likely to generate big waves reduces<br />

ihe chances <strong>of</strong> fishing activity.<br />

100

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

Saved successfully!

Ooh no, something went wrong!