Evidence Integration

There are eight evidence types, as discussed in the previous eight sections, and a single integrated plausibility value needs to be computed from them. All values are assumed to be on the same scale so this simplifies the considerations.

Some constraints the computation should meet are:

• Directly related evidence (property, subcomponent and description) should have greater weight.
• If there is no property, description or subcomponent evidence, then evidence integration produces no result.
• Other relationship evidence should be incremental, but not overwhelmingly so.
• Only types with evidence are used (i.e. some of the evidence types may not exist, and so should be ignored).
• Property, description and subcomponent evidence are complementary in that they all give explicit evidence for the object and all should be integrated.
• If supercomponent evidence is strong, then this gives added support for a structure being a subcomponent. Weak supercomponent evidence has no effect, because the subcomponent could be there by itself, or not be there at all.
• If superclass evidence is strong, then this gives added support for the object.
• Strong association evidence supports the possibility of an object's presence.
• If other identities are competing, they reduce the plausibility.
• As subclasses imply objects, the plausibility of an object must be at least that of its subclasses.

Based on these constraints, the following integration computation has been designed:

Let:
	$evd_{prop}, evd_{desc}, evd_{subcl}, evd_{supcl}, evd_{subc}, evd_{supc}, evd_{ass}, evd_{inh}$
be the eight evidence values, with weightings:
	$w_{prop}, w_{desc}, w_{subc}$
Then:
	$v_{1} = harmmean(\{(evd_{prop},w_{prop}),(evd_{desc},w_{desc}),(evd_{subc}, w_{subc})\})$
	if $evd_{supc} > 0$
	then $v_{2} = v_{1} + c_{supc}*evd_{supc}$ $(c_{supc} = 0.1)$
	else $v_{2} = v_{1}$
	if $evd_{ass} > 0$
	then $v_{3} = v_{2} + c_{ass}*evd_{ass}$ $(c_{ass} = 0.1)$
	else $v_{3} = v_{2}$
	if $evd_{supcl} > 0$
	then $v_{4} = v_{3} + c_{supcl}*evd_{supcl}$ $(c_{supcl} = 0.1)$
	else $v_{4} = v_{3}$
	if $evd_{inh} > 0$
	then $v_{5} = v_{4} + c_{inh}*evd_{inh}$ $(c_{inh} = -0.25)$
	else $v_{5} = v_{4}$
Finally, the integrated plausibility is:
	$min(max(v_{5},evd_{subcl},-1.0),1.0)$

The $-1.0$ and $+1.0$ terms in the final function ensure the result is in the correct range. The weighting constants (0.1 and $-0.25$) used above were chosen to influence but not dominate the evidence computation and were found empirically. Small changes (e.g. by 10%) do not affect the results.

The invocation network fragment executing this function is similar to those previously shown except for the use of a "gated-weight" function unit that implements the evidence increment function for supercomponent, association, superclass and inhibition evidences.