Decision Support Systems in Agriculture: Some Successes and a Bright Future - Efficient Decision Support Systems: Practice and Challenges from Current to Future - page 303

diagnosis when no individual piece of information was sufficient to provide a call for action.
It was possible to include a consistency check, if mutually contradictory facts were
observed. For example, if the probability of a nutrient deficiency for fact A was 0.9 that a P
deficiency was likely (where 0 means certainty of no deficiency, 0.5 means complete
ambivalence, and 1.0 means total certainty) and fact B had a probability of 0.2, then we have
a situation of conflicting evidence. A rule was written to send a message to list in the output
that a serious contradiction is occurring.
Table 1. Considerations in developing diagnostic questions.
We suggest that the best diagnostic information/ tools/
questions are those that build on the common knowledge that
on-site managers (e.g. farmers) have readily available together
with simple measures, both qualitative and quantitative, of
fundamental characteristics of the production system:
-The tool/question should be simple to use by lay persons.
-Results of the tool should be quick, such as the simple
observation of a symptom or property in the field.
-Cost of the tool/question should be low or of no cost.
-The tool/question should be reliable as it should reliably
indicate what action is to be taken.
Observations:
-Sometimes the result of the tool/question is that more
expertise is required.
-Incomplete or imperfect data should not completely
invalidate the diagnosis.
-The tool/question should take full advantage of the farmer,
producer, or field observer's observation and knowledge.
-The tool/question may lead to improved, better diagnostic
tools.
(Questions developed in a TPSS 650 Soil, Plant, Nutrient
Interactions by students N. Osorio, X. Shuai, W. Widmore, R.
Shirey. University of Hawai`i at Manoa)
We encountered two disadvantages of using the Bayesian accumulation of probability
framework: 1) Much of our evidence and multiple observations or measurements were
highly correlated or multicollinear. The multicollinearity contrasts with the assumed
condition of independence in classic Bayesian evidence accumulation and thus the
calculated cumulative conditional probabilities were in slight error depending on the
degree of multicollinearity. 2) One could have strong evidence both for and against a
condition as well as weak evidence for and against the condition, or even a complete lack
of information, all of which would combine to a value of 0.5. As a result, strong, but
conflicting, evidence is wholly discounted. One of our inadequate solutions to this
situation was to monitor evidence and when evidence for and against a particular
outcome differed substantially, a message was attached to the conclusion warning of the
information conflict.