 |
Decision Support for Food Related Risks: Identification of Inference Rules and Key Risk Attributes Using Fuzzy Logic. David Waters and Peter Grindrod, QuantiSci, Chiltern House, 45 Station Road, Henley-on-Thames, Oxon, RG9 1AT
ABSTRACT
Development in the field of individual and group perception of food related risks requires the identification of key risk attributes and their critical combination to form robust inference rules. In most psychometric studies statistical techniques have been employed to analyse survey data. Their limitations are well known, for example permitting exchangeability, and therefore, these passive snapshots are often of restricted management use. The processing should be done in a way that is sympathetic to the human perception process and identifies whether any underlying logic (inferential rules, linking key attributes to response) is exhibited, which could be anticipated or managed in the future.
We analyse the process by which individuals (including food
safety experts) and groups (e.g. based on ethnic, religious and
cultural backgrounds) formulate their response to perceived risks
by applying an Al technique based on fuzzy logic. This framework
is suited to the subjective, context-dependent, non unique,
classification of attributes and concepts involved in a cognitive
approach to perception. Inferences are represented by memberships
in hierarchical fuzzy sets. Such a hierarchy is inferred from the
relative strengths of different attributes. In order to identify
good rules we introduce a measure of success defined for an
arbitrary rule, which can be optimised over the set of all rules
by using a genetic algorithm. Once a 'best' rule has been fitted
to the data, we check its significance by considering its
performance with reference to a distribution of other 'best'
rules, obtained using the same target set along side an
equivalent number of random attribute sets. The main objective is
to test the homogeneity of the inter- and intra-group response
and see whether robust rules can be used for effective
communication and information dissemination strategies.
1 INTRODUCTION
In order to adopt a robust management strategy concerning
food related risks it is vital that individuals' perceptions are
extracted and presented within a quantitative framework. The main
objective is to establish the inter- and intra-group homogeneity
of response. Identifying any underlying logic, i.e. inferential
rules, linking key attributes to response / behaviour will
support the design of an effective communication strategy.
2 RISK PERCEPTION
There have been numerous psychometric studies in various
fields (nuclear, chemical, aviation etc) concerned with the
perception of risks to human life and health (Kenney, 1995;
Cohen, 1995; Slovic, 1987; Covello et al., 1988). For the purpose
of this paper it is sufficient to state the following points:
3 SURVEY DESIGN
Eliciting expert and stakeholder perception data requires an
appropriate questionnaire design. All relevant questions must be
asked, appropriate response categories offered and suitable
ordering chosen to ensure the interviewee responses are not
biased (N.B. In order to constrain the problem, question rotation
will be omitted). An example questionnaire is given in Appendix
A.
Data elicitation from consumer groups will be accomplished by
the 'in-home' test. This procedure is cost effective and ensures
a representative sample is extracted. 100 respondents will be
selected in each group. Each interview will take up to 20 minutes
to elicit 'tick-box' responses from approximately 20 questions.
Elicitation of responses from experts will be conducted
through 'in-office' interviews. Fifty experts will be sampled
consisting of 25 high-level (i.e. government, leading food safety
controller, etc.) and 25 low-level (i.e. hotel owner, publican,
restaurant, etc.) experts. Respondents will be carefully selected
to ensure a representative sample is obtained.
A pilot sample will be used to filter the number of scenarios
offered to respondents. Nine food risk scenarios will be
presented during the pilot test and the top three will be used in
the full survey. Of course this might mean that some of the
responses may be redundant due to their lack of understanding or
familiarity with the risk.
4 ANALYSIS METHOD
In order to derive a model (a set of inferential rules) in
order to rationalise individuals' perception of food related
risks it is important that the quantitative framework is
sympathetic to human decision process. It is the authors opinion
that the logical combination of attributes during the perception
process, is suited to a fuzzy mathematical analysis. The need for
such an AI approach also arises from the subjective non
quantitative data, and the unknown interdependencies of different
attributes.
In classical set theory elements are either in a set, with a
membership of unity, or not in a set, with a membership of zero.
Membership is determined by evaluating the truth of some well
defined proposition (e.g.: the set of all numbers greater than or
equal to three). In fuzzy logic, sets are defined by uncertain or
ill-defined propositions (e.g.: the set of all tall people), the
corresponding membership is allowed to take values on a sliding
scale, from zero to one, reflecting the subjective uncertainty in
the definition as well as in the information concerning the
candidate element. In fuzzy logic, memberships of intersection
and unions of sets are evaluated using minimum and maximum rules
respectively.
For each attribute and target set respondents will be asked
to assert a level of agreement or disagreement with the
proposition presented in the questionnaire; thus in effect
defining the fuzzy set of those options for which the respondent
perceives the proposition to be true. The ultimate aim is to map
the risk attribute questions onto the target sets (see Appendix
A).
The scale of agreement will be mapped directly onto the
discrete scale (0.1, 0.3, 0.5, 0.7, 0.9). Using another dataset,
alternative mappings were investigated and found to have little
impact on the results.
Two summary or target statements will be tested (i.e.
'Overall I am not concerned about this risk' and 'I avoid
exposure to this risk'). These propositions will define
memberships of target sets, representing the output of the
respondents perception and decision processes. In considering the
responses to the survey, it is the relationships between the
attributes, and their logical combinations employed in defining
the target attributes which is of central interest.
In order to identify good rules we introduced a measure of
success defined for an arbitrary rule (Grindrod et al.,
submitted). This can be optimised over the set of all possible
rules by using a genetic algorithm. Once a 'best' rule has been
fitted to the data, we check its significance by considering its
performance with reference to a distribution of other 'best'
rules, obtained by repeatedly using the same target set along
side an equivalent number of random attribute sets.
5 DISCUSSION
The pilot test is in progress and the full survey will be
completed before the end of May 1996. Using the intelligent
hybrid system which incorporates fuzzy logic and genetic
algorithms, illustrative results will be presented at the
conference. Results using other datasets are encouraging.
6 REFERENCES
Cohen B L (1995) How Dangerous is Low Level Radiation. Risk
Analysis, 15, pp. 645-654.
Covello V T, Sandman P M, and Slovic P (1988) Risk
Communication, risk statistics and Risk Comparisons; A Manual For
Plant Managers. Washington DC, Chemical Manufacturers
Association.
Keeney R L (1995) Understanding Life Threatening Risks. Risk
Analysis, 15, 6, pp. 627-637. Slovic P (1987) Perception of
Risk. Science, 236, pp. 280-28.
| |