Fuzzy Approaches To Environmental Decisions Application

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

environmental science & policy 9 (2006) 22–31

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/envsci

"
Fuzzy approaches to environmental decisions: application
to air quality

Bernard E.A. Fisher


Risk and Forecasting, Environment Agency, Kings Meadow Road, Reading RG1 8DG, UK

article info abstract

Published on line 27 December 2005 This paper considers flexible approaches to decisions designed to improve environmental
quality having regard to uncertainty. The performance of simple and complex models, for
Keywords: forecasting air quality are reviewed, and both types are shown to involve considerable
Fuzzy logic uncertainty regarded as typical of environmental systems. This means that decisions
Environmental decision-making usually depend on combining two or more quite uncertain environmental criteria, and it
Aggregation is shown that this can be approached systematically if a fuzzy logic framework is adopted.
Multi-criteria analysis Fuzzy set aggregation includes, as special cases, other decision-making frameworks, such as
Air quality models multi-criteria analysis and conventional probability based methods. Examples are pre-
Risk sented of how it can be applied to situations involving models and used to incorporate
broader factors involving risk, and socio-economic considerations.
# 2005 Elsevier Ltd. All rights reserved.

1. The problem and to consider models which allow decisions to be made,


so-called ‘fit for purpose’ models. These models may involve
Ideally environmental models should contain the best- greater uncertainty than more complex models, but they
known science, be tested against measurements, and then facilitate more readily the treatment of uncertainty, through
used for prediction and decision-making if they are sup- sensitivity analysis. Moreover, the data requirements to
ported by the necessary input data and perform well against run them are less stringent. They may also allow broader
measurements. This has led in recent years to the develop- factors to be included into the decision-making process,
ment of more complex, fundamental models, in which every incorporating risk, and optimising social and economic
part of the environmental system is described in as much factors, etc.
detail as possible, the so-called ‘reductionist’ approach. If An environmental decision ought to depend on a criterion
the development in modelling produced better predictions meeting a numerical objective with uncertainty treated
this would be the way to proceed. From examples in the explicitly (Royal Commission on Environmental Pollution,
field of air pollution, it is argued in this paper that limita- 1998). One way of including the uncertainty is to assume that
tions in process description, and the lack of detailed data on the criterion involves the membership of a fuzzy set. This is
concentrations, deposition or emissions etc., mean that this an obvious application of fuzzy logic to environmental
approach has not produced usefully better predictions. Even decision-making. In some cases, the environmental criterion
the most advanced models are still associated with large is vague or imprecise, such as when dealing with the ‘quality
uncertainties (Hunt, 2000; Funtowicz and Ravetz, 2005; of life’ (Mendes and Motizuki, 2001), the ranking of ecosys-
Saloranta, 2001). tems in terms of environmental conditions and impacts (Tran
The alternative approach to forecasting is to consider the et al., 2002), or environmental impact assessment (Enea and
decision that is likely to bring environmental improvements Salemi, 2001).

E-mail address: Bernard.Fisher@environment-agency.gov.uk.


1462-9011/$ – see front matter # 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.envsci.2005.08.006
environmental science & policy 9 (2006) 22–31 23

A fuzzy set is a generalisation of a normal set for which In this paper, we consider how conventional approaches
there is not a sharp boundary between belonging and not to air quality models can be extended within a fuzzy logic
belonging to the set. A fuzzy set is defined by its membership framework to include wider considerations, not just those
function and two examples are shown in Fig. 1, in which two concerned with the accuracy of the model. These wider,
hypothetical environmental quality criteria (relating to health generally policy related, considerations are relevant to how
and noise) along a cross-section though a city are shown, the modelling results will be used, and hence should also be
where x is the distance from the city centre. In both cases near an integral part of the decision.
the city centre (at small values of x), the environmental quality
is poorer, and therefore there is a greater possibility that
environmental quality objectives are exceeded. This is shown 2. Performance of air quality models: simple
by the higher membership functions at short distances from and complex
the city centre for the fuzzy sets describing unsatisfactory
health and noise. If the environmental quality were described 2.1. Short-range models
by a normal set, the membership function would equal 1 out to
the distance at which the environmental quality was judged A number of assessments of dispersion models’ performance
satisfactory and the membership function would be set equal to predict concentrations over short distances out to 30 km
to 0 at greater distances. have been made. Some of these have been summarised by
Qualitative factors, involving expert judgement of the Ireland (2003), and for urban areas by Fisher (2005a). They
pedigree of a model, should be part of an assessment of model show that various summary statistics can be used to judge
performance, going beyond traditional scientific approaches, model performance. Model performance also varies according
such as sensitivity analysis and model validation (Van der to the way results are interpreted. For example, for point
Sluijs et al., 2003). Wider judgements are required when sources much of the error can arise from errors in the
reviewing the formulation of a model or deciding whether it is prediction of the direction of the plume centreline. The error is
appropriate to use an established model in a novel applica- much smaller if predicted and measured maximum ground-
tion. Because of its history of testing model performance, and level concentrations are compared without regard to the
the use of objectives in air quality management, air quality location of the maxima. Generally, studies suggest that most
provides a good test case of fuzzy decision-making. However, predictions are likely to be within a factor of two of the
fuzzy decision-making is closely related to other environ- predictions, a simple, useful measure of performance, but
mental decision-making frameworks, such as multi-criteria performance is not consistently better than a factor of two.
decision analysis for flood defence (Tung, 2002), or the The uncertainty in model predictions could be based on other
weighted utility approach for abrupt climate change (Per- estimators, such as the residual sum of squares. The factor of
rings, 2003). It is not proposed that fuzzy decision-making is two is easy to explain to decision makers and would be in
superior to these other methods, but it does provide a agreement with the uncertainty associated with setting the
structured framework within which other methods may boundary of air quality management areas, recommended in
be incorporated. informal guidance produced by National Society for Clean Air
(2003).
The uncertainty associated with errors in model formula-
tion, or whether the model is appropriate to an application are
more difficult to deal with. In part, this may be addressed by
using more than one air quality dispersion model in predic-
tions (Fisher, 2003; Fisher et al., 2002). The overall uncertainty
can only be addressed by recourse to comparisons with
measurements of suitable quality and quantity. The pedigree
of the model can help (Van der Sluijs et al., 2003) and would
bring in considerations as to whether the model had been used
for similar applications and if the formulation of the model is
readily available for review by potential users. When the
structure of the environmental model is very complex, which
is often the case in regional air quality models, sensitivity
analysis to choose the best values for parameters, is not
usually undertaken, because either the run times are too long,
or the data sets of observations are too limited to evaluate each
component in detail.

Fig. 1 – Two environmental criteria A and B equal to the 2.2. Regional models
membership functions of fuzzy sets defined in terms
of the distance from the centre of a city, x. The higher The prediction of concentrations and deposition over long and
membership functions mA and mB near the city centre short distances using air quality models depends in part on the
x = 0 indicate the greater possibility that the uncertainty in the input values for the parameters in the
environmental quality is poor in this region. model. This can be assessed in a relatively straightforward
24 environmental science & policy 9 (2006) 22–31

way, if possible errors in model formulation are ignored. models are possible, because they have short running times.
Two recent studies have determined the uncertainty in Abbott et al. (2003) found that 300 runs of the model TRACK
model predictions for acid deposition (Abbott et al., 2003) using a plausible range of input parameter values, describing
and dispersion calculations (Hall et al., 2000a,b) based on emissions, processes and meteorology, encompassed almost
sensitivity studies, Monte-Carlo simulation and scena rio all the annual wet deposition measurements made at UK sites
analysis. The degree of uncertainty is again broadly within in 1997.
a factor of two i.e. changing input values within their Comparisons of model performance against measure-
bounds of uncertainty generally leads to predictions which ments were possible for both simple and complex models.
lie within a range of two of a central value, but predictions In the above study (Abbott et al., 2003), the baseline model
better than a factor of two cannot be obtained. Abbott et al. TRACK for which normal parameter input values are used, had
(2003) have also shown that the commonly applied methods correlation coefficients1 of 0.6, 0.6 and 0.4, when measured
of deriving critical loads have uncertainties of comparable and modelled annual wet deposition of sulphur, nitrate and
magnitude. reduced nitrogen were compared at 31 UK acid deposition
An example of a complex integrated atmospheric model to sites. The annual wet deposition of sulphur, nitrate and
calculate long-term annual average atmospheric concentra- reduced nitrogen for 1997 from the statistical national
tions and depositions over a region such as Europe, is the use assessment model HARM, which has in recent years been
of Models-3. Recent research (Cocks et al., 2004) has demon- used to make assessments of possible critical load excee-
strated that meso-scale pollution transport models, such as dences from a variety of sources, had slightly better correla-
Models-3, can be run to produce hourly time-series of tion coefficients of 0.7, 0.7 and 0.6. The national assessment
pollution concentration and deposition fields over long model FRAME, though not the latest version, produced a
periods, such as a year, providing direct comparisons with correlation coefficient of 0.3 for reduced nitrogen comparisons
percentile exceedences of short-term air quality objectives. for 1997. A comparison of the annual wet deposition at 48
Long-term averages derived from time-series also provide a European sites from the EMEP model yielded correlation
benchmark for comparison with older established statistical coefficients of 0.6, 0.7 and 0.7 over the period 1985–1996. The
acid deposition models, which may only be used to calculate correlation coefficients from the limited spatial analysis of the
long-term averages. Issues regarding the practical use of Models-3 annual deposition over the UK (Cocks et al., 2004)
Models-3 for regulatory assessment purposes remain, but the gave values of 0.75 and 0.5 for comparisons of the calculated
demonstration supports the viability of the concept of a and measured annual wet deposition of sulphur and nitrate at
modular assessment framework incorporating complex inter- 13 UK acid deposition sites for 1999. No correlation was found
actions. As an integrated, ‘reductionist’ model, the results for reduced nitrogen, though predictions were in the right
provide extra information on quantities, such as secondary range. One should be cautious about drawing firm conclusions
particulate matter. from these numbers, which apply to annual deposition over
The running of a meso-scale model is a very large different years, periods and spatial domains, but it is apparent
computing exercise, and a large amount of preparatory work that none of the models was markedly better at predicting
is involved in the production of the data sets used. Unlike the annual spatial patterns than any of the others.
simpler models, the attribution, or contribution, of a single However, complex models (Models-3, or EMEP) do have the
source or a set of sources to the total concentration or advantage that they can be used to predict time series of
deposition, is not a simple output of such models. The concentrations or deposition, which is not possible with the
calculations provide the concentration, as time series, at simpler models. Models-3 (Cocks et al., 2004) produced
chosen receptor points. However, as the model includes non- correlation coefficients for weekly time-series of observations
linear processes, the simple addition or subtraction of in the range 0.6–0.8 for sulphur, 0.4–0.8 for nitrate and 0.3–0.8
individual source contributions is not appropriate. In order for reduced nitrogen at UK acid deposition sites. A similar
to estimate the contribution of a set of sources, the comparison of the results of the EMEP model, at monthly
calculation would need to be repeated again, so that averaged, European acid deposition sites, appeared to suggest
computer results with, and without, the specified sources slightly lower correlation coefficients.
can be compared. The run times can be reduced by running Without dismissing the progress that ‘reductionist’
parts of the model in parallel, but run times and costs are still models may achieve in advancing the science, one may
a significant disincentive to the routine application of such consider more direct approaches to decision-making. The
models. Source-receptor relationships are available when studies cited above suggest that models of concentration or
supercomputing facilities are used, such as for the EMEP deposition over long, or short-ranges, produce results, which
model (Tarrason et al., 2003) used in international negotia- lie within, but are not much better than, a factor of two of
tions to set limits on transboundary pollution. However, observations. To the author’s knowledge this degree of
estimates of the uncertainty cannot be easily derived uncertainty is the best achievable for forecasts in all
from this model, or Models-3, because of the computing environmental media i.e. land, air and water. Given this
requirements.
1
A higher correlation coefficient does not imply better perfor-
2.3. Examples of regional model performance
mance, but for illustration has been used here to indicate degrees
of agreement. Reduced nitrogen is difficult to deal with, and
Estimates of the likely uncertainties in predicted concentra- performance depends on assumptions regarding the emissions
tions and deposition using the simpler regional transport inventory.
environmental science & policy 9 (2006) 22–31 25

degree of uncertainty in the environmental quantities


predicted by models, decisions based on models should
not be regarded as certain. Instead a fuzzy logic framework
should be applied, which additionally has the advantage
that other, broader aspects of the decision-making process
might be included within the same framework.

3. Uncertain environmental objectives

3.1. Critical load example

Although national air quality objectives, in terms of


concentrations and deposition, have been laid down, Fig. 2 – Illustration of a Monte-Carlo estimate of the
decisions regarding the management of emissions, in the membership function of the aggregated set describing
event of exceedences are not straightforward. The objectives the difference between atmospheric deposition and
themselves should not be regarded as exact boundaries critical load. The deposition and critical loads are assumed
between satisfactory and unsatisfactory air quality. Air to have simple membership functions characteristic of
quality assessments should rely on the best and most estimates lying within a ‘factor of 2’. The membership
equitable ways of ensuring improvement, taking account of function of the fuzzy set describing deposition minus
the uncertainty in current methods. Decisions made on the the critical load is shown by the straight-line graph. The
assumption that exceedences can be assessed without error membership function of the set describing exceedence
can be very sensitive to underlying assumptions, whereas assuming the deposition is greater than the critical load
methods accepting that the transition from acceptable air is the curved line. One can sees that even when the
quality to unacceptable air quality is gradual, incorporate a deposition exceeds the critical load, one cannot know
more realistic representation of present knowledge (Fisher, with complete certainty whether exceedence occurs. If
2003). uncertainties were neglected there would be a sharp
The application of the critical loads approach illustrates the transition from 0 to 1, exactly at a point where the
problem (Fisher, 2005b). A critical load assessment is based on deposition exceeded the critical load.
the idea that the acid deposition onto a 1 km  1 km grid
square in the UK should not exceed the critical load of that grid
square. The deposition from emissions in the UK compared interpreted as probability distribution or density functions.
with the critical load sets criteria on the emissions of the main However, in some cases one would wish to transform
acidifying species, sulphur oxides, nitrogen oxides and weightings in a way, which meant that they could no longer
reduced nitrogen compounds. Setting exceedences of the be interpreted as probabilities. In a fuzzy set framework, the
critical load as undesirable implies a sharp criterion, in which weighting need not lie between 0 and 1, nor need it vary
deposition less than the critical load is acceptable, and monotonically. For example when the exceedence is very
deposition greater than the critical load is not acceptable. high, the weighting is nearly one that the critical load is
Although strictly the criterion applied to critical loads exceeded. One might regard very high exceedences as outliers,
should be no exceedence anywhere, one should recognise the which should be excluded. In this case, the membership
uncertainty in the deposition and critical load. As an function for poor environmental quality could decrease again
illustration one might consider the difference between the for high values of exceedences. The advantages of using a
deposition and the critical load over a grid square, assuming fuzzy set description are that it is a flexible framework within
that both the deposition and the critical load can be described which other descriptions of environmental damage may be
by fuzzy sets with the expected factor of 2 agreement incorporated. This has advantages when including criteria,
discussed in Section 2. Then with the mean deposition greater which are uncertain, or not well defined.
than the mean critical load, the membership function of the The approach outlined above may be compared with one
fuzzy set describing exceedence is that shown Fig. 2, where x is in which a single ‘best’ prediction is compared with a
the exceedence = deposition  critical load. It has been criterion. If the prediction lies well above (or below) the
assumed that the membership functions of the deposition criterion, there is little possibility of making the wrong
and critical load fuzzy sets are unrelated, so that one can take decision. (In order to make this judgement some knowledge
random selections from each to get the aggregate fuzzy set of uncertainty is implied.) However, if the single prediction is
describing exceedence. It has been assumed for the purposes close to the criterion, one might try to apply a more detailed
of the numerical illustration, that the deposition membership model to reduce the uncertainty. However, the evidence
function ranges uniformly between 0.5 and 1.4 and is zero presented in Section 2 does not suggest that complex models
beyond these limits, while the critical load membership lead to more accurate answers, so that this does not
function ranges uniformly between 0.4 and 0.9 and is zero necessarily help the decision-making. However, complex
outside this range. models can be used to explore non-linearities between the
In the example in Fig. 2, the membership functions input parameters, which is not possible in simple models.
expressing the weighting on values in the sets could be These non-linearities could result in probability distribu-
26 environmental science & policy 9 (2006) 22–31

tions with unusual features, such as significant weighting exceedences in areas where the deposition is very much
on predictions much larger or smaller than estimates near greater than the critical load, since the straight-line member-
the centre of a range. ship function equals 1 in regions where the deposition is
Accepting that the deposition and critical load have some more than twice the critical load. Perrings (2003) remarks that
uncertainty associated with them, within a fuzzy set weighting functions often used in expected utility app-
framework one would define a membership function m(x) roaches to decision-making, have a reference point, and
for the fuzzy set of unacceptable conditions, where the greater weighting is often assigned to changes around the
critical load is exceeded, based on the ratio of the deposition reference point and less to large changes far from the refe-
to the critical load, x. Two examples are shown in Fig. 3: the rence point. In this case, the reference point is x = 1, where the
straight line assumes a linear membership function between deposition equals the critical load and the membership
the conditions when the ratio is 0 (the ideal) and 2 (moderate function is 0.5.
exceedence given the uncertainty in the assessment). For a The main question in applying a fuzzy approach is how to
ratio of deposition to critical load greater than 2, conditions choose the membership function. The two parameters x0 and
are definitely unacceptable. For intermediate values between b allow a range of shapes to be described. x0 fixes the value
0 and 2 one has a sliding scale of acceptable conditions. The where the objective is exceeded and b determines the
strict criterion for acceptability would be a step function, fuzziness; large values of b correspond with less fuzzy,
which would be represented in the figure by a vertical line sharper criteria. In this case, b has been chosen to be
where the deposition/critical load equals one (not shown in consistent with the factor of 2 performance of air quality
the figure). (A step function, set at x = 0.5 say, is an models. The choice of the membership function may also
alternative, precautionary approach.) A smooth membership depend on considerations not arising in the models used to
function has been drawn in Fig. 3 representing a gradual estimate the deposition/critical load ratio. A Pareto member-
transition between acceptable and unacceptable conditions. ship function 1/(1 + x0/x) would have a long tail in which large
The smooth curve is described in the caption to Fig. 3. It is a values of exceedence (large x) have non-trivial weightings
simple membership function with the desired properties that (and are not just equal to 1). Use of the transformation
it equals 0, at x = 0, equals 1, when x is large, and equals 0.5, x ! ebx  1 shortens the tail. In this case, it was considered
when x = x0 = 1. It depends on the parameter b. Large b values desirable to have a membership function with significant
sharpen the fit towards a crisp set. The chosen value of b = 2 variation within a small multiplicative factor of meeting the
means that the membership function equals 0.9, at x = 2, objective x0 = 1.
which corresponds to the expected uncertainty in the air The setting of the membership function requires con-
quality models reviewed in Section 2. Instead of setting sideration of subjective values, which rely on the decision
membership functions based on the accuracy of model maker’s attitude to risk. The membership function in Fig. 3 is
predictions, expert judgement may also be used to define as sensitive as the straight-line membership function to
qualitatively the membership functions of a fuzzy set. improvements in areas where the critical load is slightly
For the straight-line membership function in Fig. 3, exceeded, but is also slightly sensitive in areas where the
greatest improvement would arise from reducing deposition exceedence is grossly exceeded. If the intention is to reduce
in areas within a factor of two of where the deposition equals the membership of the fuzzy sets where gross exceedences
the critical load, x = 1. This is of little advantage in reducing occur, the membership function could then take the form
m(x) = ex. Reducing membership functions of this kind empha-
sises the regions of greatest exceedence.

3.2. Fuzzy logic framework for uncertain environmental


objectives

In the example above, both the environmental effect


(deposition) and the objective (critical load) were determined
using models based on scientific understanding. In many
situations, the decision-making needs to be extended to
include the criteria weighted according to other external
criteria. These may well not be scientifically based and may
depend on judgements involving socio-economic values, or
practicality. In such cases, a fuzzy logic framework is a
Fig. 3 – Examples of membership functions depending natural framework in which to describe an environmental
on the deposition/critical load ratio. The curved line is decision.
the membership function for the fuzzy exceedence set The combination of fuzzy sets through operations, such as
described by the function m(x) = n(x)/(n(x0) + n(x)) where union and intersection, more generally described as fuzzy
n(x) = ebx S 1, and x is the ratio of the deposition/critical aggregation, provides the framework in which criteria can be
load, b = 2 describes the fuzziness, or uncertainty, combined. The requirement to satisfy the axioms of fuzzy set
chosen to be consistent with the performance of aggregation ensures that the combinations are consistent and
environmental models. The criterion for just logical (Fisher, 2003). The need to combine criteria is common
meeting the objective is x0 = 1. to many environmental decisions.
environmental science & policy 9 (2006) 22–31 27

Having defined a membership function, the final step in the Clearly mC can be generalised to consider the aggregation of
decision-making process for the atmospheric deposition more than two sets.
example is to consider arrangements of emissions, which For n ! 1, mC ! maximum(mA, mB) corresponding to the
minimise membership of the fuzzy set of undesirable environ- commonly applied fuzzy set ‘union’ operation (the logical ‘or’
mental conditions. The environmental decision depends on operation) (Klir and Folger, 1988).
minimising a criterion, where this is the membership function For n ! 1, mC ! minimum(mA, mB) corresponding to the
m(x) of a fuzzy set. Unfortunately there is no objective way of commonly applied fuzzy set ‘intersection’ operation (the
choosing the most appropriate membership function and so logical ‘and’ operation).
this must to some extent depend on the decision maker’s For n = 1, mC becomes the arithmetic mean of the member-
attitude to risk. ship functions mA and mB.
For n = 1, mC becomes the harmonic mean of the
membership functions mA and mB.
4. Aggregation of fuzzy membership For n = 0, mC becomes the geometric mean of the member-
functions in multi-criteria situations ship functions mA and mB.
Taking an arithmetic mean with weighting factors is
Some examples are given in this section of how a fuzzy logic equivalent to multi-criteria analysis (Dodgson et al., 2000).
approach provides a framework for aggregating criteria. The Fig. 4 illustrates some of the approaches as vertices of a
aggregate, or combined membership function mC(x) could be triangle. It is included to illustrate different views to risk.
some function of the membership functions of the indivi- Multi-criteria analysis, usually involves some form of trade-off
dual environmental factors under consideration such as between different effects. If mC is determined by the ‘fuzzy
health, noise etc. and their environmental objectives, intersection’ operation, the minimum membership function is
labelled A and B for convenience. The adverse effects of used to define the aggregated set. This is the risk-taker’s
each are described by the membership function of the approach, as it suggests that compliance can be reached for all
individual fuzzy sets mA(x) and mB(x). The decision-making environmental criteria by managing the least harmful. If mC is
process depends on the aggregate membership function determined by the ‘fuzzy union’ operation, the maximum
mC(x). This could be to ensure that mC(x) is less than a critical membership function is used to define the aggregated set. This
value, say 0.5, for all x. The degree of precaution depends on is a precautionary approach as it suggests that compliance can
the choice of this critical value; lower critical values mean be reached for all environmental criteria. Other forms of
more precaution. Where advice is being sought on real world aggregation are also possible.
situations, either a policy goal or some structured consulta- Although aggregation is a way of combining fuzzy criteria,
tion of experts could be conducted to decide on the degree the decision-maker still has to decide on the most appropriate
of precaution.

4.1. Combining or aggregating fuzzy sets

The question of aggregation arises whenever individual


environmental, or social indices are grouped together to
produce a single index. For example, a single multiple index of
deprivation has been produced for England, based on
individual indices representing deprivation according to
income, employment, eduction, housing, etc. (ODPM, 2002).
A simple, weighted averaging procedure was used (see below)
with an exponential transformation to limit the effects of
cancellation. This transformation means that a very deprived
ward is still counted as deprived even if it is not very deprived
according to some of the individual indices.
Combining two fuzzy sets, such as A and B illustrated
in Fig. 1, can be performed using fuzzy set aggregation Fig. 4 – An illustration of how different choices of
in various ways, consistent with the axioms of fuzzy sets aggregation may be made. They depend on the view
(Klir and Folger, 1988). The most direct form is the general- of risk taken by the decision-maker. The weighted
ised, weighted average of the membership functions mA arithmetic mean represents some trade-off between
and mB. mC is the membership function of the aggre- criteria. The intersection and union operations represent
gated set. Weighting functions w and 1  w, with values different views to risk taking. ‘Intersection’ is the risk
between 0 and 1, can be introduced at the discretion of the taker’s approach assuming the minimum membership
decision-maker function of the criteria can be reached by all the criteria.
However, as seen in Section 4.2, it can also be used to
mC ðw; mA ; mB Þ ¼ ðwmnA þ ð1  wÞmnB Þ1=n (1) identify the few worst areas on which to focus action.
‘Union’ is a precautionary approach assuming that
mC takes certain familiar forms depending on the choice of attention should always be paid to managing the
n. One should choose values to fit asymptotic behaviour. worst of the criteria.
28 environmental science & policy 9 (2006) 22–31

way of aggregating the criteria and this depends on his


attitude to risk. The consequences can best be illustrated by an
example. One could extend the uncertainty assessment of one
of the modelling examples discussed in Section 2, to include
the additional uncertainty associated with the expert judge-
ment of the pedigree or appropriateness of the model (Van der
Sluijs et al., 2003). However, since this would involve personal
judgement of a particular model, it is better left to the
structured form of evaluation by a group of experts favoured
by the above authors. Instead an air pollution example is
presented in the next section, in which the adverse effects of
air pollution are combined with a policy objective involving
road travel.

4.2. Example of the application of fuzzy set aggregation to


air pollution and private car usage

An example of aggregation involving a scientific assessment,


air pollution exposure, and a more subjective policy
requirement, private car usage, is presented, but the
numbers and interpretation are used to illustrate the
method and should not be taken in any way as definitive.
By considering the population in each 1 km  1 km grid
square covering the UK, the relative exposure mA of the
whole population to primary PM10 particles may be calcu-
lated for a source at any location, relative to a unit source
placed in central London (Fisher, 2004). This shows where it

Fig. 6 – A criterion showing shaded areas where the relative


private car usage is highest.

would be most advantageous to reduce emissions in the


country to protect the population from air pollution (see
Fig. 5). It suggests that attention should be focused on the
areas with the highest weighting, such as areas with mA
greater than 0.4, which lie in a swathe across the country
coinciding with the main urban areas, and especially on a
few areas inside the 0.5 shaded contour, containing areas
around London.
A decision could be made to try to reduce emissions
focusing on the limited areas within the 0.5 contours, but this
may not be the best decision taking wider considerations into
account, namely that particle emissions are related to road
transport, and the distribution of road transport around the
country has a spatially varying pattern. The feasibility of
encouraging a reduction in road transport, either by encoura-
ging public transport, car sharing etc., is related to current car
usage and is likely to depend on regional factors (Department
for Transport, 2002, 2003). One can consider the relative
distances travelled per person per year along minor roads, in
each 1 km  1 km grid square (assuming travel along major
roads are subject to other types of behavioural influences).
These can be related to the membership function mB of the
Fig. 5 – A criterion showing relative exposure to PM10 fuzzy set describing the personal use of cars as ‘high’. This is
according to location. The axes show distances in shown in Fig. 6 by drawing contours showing where the ‘car
kilometres. The areas inside the 0.5-shaded contour use’ membership function is higher than 0.6. The figure shows
are areas where reductions in emissions of primary higher distances are travelled in rural areas and is quite
PM10 would have the most benefit. dissimilar from Fig. 5.
environmental science & policy 9 (2006) 22–31 29

Fig. 7 – An aggregated criterion based on taking the average Fig. 8 – An aggregated criterion based on the union of the
of the criteria in Figs. 5 and 6 by applying w = 0.5, n = 1, in criteria relating to relative exposure to PM10 and the usage
Eq. (1). This identifies the best areas for action assuming of private cars, obtained by applying w = 0.5, n = 10, in
evidence, or policy arguments, support equal benefit in Eq. (1). This criterion shows most of the country subject to
reducing the relative exposure to PM10 and the relative one or other, or both, of the individual criteria and is not
car usage. useful for discriminating where action would be most
advantageous.

The combined outcome should be to reduce areas of both ship function of the aggregate weighting equals 0.6.
high air pollution and high private car use, though these do not Because areas where sources would cause high exposure
occur in the same areas. One may consider the aggregate do not generally coincide with areas of high personal car
weighting of both factors, assuming mA is the membership usage, the areas with a high aggregate weighting lie
function of each grid square, with poor air quality, and mB is somewhere between Figs. 5 and 6. Fig. 8 (w = 0.5, n = 10)
the membership function of each grid square, with a high, shows areas where either of the criteria may be high. This
annual distance travelled by car. mA is chosen to have the covers most of the country and hence is not a useful
functional form and the parameter values shown in Fig. 3, criterion. Fig. 9 (w = 0.5, n = 10) shows where both criteria
where the x-axis is the relative air pollution exposure. This are satisfied. This is seen to be rather different from either
form has been chosen to allow for a plausible degree of of the initial criteria, Figs. 5 or 6. It focuses attention on
uncertainty. For convenience the membership function mB is just a few areas, which are highly weighted according to
chosen to have the same functional form and uncertainty as both criteria, and these are areas where policy should be
mA, where the x-axis is the relative annual distance travelled directed.
by person by car in a grid square. This involves a subjective If the aim is to identify a few areas for action, the fuzzy
judgement of what might be regarded as ‘high’ annual travel. intersection aggregation appears to satisfy the requirement.
It is an example of the use of expert or policy judgement to The example only gives a notional idea of how methods of
define a concept in a qualitative way. Since attention is fuzzy set aggregation may be used to combine the results of an
focused on change from the present situation, it seems environmental model with a broader, policy criterion. More
appropriate to choose the membership function in Fig. 3, and involved fuzzy logic rules could be developed to describe
to make it sensitive to changes in travel distance close to the responses to more complex criteria. For example, one might
current national mean travel distance. consider taking into account differences between urban and
Fig. 7 shows the effect of aggregating the two criteria rural areas, since there are fewer alternatives to the private car
using simple averaging (w = 0.5, n = 1), when the member- in rural areas.
30 environmental science & policy 9 (2006) 22–31

Acknowledgements

The valuable contribution of Dr R. Willows to the preparation


of the paper is appreciated. The views expressed are those of
the author and not necessarily those of the organisation to
which he belongs.

references

Abbott, J., Hayman, G., Vincent, K., Metcalfe, S., Dore, T.,
Skeffington, R., Whitehead, P., Whyatt, D., Passant, N.,
Woodfield, M., 2003. Uncertainty in acid deposition
modelling and critical load assessments. Environment
Agency R&D Technical Report P4-083(5)/1.
Cocks, A., Lucas, V., Rodgers, I., Teasdale, I., 2004. The
performance of Models-3 for deposition and atmospheric
concentrations over a year. Environment Agency R&D
Technical Report P4-083(5).
Department for Transport, 2002. Revised national travel survey
data for urban and rural areas. DfT December 2002,
available at http://www.transtat.dft.gov.uk/personal/
index.htm.
Department for Transport, 2003. Travel in urban and rural areas
in Great Britain, personal travel fact sheet 11. National
Statistics, Department for Transport.
Dodgson, J., Spackman, M., Pearman, A., Phillips, L., 2000. Multi-
criteria analysis: a manual. London, Department of the
Environment, Transport and the Regions. http://
www.odpm.gov.uk/stellent/groups/odpm_about/
Fig. 9 – An aggregated criteria based on taking the documents/pdf/odpm_about_pdf_608524.pdf or a hard copy
intersection of the criteria based on the relative exposure from DETR Publications Sales Centre, Unit 21, Goldthorpe
to PM10 and the use of the motor car, obtained by applying Industrial Estate, Goldthorpe, Rotherham S63 9BL.
Enea, M., Salemi, G., 2001. Fuzzy approach to the environmental
w = 0.5, n = S10, in Eq. (1). These are areas with a high
impact evaluation. Ecol. Model. 135, 131–147.
weighting according to both criteria. Fisher, B.E.A., Ireland, M.P., Boyland, D.T., Critten, S.P., 2002.
Why use one model? An approach for encompassing model
uncertainty and improving best practice. Environ. Model.
Assess. 7, 291–299.
Fisher, B.E.A., 2003. Fuzzy environmental decision-making:
5. Conclusions applications to air pollution. Atmos. Environ. 37, 1865–1877.
Fisher, B.E.A., 2004. Risk based approaches to assessing the
It is seen that conventional approaches to uncertainty, such as environmental burden of acid gas emissions. In:
sensitivity analysis, indicate that large error bounds are Proceedings of the 27th NATO/CCMS International
associated with environmental models. Rather than expend- Technical Meeting on Air Pollution Modelling and its
ing great efforts to reduce these uncertainties, it is preferable Application, 25–29 October 2004, Banff Centre, Banff,
Canada.
that environmental decision-making should be put into a
Fisher, B.E.A., 2005a. Sensitivity of urban dispersion calculations
fuzzy logic framework, which can accommodate uncertainties to urban meteorology (COST 715). Int. J. Environ. Pollut. 25,
in the model, the environmental objectives and the broader 71–79.
policy aims. Aspects of the framework are dependent on the Fisher, B.E.A., 2005b. Environmental burden of acid gas
decision-maker and wider environmental factors, but this emissions associated with the exceedence of critical loads.
may be considered an advantage as it is possible within the Atmos. Environ. 39, 1903–1915.
Funtowicz, S., Ravetz, J., 2005. Post-normal science—
framework to take a consistent approach to combining
environmental policy under conditions of complexity.
different environmental criteria. The final decision will
http://www.nusap.net.
depend on a combined environmental criterion, determined Hall, D.J., Spanton, A.M., Dunkerley, F., Bennett, M., Griffiths,
after suitable trials of the membership function of an R.F., 2000a. A review of dispersion model inter-comparisons
aggregated fuzzy set. studies using ISC, R91, AERMOD and ADMS. Environment
This does not exclude applying environmental models Agency Technical Report P353.
in a conventional statistical framework. However, one is Hall, D.J., Spanton, A.M., Dunkerley, F., Bennett, M., Griffiths,
R.F., 2000b. An inter-comparison of the AERMOD, ADMS
unlikely to be able to set results wholly in such a frame-
and ISC dispersion models for regulatory applications.
work because of practical considerations regarding any Environment Agency Technical Report P362.
actions for improvement, apart from limits on the Hunt, J., 2000. Mathematics and environmental problems—
amount of measured data available for testing model methodologies and future developments. NERC-EPSRC
performance. Workshop, July 18, 2000.
environmental science & policy 9 (2006) 22–31 31

Ireland, M, 2003. Local Air Quality Management and Climate Tran, L.T., Knight, C.G., O’Neill, R.V., Smith, E.R., Riitters, K.H.,
Change: Tools for Joined-up Policy. PhD Thesis, University Wickham, J., 2002. Fuzzy decision analysis for integrated
of Greenwich. environmental vulnerability assessment of the Mid-Atlantic
Klir, G.J., Folger, T.A., 1988. Fuzzy Sets, Uncertainty and region. Environ. Manage. 29, 845–859.
Information. Prentice Hall, New Jersey. Tung, Y.-K., 2002. In: Wu, et al. (Eds.), Risk-Based Design of
Mendes, J.F.G., Motizuki, W.S., 2001. Urban quality of life Flood Defense Systems. Flood Defence. Science Press,
evaluation scenarios: the case of Sao Carlos in Brazil. New York.
CTBUH Rev. 1, 1–11. Van der Sluijs, J., Kloprogge, P., Risbey, J.,Ravetz, J., 2003.
National Society for Clean Air, 2003. Air quality management Towards a synthesis of qualitative and quantitative
areas: a review of procedures and practice for local uncertainty assessment: applications of the numerical,
authorities. NSCA publication. unit, spread, assessment, pedigree (NUSAP) system. Paper
ODPM, 2002. Updating the English indices of deprivation at International Workshop on Uncertainty, Sensitivity and
2000. Stage 1 consultation report. Office of the Parameter Estimation for Multimedia Environmental
Deputy Prime Minister, ODPM Neighbourhood Modeling, 19–21 August 2003. Rockville, Maryland USA.
Renewal Unit. See also www.nusap.net.
Perrings, C., 2003. The economics of abrupt climate change.
Phil. Trans. R. Soc. Lond. A 361, 2043–2059. Bernard Fisher is a policy advisor on Risk and Forecasting at the
Royal Commission on Environmental Pollution, 1998. Setting Environment Agency. His interests lie in air pollution modelling,
environmental standards. 21st Report. The Stationary air quality management, acid rain and environmental inequal-
Office, Cm 4053. ities. Recent experience has involved implementing local air qual-
Saloranta, T.M., 2001. Post-normal science and the climate ity management when he was a research professor at the
change issue. Climatic Change 50, 395–404. University of Greenwich. He is a member of the Air Quality Expert
Tarrason, L., Jonson, J.E., Fagerli, H., Benecdictow, A., Wind, P., Group of the UK government Department of Environment, Food
Simpson, D., Klein, H., 2003. Transboundary acidification, and Rural Affairs, and was until recently chairman of the COST 715
eutrophication and ground level ozone in Europe. committee on ‘‘Meteorology applied to urban air pollution pro-
Part III. Source-receptor relationships. EMEP blems’’. He is a visiting professor at the Universities of Greenwich
Report 1/2003. and Hertfordshire.

You might also like