Data-Based Mechanistic Modelling and the Emulation of Large Environmental System Models - Environmental Modelling: Finding Simplicity in Complexity - page 108

approach to modelling in the hydrological literature (e.g.
Jothityangkoon et al ., 2001; Sivapalan and Young, 2005
and the references therein), for instance, is a response
to the relative failure of the alternative reductionist
('bottom-up') philosophy in this area of study. But such
scepticism is not new. It has its parallels in the envi-
ronmental (e.g. Young, 1978, 1983; Beck, 1983) and
ecosystems (e.g. see prior references cited in Silvert, 1993)
literature of the 1970s and early 1980s, where the present
first author's contributions were set within the context
of 'badly defined' environmental systems. To quote from
the first author (Young, 1983), which echoes earlier ideas
(Young, 1978), for instance:

probably be concerned that the chosen procedures could
so easily be misused: whereas the constrained parame-
ter optimization represents a quantitative and relatively
objective approach, it is submerged rather arbitrarily
within a more qualitative and subjective framework based
on a mixture of academic judgment and intuition. Such
a statistician would enquire, therefore, whether it is not
possible to modify this framework so that the analyst
cannot, unwittingly, put too much confidence in a priori
perceptions of the system and so generate overconfidence
in the resulting model.

This and the other early papers then went on to present
initial thoughts on such an objective, statistical approach
to modelling poorly defined systems that tried to avoid the
dangers of placing too much confidence in prior percep-
tions about the nature of the model. They also adumbrate
antireductionist arguments that are very similar to argu-
ments that have appeared recently in the hydrological
literature and express some of these same views within
a hydrological context (Jakeman and Hornberger, 1993;
Beven, 2000, 2001). Quite similar antireductionist views
are also appearing in other areas of science; for instance, in
a lecture presented at the University of Lancaster (Lawton,
2001), the then chief executive of the UK Natural Environ-
ment Research Council (NERC), recounted the virtues of
the top-down approach to modelling ecological systems
(although, for some reason, he did not appear to accept
that such reasoning could also be applied to other natural
systems, such as the physical environment).
In the subsequent period since the earlier papers were
published, however, the author and several colleagues
have sought to develop this statistical approach within
a more rigorous systems setting, which the author has
termed data-based mechanistic (DBM) modelling (e.g.
Young and Ratto, 2008, and the references therein). Before
discussing the DBM approach, the present chapter will
first outline the major concepts of statistical modelling
that are important in any modelling process. Subse-
quently, a typical practical example is presented that
illustrates the utility of DBM modelling in producing a
parametrically efficient (parsimonious) stochastic model
of river catchment dynamics from rainfall-flow data. This
chapter also discusses how this same methodology can be
useful not only for the modelling of environmental and
other systems directly from time series data, but also as
an approach to dynamic model reduction and the 'emu-
lation' of large computer-simulation models of dynamic
systems (see also Chapter 26).

Although such reductionist analysis is perfectly
respectable, it must be used very carefully; the dangers
inherent in its application are manifold, but they are not,
unfortunately, always acknowledged by its proponents.
It is well known that a large and complex simulation
model, of the kind that abounds in current ecological
and environmental systems analysis, has enormous
explanatory potential and can usually be fitted easily to
the meagre time-series data often used as the basis for
such analysis. Yet even deterministic sensitivity analysis
will reveal the limitation of the resulting model: many of
the 'estimated' parameters are found to be ill-defined
and only a comparatively small subset is important in
explaining the observed system behaviour.

The same paper goes on to point out that such
over-parameterization and the associated identifiability
problems are quite often acknowledged, often implic-
itly, by the reductionist simulation model builder. For
example, the modeller sometimes constrains the values
of certain 'better known' parameters and seeks to fit the
model by optimizing the chosen cost function in relation
to the remaining parameters, which are normally few in
number. However, the model then has a degree of 'surplus
content' not estimated from the available data, but based
on a somewhat ad hoc evaluation of all available prior
knowledge of the system and coloured by the analyst's
preconceived notions of its behavioural mechanisms. The
paper concludes that:

On the surface, this conventional simulation modelling
approach seems quite sensible: for example, the statis-
tician with a Bayesian turn of mind might welcome its
tendency to make use of all a priori information available
about the system in order to derive the a posteriori model
structure and parameters. On the other hand, he would