Agriculture Reference
In-Depth Information
no turkey model seems to use independent
com mercial software.
Specialized modelling software has
the advantage of making model develop-
ment fast while allowing both users and
scientists to generate the graphs and tables
they need to make informed choices. These
tools are also very flexible for the end user
who can add new outputs or variables to
the model and run sensitivity analyses to
compare different scenarios. However, the
major limitation to using such tools is that
users must be familiar with modelling and
must be trained in modelling principles and
software utilization (e.g. sensitivity analyses,
optimization) to make sure that the user
masters the model and understands its
limitations.
Depending on the user's needs and
skills, developing a friendly interface may be
more appropriate especially in today's world
where time is a limited resource. However,
these software applications do not allow
users to add new outputs to the model. Most
of the time, an IT team is needed to develop
the interface. Such an interface should be
preferred when users are non-modellers and
do not necessarily need to understand the
science of the model. However, it is still re-
commended that the model is first devel-
oped and tested on modelling software to
validate its application before allocating time
and human resources to program the model
as an independent piece of software.
Precision and accuracy
Precision 'measures how closely individual
model-predicted values are within each other',
while accuracy 'measures how closely model-
predicted values are to the true values'
(Tedeschi, 2006). The linear regression ana-
lysis of observed values (i.e.
Y
) with simu-
lated values (i.e.
X
) is a means of evaluating
both model precision and accuracy. Inter-
cept and slope estimates that do not signifi-
cantly differ from zero and unity, respect-
ively, indicate an accurate response or that
the predicted values closely fit the actual
responses. Log-transformations of the ob-
served and simulated values may be appro-
priate to avoid the effect of the linear in-
crease in performance (e.g. feed intake, live
weight) when estimating linear regression
parameters (
Fig. 8.2)
.
Accuracy can be determined as the abso-
lute or relative difference between simulated
and observed values, where the relative dif-
ference corresponds to the ratio of the abso-
lute difference to the observed value. Both
absolute and relative differences should be
calculated for the different stages of growth
to shed light on which periods are less accur-
ate than others. A challenge with turkey
growth models is the maximization of model
accuracy early in the growing period to limit
increasing cumulative inaccuracy at later
ages. Indeed, a 2% accuracy at early ages
(i.e.
10
g precision at 500 g live weight) may
be fairly satisfying but unacceptable in the
finisher phase (i.e. 400 g precision at 20 kg
live weight). Finally, the evaluation of model
accuracy at different stages of growth can help
clarify model behaviour during the overall
growing cycle. For example, with a consistent
overestimation of weight gain it may be eas-
ier to identify the cause and effect response,
whereas both over- and underestimations
make it difficult to identify the cause.
Evaluation
Model evaluation aims at determining
model precision, accuracy, robustness and
flexibility by comparing simulated outputs
with observed values over a wide range of
production scenarios. An evaluation of the
simulations should first be performed with
the data used during model calibration to
validate the consistency of model outputs.
Once the internal validation is completed,
the use of external data (i.e. data that were
not used during model development) per-
mits the assessment of model accuracy,
limitations and, therefore, its applicability.
Robustness and flexibility
Robustness refers to the ability of the model to
adapt to perturbations (Sauvant and Martin,
2010). A robust model should therefore not