326
C.H. Reick . A. Grünewald . B. Page
from a completely uncontrollable environment, taking data more frequently or in
addition at other points of the sea would let the costs explode and, finally, the
details of the interactions between the different plankton organisms and also how
they are infIuenced by environmental conditions is only insufficiently known so
that it is not clear what variables would be relevant.
We discussed that in the context of prediction the concept of generalization has
three aspects: prediction quality, predietion reliability and model correctness.
When working with clean laboratory data a good predietion quality usually is also
an indieation of reliability and model correctness. But when working with data of
low quality, as is usually the case with environmental data, only a poor prediction
quality can be expected so that the reliability and model correctness have to be
considered independently of the prediction quality. Here crossvalidation
techniques can be employed. Unfortunately, this is extremely laborous because
the same Neural Network has to be trained for many different data splittings so
that an automatization would be helpful. We showed that a major problem of
automatization is to stop the training process automatically, such that the networks
get neither under- nor overadapted. As a solution we proposed our early stopping
algorithm and showed by an example how it can be applied in practice.
Actually, our early stopping algorithm is only a first step in the direction of
automatized prediction studies. For large data sets with many variables one would
need Neural Network tools that allow for complex definitions of large sequences
of prediction experiments with networks of different structure, the automatic
execution of these experiments as weIl as their automatie documentation and
analysis. In the next years the various environmental monitoring programs with
their automatic data aquisition will yield an ever increasing fIood of data. By
employing times series predietion methods, like those based on Neural Networks,
these data could in principle be used for environmental management, e.g. to detect
and anticipate significant ecosystem changes. But this will only be possible if
appropriate Neural Network tools are available. And these have still to be
developed.
Acknowledgements
We gratefully acknowledge the provision of the zooplankton data by W. Greve,
Forschungsinstitut Senckenberg, Hamburg. This work was partially financed by
the Bundesministerium für Bildung und Wissenschaft (BMBF) under the grant
03F0181B.
References
Armstrong JS (Ed.) (2000) Principles of Forecasting. Kluwer Academic, Boston
C.H. Reick . A. Grünewald . B. Page
from a completely uncontrollable environment, taking data more frequently or in
addition at other points of the sea would let the costs explode and, finally, the
details of the interactions between the different plankton organisms and also how
they are infIuenced by environmental conditions is only insufficiently known so
that it is not clear what variables would be relevant.
We discussed that in the context of prediction the concept of generalization has
three aspects: prediction quality, predietion reliability and model correctness.
When working with clean laboratory data a good predietion quality usually is also
an indieation of reliability and model correctness. But when working with data of
low quality, as is usually the case with environmental data, only a poor prediction
quality can be expected so that the reliability and model correctness have to be
considered independently of the prediction quality. Here crossvalidation
techniques can be employed. Unfortunately, this is extremely laborous because
the same Neural Network has to be trained for many different data splittings so
that an automatization would be helpful. We showed that a major problem of
automatization is to stop the training process automatically, such that the networks
get neither under- nor overadapted. As a solution we proposed our early stopping
algorithm and showed by an example how it can be applied in practice.
Actually, our early stopping algorithm is only a first step in the direction of
automatized prediction studies. For large data sets with many variables one would
need Neural Network tools that allow for complex definitions of large sequences
of prediction experiments with networks of different structure, the automatic
execution of these experiments as weIl as their automatie documentation and
analysis. In the next years the various environmental monitoring programs with
their automatic data aquisition will yield an ever increasing fIood of data. By
employing times series predietion methods, like those based on Neural Networks,
these data could in principle be used for environmental management, e.g. to detect
and anticipate significant ecosystem changes. But this will only be possible if
appropriate Neural Network tools are available. And these have still to be
developed.
Acknowledgements
We gratefully acknowledge the provision of the zooplankton data by W. Greve,
Forschungsinstitut Senckenberg, Hamburg. This work was partially financed by
the Bundesministerium für Bildung und Wissenschaft (BMBF) under the grant
03F0181B.
References
Armstrong JS (Ed.) (2000) Principles of Forecasting. Kluwer Academic, Boston