292
J. Bobbin . F. Recknagel
15.1.1
Knowledge Generalization and Representation
Inductive models learn by generalizing the knowledge they acquire about a certain
domain. The model that is induced from available data leams to associate domain
properties that have equivalent consequences but appear differently. The
complexity of the raw data needs to be aggregated by the model.
There are many ways that knowledge can be represented, including as neural
networks, rule sets, fuzzy sets, learnt equations and regression trees. Most
previous applications of machine leaming in ecology have concentrated on
acquiring quantitatively accurate models of the systems as determined by
independent test data. However, a predictive model should achieve not only
accurate classification but also provide insight and understanding of the predietive
structure of the data (Breiman et al. 1984). Evaluating models solelyon their
predictive performance ignores any information the model can provide about the
relationships it has learnt and how they relate to the causal understanding of
ecosystem behaviour. For example, many ecological applications of neural
networks primarily focus on minimising the root mean square error (RMSE) of
ecosystem predictions based on implicit representations of the relationships by
weights of neurons rather than to explain relationships by which the model decides
its predictions. This chapter considers how ecosystem data can be used to induce
models, whieh are both predictive and explanatory.
Representing knowledge in a transparent manner is of great importance to
inductive modeling. Inductively gained models are a hypothesis derived from the
data, and certainly not a proven law for underlying mechanisms. A model derived
from measured data can only generalize to the extent that the data provided
accurately represent the universe of possible measurements. The representation of
knowledge generalised during this process is critical in terms of understanding of
predietions and knowledge discovery. Improved understanding of inductive
ecosystem models will increase our confidence in the model outputs and gradually
turn them into "grey boxes" .
In this chapter we present an evolutionary based learning algorithm whieh uses
a relatively comprehensible knowledge structure but is able to accurately predict
the total algal biomass chlorophyll-a and the abundance of algal species.
15.2
Materials and Methods
Water quality data from the hypertrophie lake Kasumigaura on the Kanto plains
North of Tokyo measured by the Japanese National Institute for Environmental
Studies from 1984 until 1993 (Takamura et al. 1992) were used in this study. The
periodicity of the measurements varied from monthly to fortnightly over the 10
years. Lake Kasumigaura has been subjected to cultural eutrophieation in the 70's
and 80's, which made blue-green algal blooms a regular annual feature of the lake
in the past. The data from 1984 until 1993 reveal that different species of
J. Bobbin . F. Recknagel
15.1.1
Knowledge Generalization and Representation
Inductive models learn by generalizing the knowledge they acquire about a certain
domain. The model that is induced from available data leams to associate domain
properties that have equivalent consequences but appear differently. The
complexity of the raw data needs to be aggregated by the model.
There are many ways that knowledge can be represented, including as neural
networks, rule sets, fuzzy sets, learnt equations and regression trees. Most
previous applications of machine leaming in ecology have concentrated on
acquiring quantitatively accurate models of the systems as determined by
independent test data. However, a predictive model should achieve not only
accurate classification but also provide insight and understanding of the predietive
structure of the data (Breiman et al. 1984). Evaluating models solelyon their
predictive performance ignores any information the model can provide about the
relationships it has learnt and how they relate to the causal understanding of
ecosystem behaviour. For example, many ecological applications of neural
networks primarily focus on minimising the root mean square error (RMSE) of
ecosystem predictions based on implicit representations of the relationships by
weights of neurons rather than to explain relationships by which the model decides
its predictions. This chapter considers how ecosystem data can be used to induce
models, whieh are both predictive and explanatory.
Representing knowledge in a transparent manner is of great importance to
inductive modeling. Inductively gained models are a hypothesis derived from the
data, and certainly not a proven law for underlying mechanisms. A model derived
from measured data can only generalize to the extent that the data provided
accurately represent the universe of possible measurements. The representation of
knowledge generalised during this process is critical in terms of understanding of
predietions and knowledge discovery. Improved understanding of inductive
ecosystem models will increase our confidence in the model outputs and gradually
turn them into "grey boxes" .
In this chapter we present an evolutionary based learning algorithm whieh uses
a relatively comprehensible knowledge structure but is able to accurately predict
the total algal biomass chlorophyll-a and the abundance of algal species.
15.2
Materials and Methods
Water quality data from the hypertrophie lake Kasumigaura on the Kanto plains
North of Tokyo measured by the Japanese National Institute for Environmental
Studies from 1984 until 1993 (Takamura et al. 1992) were used in this study. The
periodicity of the measurements varied from monthly to fortnightly over the 10
years. Lake Kasumigaura has been subjected to cultural eutrophieation in the 70's
and 80's, which made blue-green algal blooms a regular annual feature of the lake
in the past. The data from 1984 until 1993 reveal that different species of