Chapter 15 . Evolved Predictive Rules for Aigal Dynamics
15.4
Discussion
309
The application of evolutionary algorithms proved to be able to optimise symbolic
rule structures and associated parameter vectors in order to discover predictive
rule for chlorophyll-a and blue green algae species in lake Kasumigaura. The
models are comprehensible and explanatory. Tracing the decision processes that
the rule set models undertake allows to make predictions transparent. Further,
there are no assumptions required on the structure of the model itself, which
enables arbitrary criteria to be placed upon the model design to suit the questions
of interest. This was demonstrated by limiting the number of rules that a model
could contain and by using a discontinuous fitness function. The first condition
was imposed to maximise comprehensibility of the evolved rule sets, and the
second was imposed to force the rule sets to accurately predict the algal
abundance phenomena of interest.
The rules discovered for the prediction of chlorophyll-a, Microcystis and
filamentous algae Oscillatoria and Phormidium are explanatory and are likely to
contain some information about ecosystem processes related to algal growth.
However the hypothesis induced by evolutionary algorithms need to be tested by
laboratory and field experiments.
15.5
Conclusions
Knowledge representation is an important issue far inductive leaming algorithms.
Representations that allow direct expression of leamt knowledge can be used to
examine the underlying hypothesis of a discovered model. The modus operandi of
the model can then be compared to ecological theory.
In this paper we have firstly evolved a ruleset-based model for chlorophyll-a
that was compared with a ruleset extracted from the same database by the
alternative classification and regression tree (CART) method. The comparison
has demonstrated that the evolutionary algorithm model performed better than the
CART model indicated by persistently smaller RMSE.
Secondly predictive ruleset were evolved for two common groups ofblue-green
algae in lake Kasumigaura: Microcystis and the filamentous Oscillatoria and
Phormidium. These models clearly demonstrated their capability to discover
distinctions in environmental preferences of the two groups, and that these
distinctions can be learnt from data.
In essence the present study on highly complex limnological data has
demonstrated the great potential of evolutionary algorithms not only as tools for
predictive modeling but also for knowledge discovery.
15.4
Discussion
309
The application of evolutionary algorithms proved to be able to optimise symbolic
rule structures and associated parameter vectors in order to discover predictive
rule for chlorophyll-a and blue green algae species in lake Kasumigaura. The
models are comprehensible and explanatory. Tracing the decision processes that
the rule set models undertake allows to make predictions transparent. Further,
there are no assumptions required on the structure of the model itself, which
enables arbitrary criteria to be placed upon the model design to suit the questions
of interest. This was demonstrated by limiting the number of rules that a model
could contain and by using a discontinuous fitness function. The first condition
was imposed to maximise comprehensibility of the evolved rule sets, and the
second was imposed to force the rule sets to accurately predict the algal
abundance phenomena of interest.
The rules discovered for the prediction of chlorophyll-a, Microcystis and
filamentous algae Oscillatoria and Phormidium are explanatory and are likely to
contain some information about ecosystem processes related to algal growth.
However the hypothesis induced by evolutionary algorithms need to be tested by
laboratory and field experiments.
15.5
Conclusions
Knowledge representation is an important issue far inductive leaming algorithms.
Representations that allow direct expression of leamt knowledge can be used to
examine the underlying hypothesis of a discovered model. The modus operandi of
the model can then be compared to ecological theory.
In this paper we have firstly evolved a ruleset-based model for chlorophyll-a
that was compared with a ruleset extracted from the same database by the
alternative classification and regression tree (CART) method. The comparison
has demonstrated that the evolutionary algorithm model performed better than the
CART model indicated by persistently smaller RMSE.
Secondly predictive ruleset were evolved for two common groups ofblue-green
algae in lake Kasumigaura: Microcystis and the filamentous Oscillatoria and
Phormidium. These models clearly demonstrated their capability to discover
distinctions in environmental preferences of the two groups, and that these
distinctions can be learnt from data.
In essence the present study on highly complex limnological data has
demonstrated the great potential of evolutionary algorithms not only as tools for
predictive modeling but also for knowledge discovery.