Chapter 5
Ecological Applications of Adaptive Agents
F. Recknagel
5.1
Introduction
Ecologists are constantly searching for new modelling paradigms in order to
simulate realistically the distinct nature of ecosystems by computer models. The
ecosystem concept as established by Forbes (1887) had the most forming
influence on ecosystem modelling in the past century. It no longer bears elose
examination as ecosystems like lakes are known to evolve and being driven by
exogenous forces rather than existing permanently and in isolation. However, the
ecosystem approach resulted in valuable databases from monitoring as weil as
quantitative and qualitative descriptions of ecosystem dynamics and has made
ecology a predictive science (Rigler and Peters 1995). Computer models resulting
from the ecosystem concept were mainly based on differential equations (DE) for
well-defined ecological entities and processes, adjusted by measured or estimated
parameters. Radtke and Straskraba (1980) firstly tried to overcome the rigidity of
such models by parameter optimization of ecological goal functions relevant to
lake ecosystems as introduced by Straskraba (1977). The authors considered their
results as contribution to a structural self-optimising ecosystem model but
admitted that more adequate models and more suitable optimisation procedures
would be needed to make it a success. In order to overcome model rigidity,
Kaluzny and Swartzman (1985) suggested a library of alternative representations
of ecological processes from where a simulation model picks the most relevant
one for a specific ecological situation. The authors coneluded that their approach
was limited by validation data and 'the difficulty of tracing model response to
single processes' (Kaluzny and Swartzman 1985). Jorgensen and Mejer (1979)
introduced the thermodynamic entity exergy for holistic ecosystem modelling that
has led to the concept of structural dynamic models (Jorgensen 1986). It equips an
ecosystem model with aglobai rather than local goal function, namely maximizing
exergy storage, to be satisfied by optimising process parameters in the course of
simulation. Even though this approach avoids the problem of biasing by 'Iocal'
optima as faced by Radtke and Straskraba (1980), it may require more adequate
models and more suitable optimisation procedures as weil.
Machine learning techniques such as artificial neural networks (ANN)
(Rurnelhart et al. 1986) and evolutionary computation (EC) (Holland 1992) allow
looking at the same problem from a different angle. They are inductive techniques
and allow extracting empirical patterns as reflected by multivariate nonlinear time
Ecological Applications of Adaptive Agents
F. Recknagel
5.1
Introduction
Ecologists are constantly searching for new modelling paradigms in order to
simulate realistically the distinct nature of ecosystems by computer models. The
ecosystem concept as established by Forbes (1887) had the most forming
influence on ecosystem modelling in the past century. It no longer bears elose
examination as ecosystems like lakes are known to evolve and being driven by
exogenous forces rather than existing permanently and in isolation. However, the
ecosystem approach resulted in valuable databases from monitoring as weil as
quantitative and qualitative descriptions of ecosystem dynamics and has made
ecology a predictive science (Rigler and Peters 1995). Computer models resulting
from the ecosystem concept were mainly based on differential equations (DE) for
well-defined ecological entities and processes, adjusted by measured or estimated
parameters. Radtke and Straskraba (1980) firstly tried to overcome the rigidity of
such models by parameter optimization of ecological goal functions relevant to
lake ecosystems as introduced by Straskraba (1977). The authors considered their
results as contribution to a structural self-optimising ecosystem model but
admitted that more adequate models and more suitable optimisation procedures
would be needed to make it a success. In order to overcome model rigidity,
Kaluzny and Swartzman (1985) suggested a library of alternative representations
of ecological processes from where a simulation model picks the most relevant
one for a specific ecological situation. The authors coneluded that their approach
was limited by validation data and 'the difficulty of tracing model response to
single processes' (Kaluzny and Swartzman 1985). Jorgensen and Mejer (1979)
introduced the thermodynamic entity exergy for holistic ecosystem modelling that
has led to the concept of structural dynamic models (Jorgensen 1986). It equips an
ecosystem model with aglobai rather than local goal function, namely maximizing
exergy storage, to be satisfied by optimising process parameters in the course of
simulation. Even though this approach avoids the problem of biasing by 'Iocal'
optima as faced by Radtke and Straskraba (1980), it may require more adequate
models and more suitable optimisation procedures as weil.
Machine learning techniques such as artificial neural networks (ANN)
(Rurnelhart et al. 1986) and evolutionary computation (EC) (Holland 1992) allow
looking at the same problem from a different angle. They are inductive techniques
and allow extracting empirical patterns as reflected by multivariate nonlinear time