62
P.A. Whigham . G.B. Fogel
and the current observed spatial positions of the insects, as viewed by the lizard.
These types of models show promise in producing hypotheses regarding
individual behaviour and proposing theories that could be tested in the field.
A second example of individual behaviour is the study of trail following, in
particular ant foraging and the use of pheromones for trail marking. The simplest
concept is the evolution of trail following, as demonstrated by the Sante Fe trail
experiments (Koza 1992). Here an artificial ant must learn to produce a trail
following strategy that disco vers food placed along a trail scattered amongst a
144-cell grid. GP was shown to be capable of finding solutions to this problem,
even though the fitness function was based purelyon the number of food units
discovered within a certain number of evaluations. This type of model indicates
that complex spatial behaviour for an individual can be evolved and studied,
which may be used to understand observed foraging behaviour of real species.
Extending the concepts of a single population attempting to find a single best
solution, coevolutionary algorithms use a number of subpopulations where each
subpopulation evolves competing (rather than cooperating) solutions (Cohoon et
al. 1987; Whitley and Starkweather 1990). Extensions of these ideas include the
cooperative coevolutionary genetic algorithm (GA) (Potter and De Jong 1994),
where a subpopulation represents a species that solves one particular aspect of a
problem. The final complete solution is obtained by assembling representative
members from each subpopulation. The subpopulations evolve independently,
using aGA, where the goal is to have each subpopulation solve one aspect of the
problem. This has similarities to the concepts of speciation, where each
subpopulation finds a niche in the solution space to exploit. Other approaches
with coevolution allow the modelling of individual behaviour in the population to
produce models of predator-prey interactions (Cliff and Miller 1996; Haynes and
Sen 1996; Rosin and Belew 1997) and other forms of competition. Each
individual is represented as a bit string, neural network, or symbolic function that
can evolve to produce behavior based on the competition produced from other
individuals in the population. These approaches have been successful in
demonstrating concepts such as diversity, extinction and genetic drift. However,
these systems are difficult to interpret when applied to real ecosystem behavior or
when they are coupled with data based on measured systems.
Extensions to multiple populations use the explicit representation of space,
using CA or other grid-based representations, to allow spatial interactions to be
explicitly represented. An early example of this approach was 'Tierra' (Ray
1992). Here organisms are represented as simple computer programs that compete
for the memory and processing resources of the computer. The population
changed over time through reproduction with mutation, and was constrained in
terms of total size by having old individuals, or those that performed poorly, being
gradually removed. The work showed that parasites could evolve in the
population that used parts of other organisms for their own benefit.
A more complex system using self-replication and cooperating bits of computer
code that evolved in a virtual computer wOrld was called 'Avida' (Adami 1998).
P.A. Whigham . G.B. Fogel
and the current observed spatial positions of the insects, as viewed by the lizard.
These types of models show promise in producing hypotheses regarding
individual behaviour and proposing theories that could be tested in the field.
A second example of individual behaviour is the study of trail following, in
particular ant foraging and the use of pheromones for trail marking. The simplest
concept is the evolution of trail following, as demonstrated by the Sante Fe trail
experiments (Koza 1992). Here an artificial ant must learn to produce a trail
following strategy that disco vers food placed along a trail scattered amongst a
144-cell grid. GP was shown to be capable of finding solutions to this problem,
even though the fitness function was based purelyon the number of food units
discovered within a certain number of evaluations. This type of model indicates
that complex spatial behaviour for an individual can be evolved and studied,
which may be used to understand observed foraging behaviour of real species.
Extending the concepts of a single population attempting to find a single best
solution, coevolutionary algorithms use a number of subpopulations where each
subpopulation evolves competing (rather than cooperating) solutions (Cohoon et
al. 1987; Whitley and Starkweather 1990). Extensions of these ideas include the
cooperative coevolutionary genetic algorithm (GA) (Potter and De Jong 1994),
where a subpopulation represents a species that solves one particular aspect of a
problem. The final complete solution is obtained by assembling representative
members from each subpopulation. The subpopulations evolve independently,
using aGA, where the goal is to have each subpopulation solve one aspect of the
problem. This has similarities to the concepts of speciation, where each
subpopulation finds a niche in the solution space to exploit. Other approaches
with coevolution allow the modelling of individual behaviour in the population to
produce models of predator-prey interactions (Cliff and Miller 1996; Haynes and
Sen 1996; Rosin and Belew 1997) and other forms of competition. Each
individual is represented as a bit string, neural network, or symbolic function that
can evolve to produce behavior based on the competition produced from other
individuals in the population. These approaches have been successful in
demonstrating concepts such as diversity, extinction and genetic drift. However,
these systems are difficult to interpret when applied to real ecosystem behavior or
when they are coupled with data based on measured systems.
Extensions to multiple populations use the explicit representation of space,
using CA or other grid-based representations, to allow spatial interactions to be
explicitly represented. An early example of this approach was 'Tierra' (Ray
1992). Here organisms are represented as simple computer programs that compete
for the memory and processing resources of the computer. The population
changed over time through reproduction with mutation, and was constrained in
terms of total size by having old individuals, or those that performed poorly, being
gradually removed. The work showed that parasites could evolve in the
population that used parts of other organisms for their own benefit.
A more complex system using self-replication and cooperating bits of computer
code that evolved in a virtual computer wOrld was called 'Avida' (Adami 1998).