Task-Driven Species in Evolutionary Robotic Teams - Foundations on Natural and Artificial Computation - page 138

different population and evaluating them together. This way, if species are re-
quired, the subpopulations will preserve diversity. In this line we must point out
Multi-Agent ESP [13], a neuroevolutionaty algorithm that creates n populations
for deriving n ANN controllers. Each population consists of u sub-populations,
where individual ANNs are constructed by allocating a separate population for
each hidden-layer neuron of the network and a number of neuron populations are
thus evolved simultaneously. An improved version of Multi-Agent ESP is CONE
[8], which includes genotype and behavioral specialization difference metrics to
regulate genotypes between and within populations. It has been successfully
applied in collective gathering and construction tasks.
The main drawback of these coevolutionary approaches is that all of them
have been designed for off-line operation, mainly motivated by their main fea-
ture of having separate populations for each robot. The high computational
requirements of managing as many populations as robots make them inadequate
for real time operation. As a consequence, all the changes in the task and the
environment must be included in the simulation because there is no real-time
adaptation mechanism. However, due to its demonstrated capabilities for au-
tomatic specialization, in this work we will follow a coevolutionary approach
too, but avoiding the use of several populations and introducing an intrinsic
self-organization capability. The algorithm is presented in the next section.
2 Asynchronous Situated Coevolution
Asynchronous Situated Co-evolution (ASiCo) is an evolutionary algorithm in-
spired by natural evolution in terms of the use of decentralized and asynchronous
open-ended evolution. It differs from other bio-inspired approaches such as ge-
netic algorithms or evolutionary strategies where the selection and evaluation
of the individuals is carried out in a centralized manner at regular processing
intervals and where there exists an explicit fitness function. Furthermore, unlike
in traditional evolutionary algorithms, where each individual represents a solu-
tion, in ASiCo a solution is provided by the whole population. To this end, the
algorithm performs a coevolutionary process where the individuals are situated,
that is, all of them ”live” in a scenario and their interactions, including repro-
duction, are local and depend on spatial and/or temporal coincidence, making
the algorithm intrinsically decentralized. Consequently, ASiCo is highly suitable
for distributed and dynamic problems. In this sense, the scenario must be the
problem itself instead of the usual representation of it.
ASiCo has been successfully applied to different types of optimization prob-
lems, like routing [9], autonomous surveillance [10] and shipping freight problems
[11]. In these cases, the tasks were not selected due to their specialization re-
quirements but for being highly relevant distributed problems in engineering.
Here, we are more interested in analyzing in depth what happens with a task
that presents a clear advantage if coordination is achieved with individuals spe-
cialized in different sub-tasks. For a detailed explanation of the algorithm we
recommend [10]. Nevertheless, it is necessary to present its basic concepts and
operation to better understand the experiments and results.