Abstract
Holonic multiagent systems (HMAS) are a sub-field of study within artificial intelligence
(AI) which has not received much attention in recent mainstream AI
topics such as machine learning (ML). This research studies how holonism, introduced
through recursive modelling techniques, benefit multiagent systems by exploring
the relationship between recursive self-similarity and emergent intelligent
behaviour.
The scope of multiagent learning (MAL) applications is expanded into the domain
supervised learning (SL) problems by designing an abstract model for a
holonic multiagent system (HMAS) as a hierarchy of autonomous computing entities,
called holons. The self-similar structure of holons is both stable and coherent,
and likewise consists of one or more holons. It is also shown how benefits arise from
the intersection of holonic self-similarity and the property of homoiconicity resulting
in a model that is novel in terms of its parallelism and ability to scale in order
to be distributed in a cloud environment. Homoiconicity refers to the self-similar
property of a programming language where the external code written by the programmer
is a textual representation of lists, which are data structures native to the
language.
A set of models were developed and evaluated by measuring classification problem
performance measures such as the accuracy, precision, recall and f-measure. The
communication costs between interacting holons were also analysed in terms of the
replication rate and the reducer size.
The application of the model in a recursive extension to MapReduce for decision
tree learning obtained results showing that when the algorithms are applied
in a classification problem domain, they are able to perform consistent with their
expected behaviour. In the application of the model to homoiconic agent-oriented
trie data structures, namely the Patricia tries and standard tries, the output showed
results consistent with the fact that Patricia trie algorithms build a compressed trie
and, therefore, the communication cost and replication rate values are smaller than
the standard trie.
Research areas in cooperative MAL have found success in using selection-based
methods, such as Evolutionary Algorithms (EAs). This research showed that Island
genetic algorithms applied to learning decision trees can be implemented
in a holonic multiagent-based environment, resulting in a model expressing selfsimilarity
of the problem type and the architecture, achieved by integrating the
learning algorithm of decision trees directly into the evolutionary process. The
HMAS produced a stronger decision tree predictor when compared to three known
parallel GA models.