Author:
Nikolaos Chatzinikolaou
Affiliation:
School of Informatics, University of Edinburgh, United Kingdom
Keyword(s):
Genetic algorithms, Distributed computation, Multi-agent learning, Agent coordination.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence and Decision Support Systems
;
Coordination in Multi-Agent Systems
;
Enterprise Information Systems
;
Evolutionary Programming
;
Intelligent Social Agents and Distributed Artificial Intelligence Applications
Abstract:
In large scale optimisation problems, the aim is to find near-optimal solutions in very large combinatorial spaces. This learning/optimisation process can be aided by parallelisation, but it normally is difficult for engineers to decide in advance how to split the task into appropriate segments attuned to the agents working on them. This paper chooses a particular style of algorithm (a form of genetic algorithm) and describes a framework in which the parallelisation and tuning of the multi-agent system is performed automatically using a combination of self-adaptation of the agents plus sharing of negotiation protocols between agents. These GA agents are optimised themselves through the use of an evolutionary process of selection and recombination. Agents are selected according to the fitness of their respective populations, and during the recombination phase they exchange individuals from their population as well as their optimisation parameters, which is what lends the system its se
lf-adaptive properties. This allows the execution of optimal optimisations without the burden of tuning the evolutionary process by hand. The architecture we use has been shown to be capable of operating in peer to peer environments, raising confidence in its scalability through the autonomy of its components.
(More)