Date of Award

2012

Publication Type

Doctoral Thesis

Degree Name

Ph.D.

Department

Computer Science

Keywords

Applied sciences, Agent specialization, Multiagent populations

Supervisor

Ziad Kobti

Rights

info:eu-repo/semantics/openAccess

Abstract

In this dissertation, I present the Weight-Allocated Social Pressure System (WASPS). WASPS is a computational framework that when applied, can allow for the increase in agent specialization within a multi-agent population. Research has shown that specialization can lead to an overall increase in the productivity levels within a population [55]. WASPS aims to provide a mix of features from existing frameworks such as the genetic threshold and social inhibition models. It also subsumes these models, and allows hybrids of them to be created. It provides individual level behaviour as found in the genetic threshold model. As in some variations of the genetic threshold model [49], WASPS also allows for individual level learning. As found in the social inhibition models, WASPS allows for social influence, or population level learning. Unlike some models, WASPS allows agents to self-organize based on available tasks. In addition, it makes allowances for agents to allocate a resource among multiple tasks during a work period, wherein most models allow the selection of only one task.

WASPS allows the assumption that agents are heterogeneous in their task performance aptitudes. It thus aims to create skill-based agent specialization within the population. This will allow more skilled agents to allocate more resources to tasks for which they have comparative advantages over their competition. Because WASPS is self-organizing, it can handle the addition and removal of agents from social networks, as well as changes in the connections between agents. WASPS does not limit the definition of many or its parameters, which allows it to deal with changing definitions for those parameters. For example, WASPS can easily adjust to deal with changing definitions of agent skill and influence. In fact, the individual level learning can be implemented in such a way that an agent can self-optimize even when it has no competitors to influence it.

Download

Share

COinS