A Visualization Technique for Multiagent Systems
The British University in Dubai (BUiD)
In this dissertation we consider the problem of monitoring and visualizing the performance of multi-agent systems, i.e. systems of adaptive agents that change their behaviour as a result of learning and experiencing. Instead of relying on global performance measures, as have been done by current visualization techniques, we use a different technique which is a combination of dimensionality reduction and social network measures. The advantage of this technique, we claim, is that it does not only capture the performance of the multiagent system on its macro level, as is the case with the global performance metrics methods, but it can also capture the performance on the micro level, i.e. by being sensitive to the performance of individual agents.To test our technique, we first conduct several experiments to compare different combinations of dimensionality reduction techniques and network measures. Then we apply one such combination on networks of adaptive agents that play games (we use the term ”game” as in the Game theory). Our findings confirm that using dimensionality-reduced weighted network measures in visualizing the performance of multi-agent systems is informative in the sense that they proved to be sensitive to changes in the global as well as local system dynamics (for example the global network structure and the local learning of individual agents).
DISSERTATION WITH DISTINCTION
multi-agent, visualization techniques, global network structure