City Research Online

Abstract

The research conducted in this thesis is on distributed termination detection in multiagent systems.

Agents engage in complex interactions by executing behaviour specifications in the form of protocols. This work presents and experiments with a framework for making termination in a multiagent system explicit. As a side effect, the mechanism can be exploited to aid management of agent interactions, by providing visibility of the interaction process and can be extended to drive multiagent system management tasks such as timely garbage collection.

Results from previous attempts to deploy agents systems when scaling up, e.g. Agentcities, have shown and exposed a big gap between theory and practice especially in the reliability and availability of deployed systems. In particular more work needs to be done in the area of supporting agent infrastructures as much as in theoretical agent foundations.

There are two aspects to this problem of termination detection in multiagent systems, firstly, the formal verification of behaviour at compile-time and secondly, monitoring and control at run-time. Regarding the former, there has been some work on the ver- 13 ification of agent communication languages. But overall verification is difficult and often requires knowledge of internal states of agents at compile time, and as yet has not been satisfactorily solved to be deployed in real systems. The second, the runtime approach is adopted in here.

The research is not about protocol engineering but assumes correct protocols, and protocol specifications to be finite state machine graphs. Given these correct verified protocols, the thesis proposes a number of definitions culminating in identification of minimal information in the form of sub-protocols that agents being autonomous, can make available for the termination detection. An off line procedure for deriving these sub-protocols is then presented.

The thesis then considers a termination detection model, and within this model, proposes an conversationmodel encompassing protocol executions, with hierarchical conversations modelled as diffusing computation trees and defines a number of predicates to derive termination in centralised and distributed environments. Algorithms that implement these predicates are sketched and some complexity analysis is performed. The thesis then considers a prototype implementation evaluated over some defined detection delays metric.

The evaluation approach is heavily empirical, with an experimental approach adopted to evaluate various configurations of the termination detection mechanism. The evaluation employs robust resampling and bootstrapping methods to analyse and obtain distributions and confidence intervals of the detection delays metric for the termination detection mechanism.

Publication Type:	Thesis (Doctoral)
Subjects:	T Technology
Departments:	Doctoral Theses School of Science & Technology > Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses

Export

Downloads