A Load Sharing Approach Based on Refactoring of Roles in Multi-Agent Systems

1. A Load Sharing Approach Based on Refactoring of Rolesin Multi-Agent Systems Sebnem Bora, A. Murat Tiryaki and Oguz Dikenelli Ege University

2. Overview • Problem Description • Realization of Load Sharing Components in a MAS Environment • Evaluation of the Approach • Conclusion

3. Problem Description • It is almost impossible to guess load of agents. Load of the agents may increase at run time. This may lead to the agent failing with all requests unable to be performed. • All features of an agent come from the roles that are assigned to it. Overloading of an agent also comes from its role(s). • Assigning these roles to the other agents does not solve overloading. • We propose a load sharing approach based on role refactoring. • In our approach, the role on a heavily loaded agent is divided into new sub-roles with respect to a policy.

4. Problem Description(2)

5. Realization of Load Sharing Components in a MAS Environment • A load sharing algorithm has four components:Transfer,selection, location, and information policies. • A transfer policy determines whether an object may participate in a task transfer. • A selection policy selects a task for transfer, when the transfer policy decides that the object is a sender. • A location policy finds suitable senders or receivers to share load. • An information policy decides when information about the other objects in the system should be collected, where it should be collected from and what information should be collected.

6. Realization of Load Sharing Components in a MAS Environment(2) • Our selection policy selects the predefined subrole to be transfered by considering the workload of the role and a refactoring ontology. (We need a complex monitoring mechanism).

7. Realization of Load Sharing Components in a MAS Environment(3) • Our location policy has to identify the suitable agent to transfer the selected subrole. • Our information policy is periodic and centralized. Each domain agent periodically sends its workload data. • The monitoring mechanism evaluates the excess load of roles by applying our threshold policy. • Therefore, we created the “monitor role”.

8. Realization of Load Sharing Components in a MAS Environment(4) • Monitor agent: The agent plays the “monitor role” • It has two responsibilities: Evaluation of the collected information and applying of role splitting.

9. Realization of Load Sharing Components in a MAS Environment(5)

10. Realization of Load Sharing Components in a MAS Environment(6) • Evaluation of the collected information:

11. Realization of Load Sharing Components in a MAS Environment(7) • The excess workload of each role is given by the following equations:

12. Realization of Load Sharing Components in a MAS Environment(8) • Role Splitting • When the monitor agent receives the message that includes the request for rolesplitting, it initiates therole splitting plan.

13. Realization of Load Sharing Components in a MAS Environment(9) • The definitions of the new roles are passed to the reorganizingsystems' roles.

14. Evaluation of the Approach • We applied it to an agentsystem application which was implemented by SEAGENT Research Group. Thecase focuses on one of the core scenarios of the tourism domain.

15. Evaluation of the Approach(2) • The role splitting ontology

16. Evaluation of the Approach(3) • During the system execution,the travel agency role is split into two sub-roles as defined inthe design phase, when the number of requests sent to the travel agencyincreases considerably.

17. Evaluation of the Approach(4) • The performance results of the implementation

18. Conclusion • The load sharingapproach based on refactoring of roles improves system's performance underheavy load conditions. • In conclusion, first results are encouraging and indicatingthat efficient load sharing in multi-agent organizations is sustainable using role refactoring.

19. THANK YOU VERY MUCH.