Assuring agent interaction through run-time monitoring and control

A multiagent system (MAS) is composed of autonomous agents that interact and exist in a shared environment. A fundamental aspect of multiagent systems is the communication that drives collaboration and cooperation among agents. Verification of agent communication is key to predictable and efficient interaction. Verification techniques of agent interaction have mostly been applied during design time. However, a multiagent system environment is inherently complex and presents challenges such as the heterogeneity of agents and the dynamic nature of the environment. To address such challenges, run-time approaches are needed to complement design time verification. This thesis presents a run-time monitoring and control approach for improving agent communication, where the environment takes a supervisory role through a Governing Agent (GA). The role of the governing agent is to minimise the negative effects of issues in interaction through monitoring and control. Run-time monitoring of interaction of agents is modeled with an Interaction Petri net (IPN). The proposed Petri net model allows the detection of common undesired scenarios such as protocol delay, lost messages, busy-wait, transmission delay, and agent termination. When scenarios are detected, control actions are taken by the governing agent. The proposed approach is evaluated with an experimental analysis and has been shown to minimise the negative effect of undesired interaction scenarios. The GA successfully detects issues in communication and apply control actions of interaction in two different classes of case studies. The first case study is a multiagent treasure hunt with collaborative interaction, while the second presents a competitive interaction demonstrated by an auction. Furthermore, a comparative analysis with a prior Petri net model of interaction is carried out to highlight the strengths and limitations of the new IPN model.