Adaptive optimistic simulation of multi-agent systems

Simulation is an important tool for designers of multi-agent systems allowing them to learn more about the behaviour of a system or to investigate the implications of alternative agent architectures. A key issue for agent simulation is that of scalability, as agents are themselves often complex systems (e.g., with sensing, planning, inference etc. capabilities), requiring considerable computational resources. One way to address scalability issues is the application of Parallel Discrete Event Simulation (PDES) techniques to multi-agent systems (MAS). Synchronisation of parallel discrete event simulations has long been established as a difficult problem and despite many years of research no one synchronisation method has been shown to work for all types of application. The goal of synchronisation is to ensure the events of a parallel discrete event simulation are processed so as the result of the simulation is the same as the equivalent sequential simulation. This thesis presents a series of synchronisation mechanisms for optimistic simulation of multi-agent systems and in particular for the PDES-MAS framework. The thesis develops a model of accesses to the shared state of an optimistic PDES of MAS which exploits the read/write pattern of agent behaviour to reduce rollback and provide efficient state saving. Two adaptive synchronisation algorithms are then presented. The first of these algorithms uses the notion of critical accesses and a window based throttling mechanism to constrain execution of LPs. The second uses a decision theoretic approach to delay incoming read events which are likely to be rolled back. The effectiveness of these algorithms is then investigated experimentally using the ASSK kernel with a range of different multi-agent simulations. The algorithms are shown to provide significant reduction in rollbacks for a variety of different cases.