A Comparison of Interest Manager Mechanisms for Agent-Based Simulations using a Time Warp Executive

摘要

An important issue for agent-based simulation architectures concerns the means by which agents can be kept up to date regarding the position and state of other agents and objects in their local environment. We implement two distributed publish-subscribe approaches and assess their performance within our SASSY architecture. One implementation supports direct agent-to-agent communication, while in the other all messages pass through an intermediate LP. In both cases the system does not maintain global state, but facilitates communication between individual agents that may affect one another. The environment is decomposed into a set of regions, where an LP referred to as an Interest Manager (IM) manages each region. When agents enter a region they register their presence with the appropriate Interest Manager. Agents also subscribe and unsubscribe to information from neighboring regions as they move through the environment. Publish-subscribe approaches have been utilized in other simulation environments (e.g., HLA), but we believe our approach is the first to use a peer-to-peer method in an optimistic environment. We assess the efficiency of these two methods by measuring the number of rollbacks, and total runtime for various configurations of the simulation.