Agent-Environment Interactions in Large-Scale Multi-Agent Based Simulation Systems

In this paper we present the Action-Potential/Result (APR) model for agent-environment interactions in Multi-Agent Based Simulation systems (MABS) involving thousands of perception-based agents executing on a single host. The environment structure is partitioned into cells which are managed by specialized agents called controller and coordinator. The agents send their stimuli to the controller managing the cell in which they are situated. The controller assesses the received agent stimuli as well as user-triggered and events propagation stimuli. It combines them, attempts to resolve potential conflicts, communicates with adjacent controllers and the coordinator, and ensures that the updated environment state is communicated back to its agents. The APR model has been implemented as a component of the DIVAs framework and can be reused with minimal changes for the construction of agent-based simulations with DIVAs. Experimental results show a significant improvement in scalability over conventional centralized solutions.