We propose RMASBench, a new benchmarking tool based on the RoboCup Rescue Agent simulation system, to easily compare coordination approaches in controlled settings for dynamic rescue scenario. In particular, we offer simple interfaces to plug-in coordination algorithms without the need for implementing and tuning low-level agents' behaviors. Moreover, we add to the realism of the simulation by providing a large scale crowd simulator, which exploits GPUs parallel architecture, to simulate the behaviour of thousands of agents in real time. Hence, we present two key benchmarks for coordination algorithms. First we focus on a specific coordination problem where fire fighters must combat fires and prevent them from spreading across the city. We formalize this problem as a Distributed Constraint Optimization Problem and we compare two state-of-the art solution techniques: DSA and MaxSum and thus provide the first benchmarks for DCOPs in this domain. Second, we provide key results on the evacuation of crowds of 3000 civilians on different maps under different blockage conditions. Thus we define benchmark scenarios for the development of coordination algorithms for large scale crowd evacuation. Our results demonstrate that RMASBench offers powerful tools to compare coordination algorithms in a dynamic environment.
|Titolo:||RMASBench: Benchmarking Dynamic Multi-agent Coordination in Urban Search and Rescue|
|Data di pubblicazione:||2013|
|Appare nelle tipologie:||04.02 Abstract in Atti di convegno|