Scalable Distributed Simulation for Evo...

Swarms of cyber-physical systems can be used to tackle many challenges that traditional multi-robot systems fail to address. In particular, the self-organizing nature of swarms ensures they are both scalable and adaptable. Such benefits come at the cost of having a highly complex system that is extremely hard to design manually. Therefore, an automated process is required for designing the local interactions between the cyber-physical systems that lead to the desired swarm behavior. In this work, the authors employ evolutionary design methodologies to generate the local controllers of the cyber-physical systems. This requires many simulation runs, which can be parallelized. Two approaches are proposed for distributing simulations among multiple servers. First, an approach where the distributed simulators are controlled centrally and second, a distributed approach where the controllers are exported to the simulators running stand-alone. The authors show that the distributed approach is suited for most scenarios and propose a network-based architecture. To evaluate the performance, the authors provide an implementation that builds upon the eXtensible Messaging and Presence Protocol (XMPP) and supersedes a previous implementation based on the Message Queue Telemetry Transport (MQTT) protocol. Measurements of the total optimization time show that it outperforms the previous implementation in certain cases by a factor greater than three. A scalability analysis shows that it is inversely proportional to the number of simulation servers and scales very well. Finally, a proof of concept demonstrates the ability to deploy the resulting controller onto cyber-physical systems. The results demonstrate the flexibility of the architecture and its performance. Therefore, it is well suited for distributing the simulation workload among multiple servers.