Probabilistic Ants (PAnts) in Multi-Agent Patrolling

We propose a probabilistic ants (PAnts) algorithm for solving the multi-agent patrolling problem in an online and robust manner, based purely on local information. As only local information is required, this strategy can be deployed distributively. As our proposed strategy does not require a pre-processing of the map, it can be used for a map with a dynamic topology as well as dynamically changing number of agents. Our proposed strategy makes use of virtual pheromone traces which will act as potential fields, guiding each agent towards areas which have not been visited for a long time. Each agent only needs to make its decision on where to go next based on its local pheromone information. It does not need to keep a topology of the map in memory. Decision making is done probabilistically based on local pheromone information. This method is also non-intrusive to the environment and all traces are kept in virtual memory. In our experimental evaluation, we compare our method with the traditional ant algorithm as well as a variant of it. All three methods are benchmarked against the theoretical ideal for clarity.