A bacterial based distributed gradient descent model for mass scale evacuations

Most current methods for distributed agent coordination rely on large messages, complex computations and unique identifiability of communicating agents. To allow for efficient coordination in constrained environments with limited communication and computational resources, we derived a distributed gradient descent (DGD) algorithm based on how bacteria cells interact when searching for food. The method combines the (static) map of each agent with messages received from other agents. Both messages and computations are restricted to a simple vocabulary and are aggregated so no knowledge of the sender is required. In contrast to most prior DGD methods, where the interaction graph is assumed to be static, we allow for dynamic changes in the neighborhood graph for each agent. We prove that the method still correctly converges to a local optimum in such dynamic settings. We tested the usefulness of the method in simulations of mass-scale emergency evacuations in which obstacles are dynamically added to a given environment. We show that our bacterial based model can drastically reduce evacuation times in complex and realistic environments, when compared to prior models.