An evolutionary approach to swarm adaptation in dense environments

Abstract

The ability for a swarm of mobile agents to quickly adapt in unknown environments and reach a goal while avoiding obstacles and maintaining a formation is extremely important in time critical tasks. We utilize a physics-based autonomous agent framework combined with our DAEDALUS paradigm which allows the agents to learn from the neighboring agents. In traditional approaches, a swarm of agents learn the task in simulation(offline) combined with an evolutionary/genetic algorithm, and a global observer optimizes the swarm performance. In real world(online), the swarm of agents may have to rapidly adapt in unfamiliar environments. When there is no global observer and the online(real world) environment is dense with obstacles compared to offline environment, the performance feedback may be delayed or perturbed by noise, and the rules learned in simulation(offline) may not be sufficient to overcome the navigational difficulties, leaving the swarm to rapidly adapt in new environment. DAEDALUS is a paradigm designed to address these issues, by mimicking more closely the actual dynamics of populations of agents moving and interacting in a task environment. This paper presents an analysis of swarm adaptation using DAEDALUS in high obstacle density environments where agent interactions could be obstructed by obstacles.