Current models for multi-agent systems almost exclusively employ sensory modalities such as vision where agents passively receive information from the environment. Active sensing, defined as acquiring environmental information using self-generated signals, allows widespread sharing of sensory information among agents and thus gives rise to more complex interactions within engineered multi-agent systems using radar or sonar, for example. In nature, bat swarms are animal groups that successfully employ active sensing with each individual broadcasting echolocation pulses in the environment and responding to echoes. Bats flying in groups may cope with the dense sound environment through their behavior; one hypothesized strategy is the cessation of echolocation pulses in the presence of peers and “eavesdropping”, which has been demonstrated in controlled laboratory settings. In this work, we build a self-propelled-particle model with each agent avoiding obstacles in three dimensions by emitting echolocation pulses of a unique frequency. We implement a bat-inspired rule of eavesdropping to take advantage of information sharing via active sensing while reducing the energy expenditure of the group. Through a simulation study, we show that agents indeed capitalize on peers’ pulses and echoes for obstacle avoidance and we find a maximum of this effect for a set of model parameters which relate to the domain size.
Neural networks and Monte-Carlo method usage in multi-agent systems for sudoku problem solving
