MULTI-AGENT SOURCE LOCALIZATION USING PASSIVE SENSING

Abstract

Purposeful collective behaviour in multi-agent systems can be achieved from a mix of simple individualistic and social behaviours of an agent. Social behaviours are the basis of cooperation in multi-agent systems and are fundamental in achieving collective behaviour. Practical implementation of conventional social behaviour models require explicit inter-agent communication, whereas in some environments, communication bandwidth and delays are critical constraints which may compromise the intended collective behaviour. This thesis introduces three source localization algorithms. Each algorithm is a set of individualistic and social behaviours, which do not require explicit inter-agent communication and rely solely on agent's passive sensing.

The first source localization algorithm is composed of static individualistic and social behaviours. The individualistic behaviour is inspired from a bacterium's random walk while performing chemotaxis and is self-sufficient in localizing sources of interest. Self sufficiency means that an agent can localize a source on its own using only its individualistic behaviour without any team cooperation via its social behaviours. However, better localization performance can be achieved when an agent uses an optimized weighted average of both individualistic and social behaviours. The social behaviours are inspired from the long-range attraction and the short-range repulsion behaviours of a fish. The second source localization algorithm assumes an adaptive individualistic behaviour while keeping the social behaviours static. Finally, the third source localization algorithm is based on adaptive social behaviours without a self-sufficient individualistic behaviour and source localization is achieved as an emergent property of the social interactions between agents.

The agent behaviours for each source localization algorithm have been optimized using a Genetic Algorithm. Small homogeneous multi-robot systems are considered where neither the position information of the agents nor the position information of the source is available. An agent is assumed to have a single sensor to sense the source intensity and hence conducts temporal sensing to sense the gradient. For social interaction, an agent is assumed to have two sensors to detect the neighbour-majority either in its right or left sensing half. The behavioural optimization is carried out for a realistic underwater acoustic source in a range of initialization distances, neighbourhood radii and team sizes. The optimization data has been estimated by an analytical model for each localization algorithm. The performance of the collective behaviour resulting from the estimated model has been validated against agent's sensor and actuator noise along with strong multi-path interference due to variability of the environment.

Given the constraints of temporal sensing and loss of information due to noisy and simplistic passive sensing, the collective behaviours show remarkable robustness and scalability in terms of mean, median and variance of the arrival time distributions. Investigation of the team expanse in strong multi-path interference shows that team remains cohesive with minimal or no agent loss during the localization mission.