A Derivative-free Distributed Optimization Algorithm with Applications in Multi-Agent Target Tracking

Multi-agent tracking of a moving target can be modeled as a distributed optimization problem of a time-varying objective function that has an optimum at the ideal sensing states of the agents. The inputs to the objective function are some observed parameters of the target which are obtained from onboard sensory information. In particular, due to recent progress in learning-based vision techniques, these observed parameters may be of large dimensions which may limit communication capabilities, especially for large groups of coordinating agents. In cases where the analytical form of the objective function is unknown or the gradient of the objective function is difficult to estimate, which may occur due to a fast moving target or the high dimensionality of the observation parameters, gradient-based solutions may be inapplicable or computationally prohibitive to apply. In this paper, we propose a derivative-free distributed optimization algorithm based on distributed active perception for multi-agent target tracking. Our proposed method can optimize objective functions without knowledge of the gradient and does not require communication. We derive the information dynamics for general dimensions which are used to analyze the tracking convergence. Simulations and experimental results are provided.