A decentralized multi-objective optimization algorithm (Q2032002)

In this article is proposed a new gradient-based optimization algorithm and all necessary arguments, such as proof of convergence to solutions and convergence rates are presented. The algorithm is for multi-agent multi-objective set constrained problems. The agents have an initial vector of priorities (weights), and an initial vector of decision variables. The weights determine how the agents update the decision variables according with information received from the other agents. The proposed algorithm surveys the Pareto front and performs four steps at each iteration as follows:\begin{itemize} \item[1)] agent i updates its vector of priorities using those received from the other agents in the network, \item[2)] the vectors of priorities generate the matrix of information weights for the decision variable update, \item[3)] agent i updates its vector of decision variables with variables received from its neighbors, \item[4)] agent i takes a gradient descent step and projects its new decision variables onto the constraint set.\end{itemize} All necessary mathematical arguments are presented in this work or in the annex. The effectiveness of the proposed algorithm with different agents is illustrated by some numerical results, presented in the last part of the paper.