A new approach to decentralized constrained nonlinear estimation over noisy communication links (Q6058821): Difference between revisions

The considered recursive parameter estimation in a multi-agent network is based on a simple idea: if agents measure a parameter, communicate their results to neighbours and then adjust their measuring towards the average of received results (this adjustment is called innovation), then after several repetitions differences in measurements will decrease. The algorithm proposed in this paper improves several earlier ones presented, e.g. in [\textit{J. Wang} et al., ``Decentralized cooperative online estimation with random observation matrices, communication graphs and time delays'', Preprint, \url{arXiv:1908.08245}; IEEE Trans. Inf. Theory 67, No. 6, Part 2, 4035--4059 (2021; Zbl 1475.62171); \textit{A. K. Sahu} et al., ibid. 2, No. 4, 426--441 (2016; \url{doi:10.1109/TSIPN.2016.2618318})] by removing some restrictions and considering time-varying and noisy communication on unbalanced graphs. Formally, let \(\mathcal{G} = \{ \mathcal{V},{\mathcal{A}_\mathcal{G}}\} \) be a directed graph on the set of agents \(\mathcal{V} = \{ 1,2,\dots,N\} \) connected with weights \({\mathcal{A}_\mathcal{G}} = [{a_{ij}} \geqslant 0] \in {\mathcal{R}^{N \times N}}\); \({\mathcal{N}_i} = \{ j\mid {a_{ij}} > 0\} \) is the neighbourhood of agent \(i\). Agents try to estimate an \(n\)-dimensional vector \(\theta \in \Theta \subset {\mathcal{R}^n}\), using discrete-time measurements. The measurement vector of agent \(i\) at time \(k = 0,1,\dots, {m_i} < n\) is \({z_i}(k) = {f_i}(\theta ) + {v_i}(k)\), \(i \in \mathcal{V}\), where \({f_i}\) is the measurement and \({v_i}\) the noise. The agent communicates with its neighbour agents and receives neighbours' estimates corrupted by the communication noise. The setup and the measurement/communication procedure should satisfy several assumptions: \(\Theta \) is a closed convex set with a non-empty interior, the measurement function of every agent is continuously differentiable on this interior, has bounded gradient and is smooth; the communication and the measurement noise are independent; the conditional second-order moments of noises are bounded almost surely and the algorithm gain is a strictly positive converging to 0 real sequence. To handle unbalanced networks (indegree \(\ne\) outdegree in some nodes) is introduced an auxiliary matrix created from the symmetrized graph \(\widehat {\mathcal{G}}\) using a combination of the graph Laplacian and the auxiliary matrix; for convergence (in static case) is required that for every node \(i \in \mathcal{V}\) there exist real symmetric matrices \({W_i} \in {\mathcal{R}^{n \times n}}\) that are positive definite and monotone. In the time-varying case, graph weights and the neighbour sets become functions of (discrete) time, i.e. \({a_{ij}}(k)\), \({\mathcal{N}_i}(k)\) and the algorithm is modified. The algorithm is demonstrated with simulation on a network of five agents where the communication structure is switching over five configurations.