Robust Learning Over Multitask Adaptive Networks With Wireless Communication Links

Recently, adaptation, optimization and learning over multitask networks with ideal communication links have been investigated in the literature. However, in real-life applications, the communication links between the nodes of the networks are usually non-ideal. In this brief, we consider that the communication links between the nodes of the network are of the wireless type including both block fading and additive noise. In the case of multitask networks with wireless links, non-smart cooperation between the nodes leads to a degraded learning performance that is worse than the noncooperative mode. In this brief, we propose a smart cooperation policy based on an information theoretic criterion, so-called correntropy, that allows the nodes with a similar task and with high quality links to cooperate with each other and rejects the cooperation between the nodes with dissimilar tasks or with low-quality links. The theoretical learning behaviors of the proposed algorithm in the mean sense and in the mean-square sense are also derived. Finally, the computer experiments are provided to verify the theoretical findings.