Switching to Learn

Distributed estimation, detection, and learning theory in networks have attracted much attention over the past decades [1], [2], [3], [4], with applications that range from sensor and robotic networks [5], [6], [7], [8], [9] to social and economic networks [10], [11], [12]. In these scenarios, agents in a network need to learn the value of a parameter that they may not be able to infer on their own, but the global spread of information in the network provides them with adequate data to learn the truth collectively. As a result, agents iteratively exchange information with their neighbors. For instance, in distributed sensor and robotic networks, agents use local diffusion to augment their imperfect observations with information from their neighbors and achieve consensus and coordination [13], [14]. Similarly, agents exchange beliefs in social networks to benefit from each other's observations and private information and learn the unknown state of the world [15], [16]. Existing literature on distributed learning focuses mostly on environments where individuals communicate at every round. Of particular relevance to our discussion are a host of algorithms that follow the non-Bayesian learning scheme in Jadbabaie et.