Communication networks for smart grids may consist of a mixture of legacy and new links using heterogeneous technologies, such as copper wires, optical fibers, wireless and powerline communication. If nodes are connected by two or more links, such as wireless and powerline, the sender of a message must decide on which link to transmit the next message. This paper considers the problem of dynamically selecting the link, based on success/failure (acknowledgement) of previous transmissions. The novel method is based on Markov (Gilbert-Elliott) channel models of lossy and time varying links. It specifies how to employ success/failure observations to rank the links optimally, with the objective function to maximize throughput. The theory of partially observable Markov decision problems (POMDP) provides the basic framework. We compare this new method with known linear learning strategies.