Current wireless channel capacities are closely approaching the theoretical limit. Hence, further capacity improvements from complex signal processing schemes may only gain modest improvements. Multi-path communication approaches, however, combine the benefits of higher performance and reliability by exploiting the concurrent usage of multiple communication networks in areas that are covered by a multitude of wireless access networks. So far, little is known on how to effectively take advantage of this potential. Motivated by this, we consider parallel communication networks that handle two types of traffic: foreground and background. The foreground traffic stream of files should be directed to the network that requires the least time to transfer the file. The background streams are always directed to the same network. It is not clear up front how to select the appropriate network for each foreground stream. This may be performed by a static selection policy, based on the expected load of the networks. However, a dynamic policy that accounts for the network status may prove more elegant and better performing. We first propose a dynamic model that optimally assigns the foreground traffic to the available networks based on the number of fore- and background streams in both networks. However, in practice all traffic streams may be served by one application server. Thus, it may not be feasible to distinguish foreground from background traffic streams. This limitation is accounted for in our second, partial observation model that considers limited observability for dynamic network selection. We compare these static and dynamic models to each other and to the well-known Join the Shortest Queue (JSQ) model. The results are illustrated by extensive numerical experiments.