An adaptive shortest-path on-line routing algorithm

We describe a new adaptive routing algorithm for meshed-topology deflection networks. This algorithm is based on a local learning method which evolves in order to produce a local spatial representation of the traffic. We prove that we can set the parameters of the learning algorithm such that our adaptive policy is a shortest path routing. Then we show experimentally the efficiency of our algorithm. First, we compare the routing policies in a grid network, under an uniform load. Second, we create local congestion in order to show that the adaptive routing scheme avoid the overloaded region. Moreover, we propose a more realistic traffic model, and show that our algorithm is valid, even in such context. These results show the relevance of this method