Distributed learning algorithms for data network routing problem: models, convergence analysis and optimality

摘要

The behavior of the automata at the nodes of a data network is studied for an abstract network representation in which only very general functional properties are assumed. A model of a nonstationary environment is proposed with state variables as penalty parameters. The limiting behavior of the model is studied. Simulation results shown that under abnormal conditions (i.e. change of topology) the learning algorithms outperformed existing routing algorithms. Using a minimal amount of feedback, the equalizing properties of automata in equilibrium can still be used to produce optimal or nearly optimal routing.