Reinforcement learning for admission control and routing

Abstract: When a user requests. a connection to another user or a computer in a communications network, a routing algorithm selects a path for transferring the resulting data stream. If all suitable paths are busy, the user request cannot beserved, and is blocked. A routing algorithm that minimizes this blocking probability results in satisfied users, and maximizes the network operator's revenue. In some cases, it may even be advantageous to block a request from one user, to make it possible to serve other users better. This thesis presents improved and partially new algorithms, based on reinforcement learning, which optimize the way a network is shared. A main contribution of the thesis is the development of algorithms thatadapt to arrivals of user requests that are correlated over time. These methodsare shown to increase network utilization in cases where the request arrivalprocesses are statistically self-similar. Another main contribution is gainscheduled routing, which reduces the computational cost associated withmaking routing decisions. The thesis also demonstrates how to integrate theconcept of max-min fairness into reinforcement learning routing.