Dynamic resource allocation aided by reinforcement learning

Abstract

A communication system in which DRA control is aided by RL. An example embodiment may control one or more buffer queues populated by downstream and/or upstream data streams. The egress rates of the buffer queues can be dynamically controlled using an RL technique, according to which a learning agent can adaptively change the state-to-action mapping function of the DRA controller while circumventing the RL exploration phase and relying on extrapolation of the already taken actions instead. This feature may result in at least two benefits: (i) cancellation of a performance penalty typically associated with RL exploration; and (ii) faster learning of the environment, as the learning agent can determine the performance metrics of many actions per state in a single occurrence of the state. In an example embodiment, the communication system may be a DSL system, a PON system, or a wireless communication system.