Airspan Networks Inc.

Airspan IP Holdco LLC

Michael David Livingstone

Andrew Logothetis

Stuart Parrott

Qasim Khan

Antenna elements support RF communication over a communication link using frequency channels, RF processing circuits, and configuration circuitry to apply a selected configuration. Each configuration identifies processing circuit to antenna element couplings, and which frequency channel is allocated to each RF processing circuit. A reinforcement learning process is applied to dynamically alter which configuration to apply as a currently selected configuration. The reinforcement learning process maintains a future rewards record having entries that maintain, for an associated combination of link state and configuration, an estimated future rewards indication determined using a discounted rewards mechanism. A selection policy selects a configuration for a current link state, and then a new reward is observed that is dependent on how the selected configuration alters a chosen performance metric for the communication link. The estimated future rewards indication is updated by storing in the associated entry a predicted estimated future rewards indication generated by assuming, when using the discounted rewards mechanism, that all rewards that will be used in future to update the estimated future rewards indication in the associated entry will have the same value as the new reward.

Apparatus and method for configuring a communication link

An apparatus and method are provided for configuring a communication link. Wherein the apparatus has a plurality of antenna elements to support RF communication using a plurality of frequency channels, a plurality of RF processing circuits, and configuration circuitry to apply a selected configuration from a plurality of different configurations, where each configuration identifies which RF processing circuit each antenna element coupled to, and which channel allocated to each RF processing circuit. The configuration circuitry arranged to a reinforcement learning process in order to dynamically alter which of the plurality of different configurations to apply a currently selected configuration. The reinforcement learning process maintaining a future rewards record having a plurality of entries, where each entry maintains, for an associated combination of link state and configuration, an estimated future rewards indication determined using a discounted rewards mechanism. A selection policy is to select a configuration for a current link state, and a new reward is observed is dependent on how the selected configuration alters a chosen performance metric for the communication link. The estimated future rewards indication in the associated entry is then updated in dependence on the new reward.

Growing community of inventors

Berkshire, United Kingdom

Michael David Livingstone