Reinforcement learning for autonomous telecommunications networks
AssigneeCiena Corporation
InventorsEmil Janulewicz, David Côté, Gauravdeep Singh Shami, Olivier Simard, Thomas Triplet
Life of this patent
Estimating the term…
Abstract
Systems and methods include steps of determining a state of a network based on telemetry data; determining a value of a reward associated with the state; determining an action to take on the network to bring the network to a next state that is expected to have a better than or equal to value of the reward; and causing the action to be implemented in the network. The steps can also include continuing the determining steps and the causing step.
Claims18
A non-transitory computer-readable medium comprising instructions that, when executed, cause a processor to perform the steps of:
- determining a state of a network based on telemetry data;
- determining a value of a reward associated with the state;
- determining an action to take on the network to bring the network to a next state that is expected to have a better than or equal to value of the reward, wherein the action is based on Reinforcement Learning and is chosen for one of exploitation to obtain rewards and exploration to determine better actions;
- causing the action to be implemented in the network;
- determining an updated state of the network based on updated telemetry data subsequent to causing the action to be implemented in the network;
- and
- continuing the causing the action to be implemented in the network based on the updated state of the network;
- wherein the determining the state and the updated state of the network are both performed at a predetermined interval that is one of a regular interval and a varying frequency interval.
A method comprising steps of:
- determining a state of a network based on telemetry data;
- determining a value of a reward associated with the state;
- determining an action to take on the network to bring the network to a next state that is expected to have a better than or equal to value of the reward, wherein the action is based on Reinforcement Learning and is chosen for one of exploitation to obtain rewards and exploration to determine better actions;
- causing the action to be implemented in the network;
- determining an updated state of the network based on updated telemetry data subsequent to causing the action to be implemented in the network;
- and
- continuing the causing the action to be implemented in the network based on the updated state of the network;
- wherein the determining the state and the updated state of the network are both performed at a predetermined interval that is one of a regular interval and a varying frequency interval.
Dependencies are inferred from the claim text.