Apparatus and method for designing a grid-of-beams using machine learning

1. An apparatus, comprising:

at least one processor; and

at least one non-transitory memory comprising computer program code,

the at least one memory and computer program code configured, with the at least one processor, to cause the apparatus at least to:

collect network data for training a neural network;

train the neural network, using the collected data, to learn a non-discounted cumulative reward that evaluates a benefit of including a given beam into a grid-of-beams;

apply the trained neural network to select at least one beam to include in the grid-of-beams to optimize a performance metric, wherein the trained neural network is repeatedly applied until a number of beams for the grid-of-beams is selected; and

select one or more beams from the grid-of-beams to transmit to a user equipment or to receive transmission from the user equipment.

2. The apparatus of according to claim 1 , wherein the collected data comprises at least one of:

simulation data collected from statistical models or from ray tracing data for a number of realistic deployments; and

on-line data collected from measurements over a predefined grid-of-beams (GoB).

3. The apparatus of claim 1 , wherein the at least one memory and computer program code are configured, with the at least one processor, to cause the apparatus at least to:

approximate the non-discounted cumulative reward (Q) as {circumflex over (Q)}(v;w) based at least on a set of features within a feature vector, wherein v is the feature vector and w is a weight vector, wherein the feature vector v comprises at least one of:

a number of discrete fourier transform (DFT) beams of a subset of beams of a set of potential beams;

a number of merged beams of the subset of beams;

a number of tapered beams of the subset of beams;

beam pointing directions of the beams of the subset of beams;

beamwidths of the beams of the subset of beams; and

a beamforming gain of the beams of the subset of beams.

4. The apparatus of claim 3 , wherein the at least one memory and computer program code are configured, with the at least one processor, to cause the apparatus at least to:

iteratively apply the training of the neural network based on a change to either the weight vector w or the approximate {circumflex over (Q)} function.

5. The apparatus of according to claim 3 , wherein the training of the neural network comprises:

initializing the weight vector w and the non-discounted cumulative reward (Q) for states and actions; and

selecting an (s,a) pair and update w based on at least a reward given the (s,a) pair and a gradient of the approximate non-discounted cumulative reward {circumflex over (Q)}; and

repeating the selecting of the (s,a) pair and updating of w based on a change to either the w vectors or the approximate non-discounted cumulative reward {circumflex over (Q)}.

6. The apparatus of claim 1 , wherein the at least one memory and computer program code are configured, with the at least one processor, to cause the apparatus at least to:

select the at least one beam to include in the grid-of-beams (GoB), based on the non-discounted cumulative reward (Q) value learned from training the neural network.

7. The apparatus of claim 1 ,

wherein the performance metric comprises a traffic-density weighted beamforming gain, and

wherein the applying comprises applying the trained neural network to select the at least one beam to include in the grid-of-beams (GoB) based on the traffic-density weighted beamforming gain.

8. The apparatus of claim 1 ,

wherein the performance metric comprises coverage optimization, and

wherein the applying comprises applying the trained neural network to select the at least one beam to include in the grid-of-beams (GoB) based on a reference signal received power (RSRP).

9. The apparatus of claim 1 ,

wherein the performance metric comprises a geometric mean of beamforming gains, and

wherein the applying comprises applying the trained neural network to select the at least one beam to include in the grid-of-beams (GoB) based on the geometric mean of the beamforming gains.

10. The apparatus of claim 1 , wherein the neural network comprises at least one of a convolutional neural network or deep neural network.

11. A method, comprising:

collecting network data for training a neural network;

training the neural network, using the collected data, to learn a non-discounted cumulative reward that evaluates a benefit of including a given beam into a grid-of-beams;

applying the trained neural network to select at least one beam to include in the grid-of-beams to optimize a performance metric, wherein the trained neural network is repeatedly applied until a number of beams for the grid-of-beams is selected; and

selecting one or more beams from the grid-of-beams to transmit to a user equipment or to receive transmission from the user equipment.

12. The method of claim 11 , wherein the collected data comprises at least one of:

simulation data collected from statistical models or from ray tracing data for a number of realistic deployments; and

on-line data collected from measurements over a predefined grid-of-beams (GoB).

13. The method of claim 11 , wherein the method further comprises:

approximating the non-discounted cumulative reward (Q) as {circumflex over (Q)}(v;w) based at least on a set of features within a feature vector, wherein v is the feature vector and w is a weight vector, wherein the feature vector v comprises at least one of:

a number of discrete fourier transform (DFT) beams of a subset of beams of a set of potential beams;

a number of merged beams of the subset of beams;

a number of tapered beams of the subset of beams;

beam pointing directions of the beams of the subset of beams;

beamwidths of the beams of the subset of beams; and

a beamforming gain of the beams of the subset of beams.

14. The method of claim 13 , wherein the method further comprises:

iteratively applying the training of the neural network based on a change to either the weight vector w or the approximate {circumflex over (Q)} function.

15. The method of according to claim 13 , wherein the training of the neural network comprises:

initializing the weight vector w and the non-discounted cumulative reward (Q) for states and actions; and

selecting an (s,a) pair and update w based on at least a reward given the (s,a) pair and a gradient of the approximate non-discounted cumulative reward {circumflex over (Q)}; and

repeating the selecting of the (s,a) pair and updating of w based on a change to either the w vectors or the approximate non-discounted cumulative reward {circumflex over (Q)}.

16. The method of claim 11 , wherein the method further comprises:

selecting the at least one beam to include in the grid-of-beams (GoB), based on the non-discounted cumulative reward (Q) value learned from training the neural network.

17. The method of claim 11 ,

wherein the performance metric comprises a traffic-density weighted beamforming gain, and

wherein the applying comprises applying the trained neural network to select the at least one beam to include in the grid-of-beams (GoB) based on the traffic-density weighted beamforming gain.

18. The method of claim 11 ,

wherein the performance metric comprises coverage optimization, and

wherein the applying comprises applying the trained neural network to select the at least one beam to include in the grid-of-beams (GoB) based on a reference signal received power (RSRP).

19. The method of claim 11 ,

wherein the performance metric comprises a geometric mean of beamforming gains, and

wherein the applying comprises applying the trained neural network to select the at least one beam to include in the grid-of-beams (GoB) based on the geometric mean of the beamforming gains.

20. A non-transitory computer readable medium comprising program instructions, wherein the instructions cause an apparatus to perform at least:

collecting network data for training a neural network;

training the neural network, using the collected data, to learn a non-discounted cumulative reward that evaluates a benefit of including a given beam into a grid-of-beams;

applying the trained neural network to select at least one beam to include in the grid-of-beams to optimize a performance metric, wherein the trained neural network is repeatedly applied until a number of beams for the grid-of-beams is selected; and

selecting one or more beams from the grid-of-beams to transmit to a user equipment or to receive transmission from the user equipment.

21. A user equipment apparatus, comprising:

at least one processor; and

at least one non-transitory memory comprising computer program code,

the at least one memory and computer program code configured, with the at least one processor, to cause the apparatus at least to:

collect network data for training a neural network;

train the neural network, using the collected data, to learn a non-discounted cumulative reward that evaluates a benefit of including a given beam into a grid-of-beams;

apply the trained neural network to select at least one beam to include in the grid-of-beams to optimize a performance metric, wherein the trained neural network is repeatedly applied until a number of beams for the grid-of-beams is selected; and

select one or more beams from the grid-of-beams to transmit to a network node or to receive transmission from the network node.