Training and using a neural network for managing an environment in a communication network

1 . A computer implemented method for training a Student Neural Network having a plurality of parameters, wherein the Student Neural Network is for generating an action prediction matrix for an environment in a communication network, the environment comprising a plurality of nodes corresponding to entities in the communication network and a plurality of communication network resources, wherein the action prediction matrix comprises a plurality of action predictions for actions that may be executed within the environment, and wherein each action prediction of the action prediction matrix relates to at least one of a node or a resource in the environment, the method, performed by a training node, comprising:

using a Reinforcement Learning process to train a Teacher Neural Network to generate an action prediction for a resource or node in the environment;

using the trained Teacher Neural Network to generate a first training data set by, for each of a plurality of training situations:

generating a representation of a state of the environment, the state representation comprising a current version of a partial action matrix that comprises actions for execution within the environment, each action relating to at least one of a node or a resource in the environment; and

sequentially, for each node or each resource in the environment:

using the trained Teacher Neural Network to generate an action prediction relating to the node or resource based on the state representation;

adding the state representation and the generated action prediction to the first training data set;

selecting an action for execution on the node or the resource in accordance with the action prediction; and

updating the state representation to include the selected action in the partial action matrix;

the method further comprising:

generating a second training data set from the first training data set by, for each of the plurality of training situations:

generating a training situation state representation comprising values that are common to all of the state representations for the training situation in the first training data set; and

generating a training situation action prediction matrix comprising all of the generated action predictions for the training situation in the first training data set; and

using the second training data set to update values of the parameters of the Student Neural Network;

wherein using a Reinforcement Learning process to train the Teacher Neural Network to generate an action prediction for a resource or node in the environment comprises:

sequentially, for each node or each resource in the environment:

generating a representation of a state of the environment;

using the teacher Neural Network to generate an action prediction relating to the node or resource based on the state representation for the environment;

using a selection function to select an action for execution on the node or the resource in accordance with the action prediction, the selected action being any configuration or resource allocation related action applicable to the node or the resource; and

following execution of the selected action, updating parameters of the Teacher Neural Network based on an obtained measure of success of the executed action.

2 . The method as claimed in claim 1 , wherein the Teacher Neural Network is configured to:

receive as input the representation of a state of the environment, the state representation comprising a current version of a partial action matrix that comprises a plurality of actions for execution within the environment, each action relating to at least one of a node or a resource in the environment,

process the input in accordance with current values of the Teacher Neural Network parameters; and

output an action prediction relating to a node or a resource.

3 . The method as claimed in claim 1 , wherein the Student Neural Network is configured to:

receive as input a representation of a state of the environment;

process the input in accordance with current values of the Student Neural Network parameters; and

output an action prediction matrix comprising a plurality of action predictions for actions that may be executed within the environment, each action prediction of the action prediction matrix relating to at least one of a node or a resource in the environment.

4 . The method as claimed in claim 1 , wherein an action prediction relating to a node or resource comprises an action prediction vector, each element of the action prediction vector corresponding to a possible action for the node or resource, and comprising a probability that the corresponding action is the most favorable of the possible actions according to a success measure for the environment.

5 . The method as claimed in claim 1 , wherein the action prediction matrix comprises a plurality of action prediction vectors for actions that may be executed within the environment, each action prediction vector relating to a node or resource in the environment;

wherein each element of an action prediction vector corresponds to a possible action for the node or resource to which the action prediction vector relates, and comprises a probability that the corresponding action is the most favorable of the possible actions according to a success measure for the environment.

6 . The method as claimed in claim 1 , wherein a partial action matrix comprises a matrix representation of:

nodes or resources in the environment;

possible actions that may be executed on the nodes or resources; and

for each node or resource for which an action has been selected, an indication of the selected action.

7 . The method as claimed in claim 4 , wherein selecting an action for execution on the node or the resource in accordance with the action prediction comprises selecting the action on the basis of probabilities in the action prediction vector.

8 . The method as claimed in claim 1 , wherein the Teacher Neural Network is configured to output a Teacher Neural Network success prediction comprising a predicted value of a success measure for a current state of the environment, the Teacher Neural Network success prediction corresponding to selection of an action for the node or resource based on the action prediction output by the Teacher Neural Network.

9 . The method as claimed in claim 1 , wherein the Student Neural Network is configured to output a Student neural Network success prediction comprising a predicted value of a success measure for a current state of the environment, the Student Neural Network success prediction corresponding to selection of actions for nodes or resources based on the action prediction matrix output by the Student Neural Network.

10 . The method as claimed in claim 9 , wherein the success measure comprises a representation of at least one performance parameter for environment.

11 . The method as claimed in claim 1 , wherein generating the second training data set from the first training data set further comprises, for each of the plurality of training situations:

including, in the second training data set for the training situation, a final value of the Teacher Neural Network success prediction output by the Teacher Neural Network for the training situation during generation of the first training data set.

12 . The method as claimed in claim 1 , wherein generating a training situation state representation comprising values that are common to all of the state representations for a training situation in the first training data set comprises:

obtaining a final state representation for the training situation during generation of the first training data set; and

removing the partial action matrix from the obtained state representation.

13 . The method as claimed in claim 1 , wherein using the second training data set to update the values of the Student Neural Network comprises, for training situations in the second training data set:

inputting the training situation state representation to the Student Neural Network, wherein the Student Neural Network processes the training situation state representation in accordance with current values of parameters of the Student Neural Network and outputs an action prediction matrix; and

updating the values of the Student Neural Network parameters so as to minimize a loss function based on a difference between the action prediction matrix output by the Student Neural Network and the training situation action prediction matrix included in the second training data set.

14 . The method as claimed in claim 13 , wherein the Student Neural Network also outputs a Student Neural Network success prediction, and wherein the loss function is further based on a difference between the Student Neural Network success prediction output by the Student Neural Network and the Teacher Neural Network success prediction included the second training data set for the training situation.

15 . The method as claimed in claim 1 , wherein using a Reinforcement Learning process to train the Teacher Neural Network to generate an action prediction for a resource or node in the environment comprises:

generating a representation of a state of the environment, the state representation comprising a current version of a partial action matrix that comprises actions for execution within the environment, each action relating to at least one of a node or a resource in the environment; and

sequentially, for each node or each resource in the environment:

performing a look ahead search of possible future states of the environment according to possible actions for the node or resource, wherein the look ahead search is guided by the Teacher Neural Network in accordance with current values of the Teacher Neural Network parameters and a current version of the state representation of the environment, and wherein the look ahead search outputs a search action prediction and a search success prediction;

adding the current version of the state representation, and the search action prediction and search success prediction output by the look ahead search, to a Teacher Neural Network training data set;

selecting an action for the node or resource in accordance with the search action prediction output by the look ahead search; and

updating the current state representation of the environment to include the selected action in the partial action matrix; and

using the Teacher Neural Network training data set to update the values of the Teacher Neural Network parameters.

16 . The method as claimed in claim 1 , wherein the environment comprises at least one of:

a cell of the communication network;

a sector of a cell of the communication network;

a group of cells of the communication network.

17 . The method as claimed in claim 1 , wherein the representation of a current state of the environment comprises at least one of:

a value of a network coverage parameter;

a value of a network capacity parameter;

a value of a network congestion parameter;

a current network resource allocation;

a current network resource configuration;

a current network usage parameter;

a current network parameter of a neighbor communication network cell;

a value of a network signal quality parameter;

a value of a network signal interference parameter;

a value of a network power parameter;

a current network frequency band;

a current network antenna down-tilt angle;

a current network antenna vertical beamwidth;

a current network antenna horizontal beamwidth;

a current network antenna height;

a current network geolocation;

a current network inter-site distance.

18 . A system comprising a training node for training a Student Neural Network having a plurality of parameters, wherein the Student Neural Network is for generating an action prediction matrix for an environment in a communication network, the environment comprising a plurality of nodes corresponding to entities in the communication network and a plurality of communication network resources, wherein the action prediction matrix comprises a plurality of action predictions for actions that may be executed within the environment, and wherein each action prediction of the action prediction matrix relates to at least one of a node or a resource in the environment, the training node comprising processing circuitry configured to cause the training node to:

use a Reinforcement Learning process to train a Teacher Neural Network to generate an action prediction for a resource or node in the environment;

use the trained Teacher Neural Network to generate a first training data set by, for each of a plurality of training situations:

generating a representation of a state of the environment, the state representation comprising a current version of a partial action matrix that comprises actions for execution within the environment, each action relating to at least one of a node or a resource in the environment; and

sequentially, for each node or each resource in the environment:

using the trained Teacher Neural Network to generate an action prediction relating to the node or resource based on the state representation;

adding the state representation and the generated action prediction to the first training data set;

selecting an action for execution on the node or the resource in accordance with the action prediction; and

updating the state representation to include the selected action in the partial action matrix;

the processing circuitry being further configured to cause the training node to:

generate a second training data set from the first training data set by, for each of the plurality of training situations:

generating a training situation state representation comprising values that are common to all of the state representations for the training situation in the first training data set; and

generating a training situation action prediction matrix comprising all of the generated action predictions for the training situation in the first training data set; and

use the second training data set to update values of the parameters of the Student Neural Network;

wherein using a Reinforcement Learning process to train the Teacher Neural Network to generate an action prediction for a resource or node in the environment comprises:

sequentially, for each node or each resource in the environment:

generating a representation of a state of the environment;

using the teacher Neural Network to generate an action prediction relating to the node or resource based on the state representation for the environment;

using a selection function to select an action for execution on the node or the resource in accordance with the action prediction, the selected action being any configuration or resource allocation related action applicable to the node or the resource; and

following execution of the selected action, updating parameters of the Teacher Neural Network based on an obtained measure of success of the executed action.