Abstract: A computer implemented method for configuring a reinforcement learning agent to perform an efficient reinforcement learning procedure, wherein the reinforcement learning agent comprises a model trained using a machine learning process to determine actions to be performed by the reinforcement learning agent. The method comprises using the model to determine an action to perform, based on values of a set of features obtained in an environment; determining, for a first feature in the set of features, a first indication of a relative contribution of the first feature, compared to other features in the set of features, to the determination of the action by the model; and determining a reward to be given to the reinforcement learning agent in response to performing the action, based on the first feature and the first indication.

Abstract: A method of operating a device that interacts with a physical environment includes determining to take an action on the physical environment wherein the action is based on an output of a computing task, generating a risk level that estimates a risk of physical harm associated with taking the action within the physical environment, obtaining a performance indicator of a communication network between the device and a remote computing device, and determining, based on the risk level and the performance indicator, whether to offload the computing task to the remote computing device or to perform the computing task locally. Related devices are also disclosed.

Abstract: Methods and apparatus for implementing reinforcement learning are provided. A method in a client node that instructs actions in an environment in accordance with a policy includes identifying one or more critical states of the environment for which a current policy provides unreliable actions. The method further includes initiating transmission to a server of a retraining request, the retraining request having information relating to the one or more critical states. The method further includes receiving a new policy from the server, wherein the new policy is generated by the server using reinforcement learning based on the information relating to the one or more critical states, and instructing actions in the environment in accordance with the new policy.

Abstract: A method performed by a first network node in a communication network is provided. The method includes determining, from a machine learning model at the first network node, a value for a network parameter for a cellular network for operation of a communication device based on a set of observations of the communication device, a safety level of the communication device, and a key performance indicator, KPI, of the cellular network. The method further includes signaling the value of the network parameter to a core network node in the cellular network for a resource allocation of the cellular network. A method performed by a core network node in a cellular network is also provided.

Abstract: A management node for use with a cognitive layer (CL) having agents. The agents have respective agent information indicating type(s) of input data required by a model implemented by the agent, parameter(s) of the system that are to be improved by the agent, type(s) of output data provided by the model and a data distribution for the agent. A method includes (i) selecting a set of similar agents that improve a first parameter of the system: (ii) for a first agent and a second agent in the selected set of similar agents, comparing the data distribution for the first agent to the data distribution for the second agent to determine a relationship: (iii) initiating generation of the candidate agent(s) based on the relationship; and (iv) determining whether to replace one or both of the first agent and the second agent with the one or more candidate agents.

Abstract: In one aspect, a method of determining a risk of conflict between a movable device and potential obstacles is provided. The method includes dividing a space into a plurality of positions. At each of a plurality of successive times, the method further includes: determining a conflict function for the movable device in a first position at a respective time, the first position having one or more neighbouring positions, wherein the conflict function is determined based on whether or not an obstacle is present in any of the first position and the one or more neighbouring positions; and determining a respective risk value for at least one of the first position and the one or more neighbouring positions using the conflict function and a risk value associated with a second position, wherein the movable device is planned to move from the first position to the second position at a subsequent time.

Abstract: A method performed by a risk management node for autonomous devices. The risk management node may determine state parameters from a representation of an environment. The representation of the environment may include an object, an autonomous device, and a set of safety zones. The risk management node may determine a reward value based on evaluating a risk of a hazard with the object based on the determined state parameters and current location and speed of the autonomous device relative to a safety zone from the set of safety zones. The risk management node may determine a control parameter based on the determined reward value, and may initiate sending the control parameter to the autonomous device to control action of the autonomous device. The control parameter may be dynamically adapted to reduce the risk of hazard with the object based on reinforcement learning feedback from the reward value.

Abstract: A Cognitive Manager (CM) controls a RAN. The CM obtains from the RAN a set of local needs identifying a radio communication performance criterion that is to be satisfied. A set of neighbor needs are obtained from a neighbor CM. An event message is received from the RAN containing a radio metric that does not presently satisfy at least one of the set of local needs. A set of present neighbor improvement action drives being performed or will be performed to satisfy the set of neighbor needs, are obtained from the neighbor CM. A combination of the set of local needs, the set of the neighbor needs, the set of present neighbor improvement action drives, and the radio metric, is used to determine a local improvement action drive that is to be performed by the RAN. An adjustment report identifying the local improvement action drive is provided to the RAN.

Abstract: A method, performed by a first network node, for handling a maintenance of one or more second network nodes is described herein. The first network node and the one or more second network nodes operate in a communications network. The first network node obtains, respectively, from each of one or more third network nodes operating in the communications network, and for a respective second network node of the one or more second network nodes, one or more predictive models for each of: a) a performance of the respective second network node, and b) a traffic load of the respective second network node. The first network node then determines one or more plans to maintain the one or more second network nodes, based on the obtained one or more predictive models.