Method performed by a network node for training a reinforcement learning algorithm with multiple reward functions to perform a task in a communication network

Abstract

The disclosure provides methods and apparatus for training and executing a reinforcement learning, RL, algorithm configured with multiple reward functions to perform a task in a communication network (1000) The method (100, 500) comprising, the first network node (1100) transmitting (103, 503) a first message, to the second network node (1200), comprising parameters of the RL algorithm. The first network node (1100) receiving (104, 507) a second message, from the second network node (1200), comprising one or more observations. The first network node (1100) updating (105, 508) the parameters of the RL algorithm based on the received one or more observation. The first network node (1100) performing (106, 509) the transmitting, the receiving and the updating iteratively until a stopping criterion exceeds a threshold and storing (107, 510) the updated parameters of the RL algorithm.

Description

[0001] REINFORCEMENT LEARNING WITH MULTIPLE REWARD FUNCTIONSTECHNICAL FIELDThe disclosure relates to methods performed by a first network node and a secondnetwork node, computer programs and computer program products comprising thecomputer programs for training and executing a reinforcement learning, RL, algorithmconfigured with multiple reward functions to perform a task in a communicationnetwork. BACKGROUNDIn telecommunication, reinforcement learning, RL, optimizes various aspects ofnetwork management, from resource allocation to traffic management and beyond.RL is a fundamental sequential decision-making framework. At its core, RL is basedon the concept of an agent learning by direct interaction with an environment. This process is typically formalized as a Markov Decision Process (MDP), which consists of states, actions, transition probabilities, and rewards.In simple terms, states represent configurations of the environment, and encapsulateall relevant information necessary for decision-making; actions denote the possibledecisions that the agent can take while in each state; transition probabilities capture the environment's dynamics, specifying the probability of transitioning from one stateto another upon taking a particular action; rewards serve as numerical feedbacksignals provided by the environment to the agent. In other words, rewards quantify the desirability of the agent's actions, guiding its decision-making process towards achieving predefined objectives.Mathematically, an RL problem is specified by an MDP ^, which consists of(i) a state space ^, representing the set of all possible states that theenvironment can be in; (ii) an action space ^, comprising of the set of all possible actions thatthe agent can take; (iii) a transition probability ^, capturing the dynamics of the environmentand defining the probability of transitioning from one state to another given a specific action, i.e., ^(^^^^|^^, ^^) represents the probability oftransitioning to a state ^^^^ at time step ^ + 1, given that the agenttakes action ^^in state ^^; (iv) a reward function ^, quantifying the immediate feedback the agentreceives from the environment upon taking an action in a particular state. Mathematically, ^^ = ^(^^ , ^^ , ^^^^) represents the rewardobtained when transitioning from state ^^to state ^^^^after performing action ^^;(v) the discount factor ^ ∈ [0,1] accounts for the relative importance offuture rewards compared to immediate rewards. In RL, a proper choice of the reward function is fundamental to attain the desiredobjective. The performance of a deterministic Markovian policy ^: ^ → ^ that mapsstates to actions is measured by the value function, defined as ^^ = is the ^ function of ^. Using thisobjective one can also define the optimal value as ^⋆(^) = m^ax ^^(^) and the optimal^-value ^⋆(^, ^) = max ^^^(^, ^). Therefore, the optimal policy ^⋆ can be defined as^⋆(^) = ^^^^^^ ^⋆^ (^, ^). The effectiveness of RL algorithms and the behaviour of theoptimal policy heavily relies on a careful selection and design of the reward functions. Policies are tailored to exploit the specific structure encoded within the reward function. When an RL algorithm is trained with a given reward and such reward function undergoes a change, the implications for the policies trained with the previous reward function is severe, yielding a degradation in performance. This fact stems from the tight coupling between the reward function and the learned policies, where policies are optimized to maximize the rewards provided by the specific reward function they were trained on. Adapting policies to the updated reward function typically requires re-training the RL agent from scratch or fine-tuning existing policies. This incurs additional computationalcosts and time overhead, especially in complex environments or with large-scale RLsystems. Reward-free RL [1,2] is a variant of the classical RL setting consisting of two phases: (i) in the first phase (a.k.a. exploration) the learner collects samples from an unknown environment without knowing the reward function; (ii) in the second phase (a.k.a.planning or inference) a reward function is revealed, and the objective is to computea corresponding (approximately) optimal policy for such reward. This setting diverges from the classical RL paradigm and highlights the role of exploration. In fact, now the learner needs to account for all possible rewards since it is not known a-priori which state-action pairs will lead to high rewards in the planning phase. In such a setting, the exploration strategy is in general dictated by the hardestto reach state in the environment.Another setting, unsupervised RL [8, 9, 10], similar to the reward-free setting consists of two phases: 1. the first phase is like the first phase of reward-free RL, where a model is trainedwithout knowledge of the reward function. 2. in the second phase we use a reward function to fine tune the model obtainedfrom the first phase. Also, in this setting the role of exploration is essential in the first phase. Most of the work in the RL literature has focused on the single reward function paradigm. These algorithms have several shortcomings: Firstly, they can only beexecuted with a specific reward function used during training. This prevents suchalgorithms to optimize, at runtime, to different reward functions (different objective andkey performance indicators, KPIs, such as quality of service or network intents in radioaccess networks). Secondly, single-reward function algorithms are not suited to solvethe data collection, or exploration problem, in presence of multiple reward functions.Moreover, due to the specificity of reward function, these algorithms typically aresignificantly sensitive to modifications or misspecifications in reward functions andrequires training the RL agents from scratch for different reward functions, achallenging task in general. SUMMARYAn object of the invention is to enable a reinforcement learning, RL, algorithm toperform a task in a communication network with dynamic criteria for performing thetask.In a first aspect of the invention, a first network node performs a method for training aRL, algorithm to perform a task in a communication network. The RL algorithm beingconfigured with multiple reward functions wherein each reward function of the multiplereward functions is associated with the task. The communication network comprisesthe first network node and a second network node. The method comprises transmittinga first message, to the second network node, comprising parameters of the RL algorithm. The method comprising receiving a second message, from the second network node, comprising one or more observations, wherein each observation of the one or more observations comprise rewards provided by the multiple reward functions.The method comprises updating the parameters of the RL algorithm based on thereceived one or more observations comprising the rewards. The method comprises performing the transmitting, the receiving, and the updating iteratively until a stopping criterion exceeds a threshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm. The method comprises storing the updated parameters of the RL algorithm.In a second aspect of the invention, the second network node performs a method fortraining the RL algorithm to perform the task in the communication network. The RLalgorithm being configured with multiple reward functions wherein each rewardfunction of the multiple reward functions is associated with the task. The communication network comprises the first network node and the second networknode. The method comprises receiving the first message, from the first network node,comprising parameters of the RL algorithm. The method comprises determining basedon the received parameters of the RL algorithm, one or more actions associated with the task. The method comprises executing the determined one or more actions associated with the task. The method comprises obtaining one or more observations based on the execution of the determined one or more actions, wherein each observation of the one or more observations comprise rewards provided by themultiple reward functions. The method comprises transmitting the second message,to the first network node, the second message comprising the one or moreobservations. The method comprises performing the receiving, the determining, theexecuting, the obtaining and the transmitting iteratively until the stopping criterion exceeds the threshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm.In a third aspect of the invention, the first network node performs a method forexecuting the RL, algorithm to perform the task in the communication network. The RLalgorithm being configured with multiple reward functions wherein each rewardfunction of the multiple reward functions is associated with the task. Thecommunication network comprises the first network node and the second networknode. The method comprises receiving a request message, from the second networknode, comprising a request for a policy corresponding to each reward function of themultiple reward functions. The method comprises verifying availability of thepolicy corresponding to each reward function of the multiple reward functions. Themethod comprises transmitting a response message, to the second network node,comprising an indication regarding the availability of the policy corresponding to eachreward function of the multiple reward functions.In a fourth aspect of the invention, the second network node performs a method forexecuting the RL algorithm to perform the task in the communication network. The RLalgorithm being configured with multiple reward functions wherein each reward function of the multiple reward functions is associated with the task. Thecommunication network comprises the first network node and the second networknode. The method comprises transmitting the request message, to the first networknode, comprising the request for the policy corresponding to each reward function ofthe multiple reward functions. The method comprises receiving the responsemessage, from the first network node, comprising the indication regarding availabilityof the policy corresponding to each reward function of the multiple reward functions. According to a fifth aspect of the invention, there is presented a computer programcomprising instructions which when executed on a processor of the first network node,causes the first network node to perform a method according to any of theembodiments of the first aspect or the third aspect.According to a sixth aspect of the invention, there is presented a computer programproduct which comprises a computer readable storage medium on which a computerprogram according to the fifth aspect is stored.According to a seventh aspect of the invention, there is presented a computer programcomprising instructions which when executed on a processor of the second networknode, causes the second network node to perform a method according to any of theembodiments of the second aspect or the fourth aspectOther objectives, features and advantages of the enclosed embodiments will be apparent from the following detailed disclosure, from the attached dependent claims as well as from the drawings. Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a / an / the element, apparatus, component, means, module, action, etc." are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, module, action, etc., unless explicitly stated otherwise. The actions of any method disclosed herein. BRIEF DESCRIPTION OF THE DRAWINGSA more complete understanding of the present embodiments, and the attendant advantagesand features thereof, will be more readily understood by reference to the following detaileddescription when considered in conjunction with the accompanying drawings wherein:Figure 1 illustrates a method performed by a first network node for training a reinforcement learning, RL, algorithm to perform a task in a communication network. Figure 2 illustrates a method performed by the second network node for training the RL algorithm to perform a task in the communication network. Figure 3 depicts a method performed by the first network node for executing the RL algorithm to perform a task in the communication network. Figure 4 depicts a method performed by the second network node for executing the RL algorithm to perform a task in the communication network. Figure 5 refers to a communication between the first network node and the second network node for training the RL algorithm. Figure 6 refers to a communication between the first network node and the second network node for executing the RL algorithm. Figure 7 refers to Multiple-Rewards Navigate and Stop, MR-NaS, training scheme. Figure 8 refers to Bootstrapped Multiple Rewards Best Policy Identification, DBMR-BPI, training scheme.Figure 9 is a block diagram illustrating an example first network node, an example secondnetwork node and an example communication network.Figure 10 depicts performance of the various algorithms in terms of (1) throughput, (2) BLER(block error rate) and (3) spectral efficiency.Figure 11 depicts the distribution of the number of transmissions attempted per packet andthe action selection for each algorithm.Figure 12 illustrates a distribution of normalized rewards collected during training by thealgorithms.Figure 13 depicts the estimated sub-optimality gap for each reward, and the number of timesa certain reward was chosen for exploration by DBMR-BPI. DETAILED DESCRIPTIONRL offers a powerful framework to learn control strategies in communication networksystems. When optimizing these systems, the ability to learn optimal policy for specificclasses of reward functions is a critical feature in use-cases such as scheduling,network selection, resource allocations, etc., and ensures good performance across awide range of reward functions. Finding algorithms that can adapt quickly to differenttypes of reward function is a problem of great significance.Aspects of the disclosure address Best Policy Identification, BPI, problem for multiplereward functions. In classical BPI, the goal is to learn the best policy for a given rewardfunction as quickly as possible by interacting with an environment. Aspects of thedisclosure provide a solution for different network optimization applications that requirea given degree of flexibility in defining a reward function. An important problem solvedby the proposed methods is offering differentiated services in radio access networks,RAN, where a network operator may optimize the performance of user devices or localnetwork areas by different reward functions, where each reward function may be afunction of one or multiple Key Performance Indicators, KPIs. Examples of the disclosure propose to solve the BPI problem in multiple-rewardfunctions setting for both Tabular (MDPs with a finite number of states and actions)and Deep Reinforcement Learning (Deep RL, for MDPs with a continuous statespace). This represents a major technical improvement with respect to the state-of-the-art BPI algorithms that consider RL applications with a fixed reward function.Additionally, the embodiments of the disclosure propose methods to efficiently train, ina communication network, an RL algorithm that can produce different estimatedpolicies as function of multiple reward functions. The invention further disclosessignalling methods executed by a first network node and a second network node inthe communication network for training as well as executing the RL algorithmconfigured with multiple reward functions.An actor-learner split architecture is disclosed, wherein the first network node (thelearner) and the second network node (the actor) coordinate to train the RL algorithmwith multiple reward functions in the communication network. Furthermore, theinvention discloses a method for the second network node to request a policy to thefirst network node based on at least one reward function. Also, the effectiveness of theproposed algorithms is demonstrated using downlink link adaptation task.Figure 1 illustrates a method 100 performed by a first network node 1100 (not shown;refer Figure 5 or 6), for training a reinforcement learning, RL, algorithm to perform atask in a communication network 1000 (not shown; refer Figure 9), the RL algorithmbeing configured with multiple reward functions wherein each reward function of the multiple reward functions is associated with the task, the communication networkcomprising the first network node 1100 and a second network node 1200 (not shown;refer Figure 5 or 6). The method 100 comprises transmitting 103 a first message to the second networknode 1200. The first message comprises parameters of the RL algorithm.The method 100 comprises receiving 104 a second message, from the secondnetwork node 1200, comprising one or more observations, wherein each observation of the one or more observations comprise rewards provided by the multiple reward functions.The method 100 comprises updating 105 the parameters of the RL algorithm basedon the received one or more observations comprising the rewards.The method 100 comprises performing 106, the transmitting 103, the receiving 104,and the updating 105 iteratively until a stopping criterion exceeds a threshold, whereinthe stopping criterion is based on at least one of the parameters of the RL algorithm.The method 100 comprises storing 107 the updated parameters of the RL algorithm.In an embodiment, the first network 1100 may comprise one or more criteria forperforming the task. In another embodiment, the method 200 comprises the first network node 1100 receiving 101, from the second network node 1200, one or more criteria for performing the task. In an embodiment, the method 200 comprises transforming 102 the one or more criteria into the multiple reward functions.Figure 2 illustrates a method performed by the second network node 1200 (not shown;refer Figure 5 or 6), for training the RL algorithm to perform a task in the communicationnetwork 1000 (not shown; refer Figure 9). The method 200 comprises receiving 202 afirst message, from the first network node 1100. The first message comprisingparameters of the RL algorithm. The method 200 comprises determining 203, basedon the received parameters of the RL algorithm, one or more actions associated withthe task. The method 200 comprises executing 204 the determined one or moreactions associated with the task. The method 200 comprises obtaining 205 one ormore observations based on the execution of the determined one or more actions, wherein each observation of the one or more observations comprise rewards providedby the multiple reward functions. The method 200 comprises transmitting 206 asecond message, to the first network node 1100. The second message comprising theone or more observations. The method 200 comprises performing 207 the receiving202, the determining 203, the executing 204, the obtaining 205 and the transmitting206 iteratively until the stopping criterion exceeds a threshold, wherein the stoppingcriterion is based on at least one of the parameters of the RL algorithm. In an embodiment, the method 200 comprises transmitting 201 to the first network node 1100, one or more criteria for performing the task.In an embodiment, the task may comprise a radio access network, RAN, parametercontrol task such as antenna tilt optimization, downlink link adaptation, etc. In an embodiment, the criteria for performing the task may comprise Reference Signal Received Power, RSRP. In an embodiment, the criteria for performing the task may comprise Signal to Noise plus Interference Ratio, SINR. In an embodiment, the criteria for performing the task may comprise Transport Block Size, TBS. In an embodiment, the criteria for performing the task may comprise spectral efficiency. In an embodiment, the criteria for performing the task may comprise a number of transmissions.In an embodiment one or more criteria is referred to as communication network intents.In intent-based operations for communication networks, the term “intent” refers to objectives, requirements, and constraints for a service or network operation; see, for example, 3GPP TS 28.312. Example of intents may be user-specific quality of service (QoS) parameters, such as the 5G QoS Identifier, 5QI, parameters of a 3GPP NG- RAN system, defining service requirements in terms of reliability, latency, throughput etc. In an embodiment, the criteria for performing the task may comprise a combination of the criteria described above. In an embodiment, the parameters of the RL algorithm comprise a sampling rule foraction selection ^^, wherein ^ indicates a value of an iteration or timestep in the RLalgorithm training. In an embodiment, the parameters of the RL algorithm comprise an average cumulative allocation ^∗^. In an embodiment, the parameters of the RLalgorithm comprise a first policy ^^∗. In an embodiment, the parameters of the RLalgorithm comprise a second policy ^^,^. In an embodiment, the parameters of the RLalgorithm comprise a policy weighting factor ^^. In an embodiment, the parameters ofthe RL algorithm comprise an indication of number of actions for the second networknode 1100 to perform based on the sampling rule for action selection ^^. The indicationof number of actions for the second network node 1200 to perform based on thesampling rule for action selection ^^comprises a minimum or a maximum number ofactions for the second network node 1200 to perform based on the sampling rule foraction selection ^^.Each observation in the one or more observations comprises an action ^^ executed atthe ^^^ timestep or iteration of the RL algorithm training. Each observation in the oneor more observations comprises a set of rewards ^^ provided for the execution of theaction ^^, wherein each reward in the set of rewards ^^ is provided by one reward function of the multiple reward functions. In an embodiment, each observation in theone or more observations comprises a new state ^^^^ observed upon executing theaction ^^. An observation is denoted by wherein ρ indicates atotal number of reward functions.The sampling rule for action selection denotes an exploration strategy of the RLalgorithm. The goal of the RL algorithm training is to update the sampling rule for actionselection ^^ in order to output a policy ^^⋆ ^,^corresponding to each reward function ofthe multiple reward functions, wherein ^^ ⋆^,^ is the estimated optimal policy for areward function ^ at a stopping time τ (final value of the iteration or timestep), whichdefines when to stop the algorithm training. The process of updating the sampling rulefor action selection ^^ in order to output the policy ^^⋆ ^,^corresponding to each rewardfunction of the multiple reward functions is denoted as BPI for multiple rewardfunctions, MR-BPI.A set of reward functions denoted by ℛ is generally described as a finite set of discretereward functions. The methods 100 and 200 leverage a lower bound on samplecomplexity to design a provably sample efficient algorithm which identifies an optimalpolicy for all reward functions specified in the set of reward functions ℛ in aenvironment with finite number of states and finite number of action. Samplecomplexity refers to a number of training data samples required to learn an optimalpolicy for each reward function so that the probability that one of the policies learnt isnot the optimal one is lower than some value ^. This setting is also known as PC,probably correct framework. A fundamental limit (the lower bound) on samplecomplexity achieved by any algorithm identifying the best policy for all possible rewardfunctions and a pre-specified error probability are calculated in order to perform MR-BPI.The RL algorithm according to methods 100, 200 is exemplified using a discountedMDP ^ = (^, ^, ^, ^), where ^ is a state space, ^ is an action space, ^: ^ × ^ → Δ(^)is a transition function, which maps state-action pairs to a distribution over states Δ(^),and ^ ∈ (0,1) is a discount factor. ℳ denotes a set of MDPs over (^, ^) correspondingto a set of reward functions ℛ and ^^ = (^, ^, ^, ^, ^) denotes an MDP when consideringa reward function ^: ^ × ^ → [0,1], ^ ∈ ℛ. The objective is the same as before, and we define = ^, ^] to be the value of a policy ^ in state ^ withreward ^ (sim. max ^ ^ ^^ (^) isthe optimal value in state ^ with reward ^ (sim. ^⋆^(^, ^) = m^ax^^^ (^, ^)). For a givenreward function ^, Π⋆(^^) denotes a set of optimal policies. For a given reward function^, there can be one or more optimal policies.A reward ^ is selected from a finite set of reward functions ℛ = {r^, r^, … , r^}, ^ > 0,known beforehand. For tabular MDPs, reward functions are represented as vectorswithin [0,1]^×^, with each coordinate ^ corresponding to a reward for the ^-th state-action pair (thus, a reward in each state-action pair is bounded in [0,1]), while forcontinuous-state MDPs, reward functions are represented as bounded functions.Multiple-reward functions Best Policy Identification (MR-BPI)In MR-BPI, the MDP ^ is initially unknown, (precisely, the transition function isunknown), and an objective of the learner node, the first network node 1100, is to learnthe best policy for a reward function ^ as quickly as possible by interacting with anenvironment, such as the communication network 1000. The objective is oftenformalized in a Probably Correct (PC) framework [3,4], where the learner node has tostop interacting with the environment when the learner node can output an optimalpolicy for any reward with a confidence value of at-least 1 − ^.The MR-BPI algorithm is said to be PC with confidence value ^ ∈ (0,1), if the MR-BPIalgorithm returns a best policy for any reward function ^ ∈ ℛ with probability greaterthan 1- δ, or more formally, if, for all MDPs ^, it satisfies^ ^^ < ∞, ^∀^ ∈ ℛ, π^ ⋆^,^ ∈ Π (^^)^^ ≥ 1 − ^,where ^^^,^ is the estimated optimal policy for a reward function ^ at a stopping time τ(final value of the timestep).Lower Bound: To perform MR-BPI, the lower bound of the sample complexity of any^-PC algorithm for finite state-action spaces MDPs needs to computed. To computethe lower bound, an exploration strategy ^ (^, ^) denoting a frequency of visit of astate-action pair (^, ^) is introduced. (^, ^) is also referred to as allocation. Tocompute the lower bound, a set of confusing models ^^^^ (^) for ^^ is introduced.^^^^(^) = {^^: Π⋆(^^) ∩ Π⋆(^^^ ) = ∅, ^(⋅ |^, ^) ≪ ^^(⋅ |^, ^) ∀(^, ^)},where ^^ = (^, ^, ^^^ , ^, ^) is an MDP with reward ^.The confusing models are statistically similar to the original model ^^but they offer adifferent set of optimal policies for a given reward function ^. However, without furtherassumptions, it is possible that there does not exist any confusing model for a givenreward function. For example, this scenario happens whenever the reward function isidentical in all state-action pairs (therefore all the actions are optimal).It is assumed that is not degenerate if ∀^ ∈ ℛ there exists a model ^′ withtransition function ^′ such that ^(⋅ = Based on the above statements, the lower bound is computed as: ^The above result can be interpreted as a zero-sum game between an agent choosingthe exploration strategy or allocation ^ according to ^, and an opponent that choosesa confusing model ^′. For the lower bound in MR-BPI the opponent needs also toconsider the set of reward functions when choosing the confusing model, which makesthe problem more difficult to solve.Lower Bound relaxation: Unfortunately, even in classical BPI, comprising one rewardfunction, computing characteristic rate ^⋆(^) is a non-convex problem [3]. Instead, asin prior art [3,4,5], a convex upper bound ^⋆(^) of ^⋆(^), called as relaxedcharacteristic rate is computed. Sampling according to ^⋆(^) guarantees that anoptimal policy is identified for a given reward function but at the cost of an increasedover-exploration. This relaxed characteristic rate is typically characterized by minimum sub-optimalitygaps Δ ⋆^(^, ^) = − ^^(^, ^), which are, respectively, the optimal value function^⋆(^) and the optimal action-value function ^⋆^ ^(^, ^), for a reward function ^ and a givenstate-action pair (^, ^). One would then assume that it is possible to obtain a similarresult also in this setting by minimizing these gaps over the set of reward functions.Unfortunately, there are some challenges when applying this logic. The main issue isthat, even with a convex set of reward functions, the minimum sub-optimality gap maybe non-convex and possibly discontinuous, and the discontinuity comes from the fact that in a certain state (or multiple states) all the actions in that state become optimalfor specific reward functions.To define the relaxed characteristic rate, a few definitions of problem-dependentquantities for MDPs are introduced. In a given state ^ ∈ ^, and for a given rewardfunction ^, a set of optimal actions in ^ is indicated by ^ {^ ∈^: ^⋆^(^, ^) = ^⋆^ (^)}. Let Δ^ ≔ min Δ (^, ^) be the minimum gap for a^,^∉^(^;^ )^^given reward function. Additionally, let Var^(^, ^) ) −E ^^^∈^(⋅|s, a [V⋆) ^(s^^)]] be the variance of the optimal value function in the next state,and be the maximum deviationstarting from (^, ^) for a given reward function ^. We also define a reward-specific The following result states an upper bound on ^(ω; ^) ≤ ^(ω; M) for all ω: ^^ is a function of the discount factor ^, the maximum deviation MD^(^, ^) and thevariance Var^(^, ^). It is denoted that an optimal allocation ω⋆ = arg ^ ∈ in ^f(^) U(ω ; ^).The main idea is that the exploratory policy or allocation should follow the samplingstrategy as dictated by the solution to the relaxed characteristic time ^⋆=arg^∈i^n(f^)^(^; ^). However, as mentioned in previous works [7, 6], the model MDPis unknown, and thus, ^⋆ cannot be computed directly. Alternatively, using certaintyequivalence principle, and the current estimate of the model ^^in place of the truemodel ^ the allocation at step ^ can be computed as ^⋆^ = arg^∈i^n(f^^where in ^(^, ^^) all the MDP-related quantities, such as Var^, MD^, ^^ , Δ^, arecomputed with respect to the MDP ^^with reward ^, that we also denote by ^^,^. Since the model ^^is uniquely identified by its transition function, the estimate ^^is equivalently given by the empirical transition function at time ^: otherwise. Figure 5 refers to the communication between the first network node 1100 and thesecond network node 1200 in accordance with the methods 100, 200 and theirembodiments described above.Figure 7 refers to a training scheme named Multiple-Rewards Navigate and Stop, MR-NaS (algorithm for performing MR-BPI) for training the RL algorithm for tabular MDPsto perform a task in the communication network 1000. The method steps of the trainingscheme are divided between the first network node 1100 and the second network node1200 during the training process.In an embodiment, the first network 1100 may comprise one or more criteria for performing the task. In another embodiment, referring to Figure 5, step 501, the second network node 1200may transmit one or more criteria for performing the task to the first network node1100.In an embodiment, the task may comprise the RAN parameter control task such asantenna tilt optimization, downlink link adaptation, etc. In an embodiment, the criteria for performing the task may comprise Reference Signal Received Power (RSRP). In an embodiment, the criteria for performing the task may comprise Signal to Noise plus Interference Ratio (SINR). In an embodiment, the criteria for performing the task maycomprise Transport Block size (TBS). In an embodiment, the criteria for performing thetask may comprise spectral efficiency. In an embodiment, the criteria for performing the task may comprise the number of transmissions. In an embodiment one or more criteria can be referred to as communication network intents. In intent-based operations for communication networks, the term “intent” refers to objectives, requirements, and constraints for a service or network operation; see, for example, 3GPP TS 28.312. Example of intents may be user-specific quality of service (QoS) parameters, such as the 5G QoS Identifier, 5QI, parameters of a 3GPP NG-RAN system, defining service requirements in terms of reliability, latency, throughput etc. In an embodiment, the criteria for performing the task may comprise a combination of the criteria described above. Referring to Figure 5, step 502, the first network node 1100 may transform the one or more criteria into the multiple reward functions. The first network node may transform the one or more criteria into the multiple reward functions by setting threshold to each criteria of the one or more criteria. The first network node may transform the one or more criteria into the multiple reward functions by setting threshold to a combination of criteria of the one or more criteria.In an embodiment, method 500 may further comprise, step 701 of Figure 7,initialization of confidence level δ ∈ (0, 1 / 2), exploration terms α, β ∈ [0,1] with α + β ≤1, an iteration counter or a timestep counter ^ ← 1 and a state-action counter ^^(^, ^) ←0 for all states and actions.In an embodiment, method 500 may further comprise, step 702 of Figure 7, computingallocationAt this step, the optimization problem defining is solved, i.e.,with the plug-in estimate of M^as described above, and the estimated optimal allocation is obtained. An average cumulative allocation is alsocomputed.In step 503, the first network node 1100 transmits the first message comprising theparameters of the RL algorithm.In an embodiment, the parameters of the RL algorithm may comprise the first policy^⋆^, the second policy ^^,^ and the policy weighting factor ^^, wherein = The second network node 1200 may compute the sampling rule for action selection ^^according In an embodiment, the parameters of the RL algorithm may comprise the averagecumulative allocation ^∗^, the second policy ^^,^ and the policy weighting factor ^^. Thesecond network node 1200 may compute the first policy ^⋆^(^|^) = and use the computed first policy to compute the sampling rule for action selection ^^according to The sampling rule for action selection is defined through a convex combination of two components and ii) The first component tracks an action that is according to the sampling rule, which is defined through ^^⋆computed at step 702. The latter component is a forcing component that ensures that each action issufficiently often. More precisely, ,^ with ^^(^^) denoting a number of visits to state ^^up to time ^ ^^^^(∑^ ^^(^,^) )^ ^^^ |^^(^,^)^^ ^^^ ^^(^,^)| . In an embodiment, the parameters of the RL algorithm may comprise an indication of number of actions for the second network node 1100 to perform based on the sampling rule for action selection ^^. The indication of number of actions for the second network node 1200 to perform based on the sampling rule for action selection ^^comprises a minimum or a maximum number of actions for the second network node 1200 to perform based on the sampling rule for action selection ^^.In Figure 5, step 504, and Figure 7, step 703, the second network node 1200determines, based on the received parameters of the RL algorithm, one or moreactions associated with the task. Determining one or more actions associated with thetask comprises determining one or more actions ^^ based on the sampling rule ^^ anda state ^^ observed by the second network node (1100), ^^ ∼ ^^(⋅ |^^).In Figure 5, step 505, and Figure 7, step 703, the second network node 1200 executesthe determined one or more actions associated with the task.As shown in Figure 7, step 704, the second network node 1200, upon executing thedetermined one or more actions associated with the task, may observe a transition from state ^^to state ^^^^.Referring to Figure 5, step 506, the second network node 1200 obtains one or moreobservations based on the execution of the determined one or more actions, whereineach observation is associated with an action. In an embodiment an observation maycomprise an action ^^ executed at the ^^^ timestep of the RL algorithm training, a setof rewards ^^ provided for the execution of the action ^^, wherein each reward in theset of rewards ^^is provided by one reward function of the multiple reward functions and the state ^^^^observed upon executing the action ^^.In Figure 5, step 507, the second network node 1200 transmits the second message,to the first network node 1100, the second message comprising the one or more observations. In step 508 and step 705, the first network node 1100 updates the parameters of the RL algorithm based on the received one or more observations comprising the rewards.Updating the parameters comprises setting ^ ← ^ + 1 , ^^(^^, ^^) ← ^^(^^ , ^^) + 1 andupdating As per steps 509 and 706, the first network node 1100 performs the transmitting 503, the receiving 507 and the updating 508 iteratively until the stopping criterion exceeds a threshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm. As per steps 509 and 706, the second network node 1200 performs the receiving 503,the determining 504, the executing 505, the obtaining 506 and the transmitting 507iteratively until the stopping criterion exceeds a threshold, wherein the stoppingcriterion is based on at least one of the parameters of the RL algorithm.The stopping criterion is defined through a relaxed generalized likelihood ratio term^^(^^ / ^; ^^) and a threshold function ^(^^, ^). The term ^^ / ^ refers to the rates of visitof each state-action pair, while ^^ is an estimate of the MDP at the ^^^ timestep of theRL algorithm training. The term is computed as where ^^ / ^ is theallocation and ^^ is the MDP considered. The threshold function is selected toguarantee that the algorithm is ^-PC and is generally chosen + In step 509, when the stopping criterion is satisfied the first network node 1100 storesthe updated parameters of the RL algorithm. End of the RL algorithm training is markedby the stopping criterion being satisfied.The updated sampling rule for action selection at the end of the RL algorithm trainingis denoted by ^^ ⋆^,^ wherein τ denotes the stopping time or the final value of thetimestep. ^^ ⋆^,^ is updated during the iterative process as shown in Figures 5 and 7corresponding to each reward function of the multiple reward functions, thus, ^^ ⋆^,^(output, in step 707, of MR-NaS scheme) is the policy for any reward function ^ fromthe multiple reward functions ℛ. The policy ^^ ⋆^,^ is capable of generalizing to eachreward function of the multiple reward functions.The method steps described by Figures 500 and 700 is depicted in the followingpseudo code: MR-NaS training schemeInitialise: Confidence value ^; exploration terms (^, ^); a set of reward functions ℛ1: Set counter all (^, ^) ∈ ^ ×2: Set exploration term ^^ = 1 / ^^(^^)^and observe ^1.3: while 4: Compute let 5: Set 6: Determine and execute ^ ∼ ^^(・|^^) and observe ^^^^ ∼ ^7: Update 8: end whileOutput: a policy ^^ ⋆^,^ for a given reward function ^Extension to Deep-Reinforcement Learning The method presented in Figure 7 is valid for MDPs with finite state-action spaces. Inthis section, extending MR-NaS for Deep Reinforcement Learning, Deep-RL, to enablemultiple-rewards exploration in MDPs with continuous state spaces is disclosed. Theconcern with extending MR-NaS to Deep-RL is the exploration phase, where an agentinteracts with the environment to collect a dataset ^ = {(^^ , ^^)}^ without knowingwhich reward function will be chosen when the Deep-RL algorithm is deployed. Anadapted version of MR-NaS in the exploration phase is proposed to collect a dataset^ from the environment. Extending MR-NaS to Deep-RL is not straightforward. Firstly,MR-NaS is a model-based method; secondly there is no closed-form solution toarg ⋆^∈i^n(f^)^ (^; ^) .In [5], it has been suggested that a model-free solution can be achieved by (i) using agenerative solution for arg in ⋆^f ^ (^; ^), and (ii) by estimating a parametric uncertaintyof the MDP-specific quantities (sub-optimality gaps, etc). In other words, the generative solution may be close enough to the model-based one. However, thissolution may be biased because the true sub-optimality gaps are not known, andtherefore, the parametric uncertainty has to considered in the estimates.The approach in [5], termed as Deep Bootstrapped Model Free- Best PolicyIdentification, DBMF-BPI, discloses learning an ensemble of ^ functions of the optimalpolicy and, similarly, learning a ^-th moment of those ^-functions for a single rewardfunction. ^-th moment of the ^-function is called ^-function. To extend DBMF-BPI tomultiple reward functions, the following changes are proposed:- The ^-function (also the ^-function) needs to be trained for multiple rewardfunctions. In the finite action setting this implies that there are |^||ℛ| values toestimate in every state, wherein |^| denotes a total number of actions and |ℛ|denotes a total number of rewards. Alternatively, one can choose to feed asinput to the deep neural networks a one-hot encoding of a chosen reward, butit is less effective in terms of model capacity. -During exploration phase, choosing reward functions to explore based on adifficulty of the reward functions.The proposal is to encode an exploration difficulty of each reward function through acategorical distribution, that is, more probability mass is assigned to more difficultreward functions. An agent samples from this distribution which reward function toexplore at each timestep. For stability, this sampling does not need to be repeated ateach iter, but can be done periodically (e.g., every 1 / (1 − γ) steps).After sampling the reward function ^ from the set of reward functions ℛ, ^ ∼ ℛ, thechosen reward function is used by the DBMF-BPI [5] algorithm to compute the exploration strategy.A linear combination of reward function can also considered.To estimate the exploration difficulty of a reward function or a linear combination ofreward functions, an estimate of the minimum sub-optimality gap Δ^,^ at time ^ for thereward function or the linear combination of reward functions is used. The explorationdifficulty of the reward function or the linear combination of reward functions,proportional is to 1 / Δ^^,^.Additionally, a forced tracking of the reward functions can be used to ensure that eachreward function is sampled sufficiently often during the exploration phase. One way toensure that is by sampling each reward function with a uniform probability. Or, ^^ ={^ ∈ ^: √^ − ^^,^ / 2 < 0} can be a set of under-sampled reward functions at timestep ^,where ^^,^ is a number of times the algorithm has chosen a reward function ^ up tothe previous timesteps (0 to ^ − 1) or till the ^^^ timestep and choose ^ = ^^^ ^^^^^^,^whenever ^^ is non-empty. Otherwise, reward function ^ is chosen with a probabilityinversely proportional to the minimum sub-optimality gap, ∝ 1 / Δ^^,^. This forcedtracking procedure guarantees that each reward function is sampled or chosen at arate of ^^√^^. Following embodiments will be described with reference to Figure 5 and Figure 8.Figure 8 refers to a training scheme named Deep Bootstrapped Multiple Rewards BestPolicy Identification, DBMR-BPI, for training RL algorithm in continuous state-actionsspace with multiple reward functions to perform a task in the communication network1000. The method steps of the training scheme are divided between the first network node 1100 and the second network node 1200. In an embodiment, the first network 1100 may comprise one or more criteria for performing the task. In another embodiment, referring to Figure 5, step 501, the second network node 1200 may transmit one or more criteria for performing the task to the first network node 1100. In an embodiment one or more criteria can be referred to as communication networkintents. In intent-based operations for communication networks, the term “intent” refersto objectives, requirements, and constraints for a service or network operation; see, for example, 3GPP TS 28.312. Example of intents may be user-specific quality ofservice (QoS) parameters, such as the 5G QoS Identifier, 5QI, parameters of a 3GPPNG-RAN system, defining service requirements in terms of reliability, latency, throughput etc. In an embodiment, the task may comprise the radio access network parameter controltask such as antenna tilt optimization, downlink link adaptation, etc. In an embodiment,the criteria for performing the task may comprise Reference Signal Received Power (RSRP). In an embodiment, the criteria for performing the task may comprise Signal to Noise plus Interference Ratio (SINR). In an embodiment, the criteria for performing the task may comprise Transport Block Size (TBS). In an embodiment, the criteria for performing the task may comprise spectral efficiency. In an embodiment, the criteriafor performing the task may comprise a number of transmissions.In an embodiment, the criteria for performing the task may comprise the combinationof the criteria described above. Referring to Figure 5, step 502, the first network node 1100 may transform the one or more criteria into the multiple reward functions. The first network node may transform the one or more criteria into the multiple reward functions by setting threshold to each criteria of the one or more criteria. The first network node may transform the one ormore criteria into the multiple reward functions by setting threshold to the combinationof criteria of the one or more criteria.In an embodiment, method 500 may comprise step 801 of Figure 8. In step 801, thefirst network node 1100 may initialise an ensemble of neural networks as described inDBMF-BPI [5]: an ensemble of Q-functions and an ensemble of M-functions. The firstnetwork node 1100 may initialise an estimate of the sub-optimality gap Δ^,^ for eachreward function ^ ∈ ℛ.The method 500 may comprise step 802 of Figure 8. At every timestep ^, orperiodically, a reward function ^ is chosen by the first network node 1100. Thereward function ^ can be chosen randomly and / or deterministically.If randomly, the probability of choosing a reward function ^ is inversely proportional tothe minimum sub-optimality gap, A categorical distribution can be used to sample / choose a reward function for exploration. This categorical distributioncan be mixed with a uniform distribution to guarantee a minimum probability. Anotherchoice is a SoftMax distribution of the minimum sub-optimality gap Δ^,^.If deterministically, the first network node 1100 can choose the least chosen reward.The method 500 may comprise step 803 of Figure 8. In step 803, the first networknode 1100 may compute allocation for the chosen reward function and using the computed allocation , the sampling rule for action selection π^ is computedaccording to DBMF-BPI [5].In step 503, the first network node 1100 transmits the first message comprising theparameters of the RL algorithm. The parameters of the RL algorithm may comprise the sampling rule for action selection ^^. The parameters of the RL algorithm may comprise the allocation ^∗^ . The parameters of the RL algorithm may comprise theestimate Δ^,^for each reward function. In Figure 5, step 504, and Figure 8, step 804, the second network node 1200 determines, based on the received parameters of the RL algorithm, one or moreactions associated with the task. Determining one or more actions associated with thetask comprises determining one or more actions ^^ based on the sampling rule ^^ anda state ^^ observed by the second network node 1100, ^^ ∼ ^^(⋅ |^^).In Figure 5, step 505, and Figure 8, step 704, the second network node 1200 executesthe determined one or more actions associated with the task. As shown in Figure 8, step 804, the second network node 1200, upon executing the determined one or more actions associated with the task, may observe a transition from state ^^to state ^^^^. Referring to Figure 5, step 506, the second network node 1200 obtains one or more observations based on the execution of the determined one or more actions, whereineach observation is associated with an action. The observation comprises: -an action ^ ^^ executed at the ^ ^ timestep of the RL algorithm training,- a set of rewards ^^ = provided for the execution of the action ^^, whereineach reward in the set of rewards ^^is provided by one reward function of the multiple reward functions and ^ is equal to the total number of reward functions(^ = |ℛ| ) and- a state ^^^^ observed upon executing the action ^^In Figure 5, step 507, the second network node 1200 transmits the second message,to the first network node 1100, the second message comprising the one or more observations.The method 500 may comprise the step 805 of Figure 8. In step 805, the first networknode 1100 may store the one or more observations to a buffer ^ . Buffer ^ is databuffer to store observations provided by the second network node 1200.The method 500 may comprise the step 806 of Figure 8. In step 806, the first networknode 1100 may randomly sample a batch of ^ observations from the buffer ^. The firstnetwork node 1100 may use the sampled batch of ^ observations to train theensemble of Q-functions or Q networks and the ensemble of M-functions or M-networks as in DBMF-BPI [5] for each reward function ^ ∈ ℛ.In step 508, the first node node 1100 updates the parameters of the RL algorithmbased on the received one or more observations comprising the rewards. Updatingthe parameters of the RL algorithm comprises updating the estimate Δ^,^ for eachreward function using the sampled batch of ^ observations and the trained ensembleQ networks. As per steps 509 and 807, the first network node 1100 performs the transmitting 503, the receiving 507 and the updating 508 iteratively until the stopping criterion exceeds a threshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm. As per steps 509 and 807, the second network node 1200 performs the receiving 503, the determining 504, the executing 505, the obtaining 506 and the transmitting 507 iteratively until the stopping criterion exceeds a threshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm.In an embodiment stopping criterion can be a user defined stopping criterion. In anembodiment stopping criterion can be a number of observations collected and stored in the buffer ^, with the threshold denoting a minimum or maximum number of observations that can be collected and stored in the buffer ^. In step 509, when the stopping criterion is satisfied the first network node 1100 stores the updated parameters of the RL algorithm. End of the RL algorithm training is marked by the stopping criterion being satisfied.DBMR-BPI training scheme1: Initialize buffer ^, parameters of the algorithm.2: for ^ = 0, 1, 2, ..., do3: Choose reward ^ with probability ∝^ ^^and compute allocation ^^using DBMF- ^,^BPI with reward ^4: Determine and execute ^^ ∼ and observe ^^^^ ∼ ^(・|^^, ^^)5: Add observation (^^, ^^ , ^^,^ , ... , ^^,^, ^^^^) to the buffer ^, with ^^,^ = ^^(^^, ^^), ^ =1, ... , |^|6: Train the ensemble Q-networks according to DBMF-BPI and update estimates ^^,^ for each reward function ^ ∈ ^7: end forThe output, in step 808, of the DBMR-BPI training scheme as depicted by Figures 5and 8 is the trained ensemble of Q-networks and the buffer ^. The ensemble of Q-networks trained according to the method steps of Figure 5 and 8 refer to the policycorresponding to each reward function of the multiple reward functions in the contextof Deep-RL with multiple reward functions. The ensemble of Q-networks trainedaccording to the method steps of Figure 5 and 8 is capable of generalizing to eachreward function of the multiple reward functions.Figure 3 depicts a method 300 performed by the first network node 1100, for executingthe RL algorithm to perform a task in the communication network 1000. The RLalgorithm being configured with multiple reward functions wherein each rewardfunction of the multiple reward functions is associated with the task. The method 300 comprises receiving, 301, a request message, from the second network node 1200.The request message comprising a request for the policy corresponding to eachreward function of the multiple reward functions. In an embodiment, the requestmessage indicates the multiple reward functions. In an embodiment, the requestmessage indicates one or more criteria for performing the task. In an embodiment, themethod 300 comprises, the first network node 1100, transforming the one or morecriteria into the multiple rewards functions. The method 300 comprises verifying, 302,availability of the policy corresponding to each reward function of the multiple rewardfunctions. The method 300 comprises, the first network node 1100, transmitting 304 aresponse message to the second network node 1200. The response message comprising an indication regarding the availability of the policy corresponding to each reward function of the multiple reward functions. In an embodiment, the response message comprises the policy corresponding to each reward function of the multiple reward functions. In an embodiment, the method 300 comprises, the first network node 1100,transmitting 304 the policy corresponding to each reward function of the multiplereward functions. The policy corresponding to each reward function of the multiplereward functions is transmitted by the first network node either through the responsemessage 303 or through a separate message 304. The policy corresponding to eachreward function of the multiple reward functions may comprise ^^ ⋆^,^ outputted by MR-NaS training scheme (as shown in Figure 5 and 7). The policy cor1843responding toeach reward function of the multiple reward functions may comprise ensemble of Q-networks outputted by DBMR-BPI training scheme (as shown in Figures 5 and 8).Figure 4 depicts a method 400 performed by a second network node 1200, forexecuting the RL algorithm to perform a task in the communication network 1000. TheRL algorithm being configured with multiple reward functions wherein each rewardfunction of the multiple reward functions is associated with the task. The method 400 comprises, the second network node 1200, transmitting 401 a request message, tothe first network node 1100. The request message comprising a request for the policycorresponding to each reward function of the multiple reward functions. In anembodiment, the request message indicates the multiple reward functions. In anembodiment, the request message indicates one or more criteria for performing thetask. The method 400 comprises, the second network node 1200, receiving 402 aresponse message, from the first network node 1100. The response messagecomprising an indication regarding availability of the policy corresponding to eachreward function of the multiple reward functions. In an embodiment, the responsemessage further comprises the policy corresponding to each reward function of themultiple reward functions. In an embodiment, the method 400 comprises, the secondnetwork node 1200, receiving 403 the policy corresponding to each reward function ofthe multiple reward functions. The policy corresponding to each reward function of themultiple reward functions is received by the second network either through theresponse message 402 or through a separate message 403. The method 400comprising, the second network node 1200, executing, 404, the policy correspondingto each reward function of the multiple reward functions. the policy corresponding toeach reward function of the multiple reward function is executed to perform the task inthe communication network 1000. The policy corresponding to each reward functionof the multiple reward functions may comprise ^^ ⋆^,^ outputted by MR-NaS trainingscheme (as shown in Figure 5 and 7). The policy corresponding to each rewardfunction of the multiple reward functions may comprise ensemble of Q-networksoutputted by DBMR-BPI training scheme (as shown in Figures 5 and 8).Figure 6 corresponds to the communication between the first network node 1100 and the second network node 1200 in order to execute methods 300 and 400, respectively, as discussed above.Figure 9 illustrates examples of the first network node 1100, the second network node1200, and the communication network node 1000. The first network node 1100illustrated in Figure 9 implements the method 100, 300, 500, 600 as illustrated inFigures 1, 3, 5 and 6, respectively, according to examples of the disclosure, forexample on receipt of suitable instructions from a computer program 1101. The firstnetwork node 1100 comprises a processor or processing circuitry 1102, and a computer program product 1104 in the form of a memory 1103. The processing circuitry 1102 is operable to perform some or all of the steps of the embodiments of methods 100, 300, 500 and 600 as discussed above with reference to Figures 1, 3, 5 and 6. The memory 1103 contains instructions executable by the processing circuitry1102 such that the first network node 1100 is operable to perform some or all of theembodiments of the method 100, 300, 500 and 600 as illustrated in Figures 1, 3, 5 and 6. The instructions may also include instructions for executing one or more telecommunications and / or data communications protocols. The instructions may be stored in the form of the computer program 1101. In some examples, the processor or processing circuitry 1102 may include one or more microprocessors or microcontrollers, as well as other digital hardware, which may include digital signal processors (DSPs), special-purpose digital logic, etc. The processor or processing circuitry 1102 may be implemented by any type of integrated circuit, such as an Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA), etc. The memory 1103 may include one or several types of memory suitable for the processor, such as read-only memory (ROM), random-access memory, cache memory, flash memory devices, optical storage devices, solid state disk, hard disk drive, etc.The second network node 1200 illustrated in Figure 9 implements the method 200,400, 500, 600 as illustrated in Figures 2, 4, 5 and 6, respectively, according to examples of the disclosure, for example on receipt of suitable instructions from a computer program 1201. The second network node 1200 comprises a processor or processing circuitry 1202, and a computer program product 1204 in the form of a memory 1203. The processing circuitry 1202 is operable to perform some or all of the steps of the embodiments of methods 200, 400, 500 and 600 as discussed above with reference to Figures 2, 4, 5 and 6. The memory 1203 contains instructions executable by the processing circuitry 1202 such that the first network node 1200 is operable to perform some or all of the embodiments of the method 200, 400, 500 and 600 as illustrated in Figures 2, 4, 5 and 6. The instructions may also include instructions for executing one or more telecommunications and / or data communications protocols. The instructions may be stored in the form of the computer program 1201. In some examples, the processor or processing circuitry 1202 may include one or more microprocessors or microcontrollers, as well as other digital hardware, which may include digital signal processors (DSPs), special-purpose digital logic, etc. The processor or processing circuitry 1202 may be implemented by any type of integrated circuit, such as an Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA), etc. The memory 1203 may include one or several types of memory suitable for the processor, such as read-only memory (ROM), random-access memory, cache memory, flash memory devices, optical storage devices, solid state disk, hard disk drive, etc. The communication network 1000 illustrated in Figure 9 enables connectivity between user equipments, UEs, and communication network nodes, such as the first networknode 1100 and the second network node 1200. In that sense, the communicationnetwork 1000 may be a 3rd Generation Partnership Project, 3GPP, network and may be configured to operate according to predefined rules or procedures, such as specific standards that comprise, but are not limited to: Global System for Mobile Communications (GSM); Universal Mobile Telecommunications System (UMTS); Evolved Packet System (EPS), 5th Generation System (5GS) or any applicable future generation standard (e.g.,3GPP 6G). The communication network 1000 here comprises one or more radio access network (RAN) nodes, and one or more core network nodes of an Evolved Packet Core (EPC) or an 5th Generation Core (5GC). A RAN node 102 may comprise a base station, eNodeB of an LTE network, gNodeB of a New Radio network, or any other current or future implementation of functionality facilitating the exchange of radio network signals between nodes of the communication network 100 and / or UEs. Moreover, as will be appreciated by those of skill in the art, a network node is not necessarily limited to an implementation in which a radio portion and a baseband portion are supplied and integrated by a single vendor. Thus, it will be understood that network nodes include disaggregated implementations or portions thereof. For example, in some embodiments, the communication network comprises one or more Open-RAN (ORAN) network nodes. An ORAN network node is a node in the communication network that supports an ORAN specification (e.g., a specification published by the O-RAN Alliance, or any similar organization) and may operate alone or together with other nodes to implement one or more functionalities ofany node in the communication network, including one or more access network nodesand / or core network nodes. In an embodiment, the first network node 1100 is a RAN node in the communication network 1000. In an embodiment, the first network node 1100 is an Orchestration and Management, OAM, node. In an embodiment, the first network node 1100 a Service Management and Orchestration, SMO, node. In an embodiment, the first network node1100 is a non-real-time RAN Intelligent controller, Non-RT RIC. In an embodiment, thefirst network node 1100 comprises an rApp. rApp is defined by Open RAN. The rAppsare software application designed to run on the Non-RT RIC to realize different RAN management and optimization use cases in an automated way, with control loops on a time scale of one second and longer. In an embodiment, the second network node 1200 is a RAN node. In an embodiment, the second network node 1200 comprises an rApp. In an embodiment, the secondnetwork node 1200 comprises an xApp. xApp is defined by Open RAN. The xApps aresoftware application designed to run on the Non-RT RIC to realize different RAN management and optimization use cases in an automated way, with control loops on a time scale of one second and longer. It is to be understood that the singular form "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms, "comprises" "comprising", "includes" and / or "including" when used herein, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. While various embodiments of the disclosure are described, it should be understood that they have been presented by the way of example only, and not limitations. Thus, the breadth and scope of the disclosure should not be limited by any of the above- described exemplary embodiments. Moreover, any combination of the above- described elements in all possible variations thereof is comprised by the disclosure unless otherwise indicated herein or otherwise clearly contradicted by context. Additionally, while the processes described above and illustrated in the drawings are shown as a sequence of steps this was done for the sake of illustration. It is contemplated that some steps may be added, some steps omitted, the order of the steps may be re-arranged, and some steps may be performed in parallel. The disclosed methods and embodiments provide an optimal policy for a class ofreward functions. For finite MDPs, MR-NaS allows to identify the best policy fordifferent reward functions, as opposed to classical RL, which focus on solving a controlproblem with a pre-specified reward function. For Deep RL, DBMR-BPI is able toexplore for multiple specified reward functions, a capability currently lacking in modernDeep-RL algorithms. The proposed solution offers tunability by accommodatingdifferent classes of reward functions and by allowing for reward functions toparametrized by weight parameters. The proposed framework is flexible that it can beapplied to a wide variety of tasks which can be formulated as RL problem, includingnetwork optimization tasks such as RAN parameters optimization, network selection,scheduling, resource allocations and beyond (e.g., robotics, autonomous driving, and control of dynamical systems). MR-NaS, the proposed algorithm for the finite MDP setting, is sample efficient, i.e., it returns the optimal policy for all considered rewardfunctions using a limited number of samples.Experimental results:In this section, application of DBMR-BPI training scheme to the Downlink linkadaptation problem in radio networks is demonstrated.Link adaptation, or adaptive modulation and coding (AMC), is a technique used in wireless communication systems, such as the 3GPP HSDPA, LTE or NG-RAN, todynamically adapt the transmission rate of a communication link to the time- andfrequency-varying channel conditions. Modulation and Coding Schemes (MCS) effectively adapt the transmission rate by matching the modulation and coding parameters used for communication (such as modulation order, code rate, etc.) to the conditions of the radio link, such as propagation loss, channel strength, interference from other signals concurrently transmitted in the same radio resources, etc.Link adaptation is a dynamic process that is performed in each transmission timeinterval (e.g., on a sub-millisecond time-scale in the 3GPP NG-RAN system) wherein a transmission is scheduled between a base station (a radio node that facilitates wireless communication for user equipment within its coverage area) and a user equipment (UE, a mobile device such as a smartphone or tablet). Link adaptation algorithms attempt to optimally adapt a transmission data rate chosen for a link according to interference conditions of the link. These algorithms select the most appropriate modulation order and coding rate (also known as MCS) based on the latest information available about the state of the communication system as well as the state of the link (in terms of Channel Quality Indicator, CQI). Problem In wireless communication systems, there is a fundamental trade-off between throughput and spectral efficiency, which can be understood in terms of resource elements. Resource elements are the smallest units of time-frequency resources allocated for transmission, encompassing a specific duration of time and a specificfrequency bandwidth. An objective is to explore the communication networkenvironment using DBMR-BPI to design policies that optimize this trade-off based ona desired communication network intent. Higher throughput, an amount of datatransmitted successfully per unit of time, often requires higher-order modulation schemes and denser coding rates, which can reduce spectral efficiency—the effective use of available spectral resources. Conversely, enhancing spectral efficiency through lower-order modulation schemes and robust coding rates can ensure more reliabletransmissions but may limit throughput. As measuring throughput directly ischallenging due to various influencing factors, Transport Block Size (TBS) isconsidered. TBS represents the amount of data that can be transmitted in a given time interval based on the selected Modulation and Coding Scheme (MCS). However, increasing TBS does not necessarily result in higher throughput, as a larger TBS canlead to a higher Block Error Rate (BLER)— a percentage of transport blocks that fail tobe decoded correctly—impacting the reliability of data transmission and reducing an actual amount of successfully received data. Environment, RL formulation and setupA Time-Division Duplexing (TDD) 5G system operating at a 3.5GHz carrier frequency,with PHY layer numerology μ = 0 of the 3GPP technical specifications 38.211 andsingle-user Multi-Input Multi-Output (SU-MIMO) transmission is considered. Each base station is configured as massive MIMO (mMIMO) with an 8x4x2 antenna array.1 site, 3 sectors, and a fixed number of 10 UEs with full buffer traffic experiencingstable interference conditions are considered. 150 randomized scenarios thatrandomize across several variables (e.g., UEs movement and traffic) weregenereated. Each scenario lasted 5 seconds. The downlink link adaptation problem can be formulated as a RL problem with multiple reward functions.The base station, represents the RL agent or the learner node or the second networknode 1200, depending on the channel measurements transmitted by the UEs, choosesthe MCS. The first network node 1100 may represent an OAM node. The first networknode 1100 may represent an SMO node.The state includes current CQI values, number of the antennas of the UE, estimated speed of the UE, rank indicators, and other relevant metrics that affect link quality, including a windowed history of these values. Actions involve selecting different MCS values, which dictate the modulation order and coding rate for the packet transmission. The transmission of a packet from a UE is modelled as an episodic MDP of at most 5steps. At each time step t, the agent chooses an action (the MCS). If the transmissionis successful, the agent receives an ACK and the episode terminates. Otherwise, the agent receives a NACK and observes the new channel measurements (next state) until the episode is done. Reward designTwo reward functions, that optimize for different communication network intents, areconsidered. For each episode, let ^^ be a boolean value indicating if a packet wassuccessfully transmitted (ACK) or not (NACK). Let ^^^^ indicate the transport blocksize for an ^^^ transmission, and let ^^^,^ indicate a number of resource elements forthe ^^^transmission. 1. Transport block size (TBS) reward function: ^^,^ = ^^^^・ 1^^^^, thisreward function encourages the agent to choose actions that maximize thetransport block size (at the expense of other metrics, i.e., error rate, spectralefficiency), measured as the number of bits of successfully transmitted packets.2. Spectral efficiency- resource elements (RE) reward function: ^^,^ = 0 if thetransmission is not terminal; otherwise, ^^,^ = − ∑^^^ ^^^,^ ,representing a total number of resource elements used up to the final time step ^ of the transmission episode.These rewards, if weighted together, can optimize for different intents, betweenthroughput, efficiency and transmission rate.The DBMR-BPI algorithm was evaluated along with RND, APT and Disagreementalgorithms. Numerical resultsThe performance of the various algorithms, DBMR-BPI, Random Network Distillation

[0011] RND, Disagreement

[0012] and Active pre-training, APT

[0013] , in terms of (1)throughput, (2) BLER (block error rate) and (3) spectral efficiency (which measuresthe number of resource elements used) is shown in Figure 10.The distribution of the number of transmissions attempted per packet and the actionselection (MCS index) for each algorithm are represented in Figure 11.From Figures 10 and 11, it can be seen that DBMR-BPI achieves better performanceboth in terms of throughput and spectral efficiency. The MCS index distribution revealsthat DBMR-BPI tends to use all the available actions during the first transmission (notethat only a last 5 MCS index values are used during re-transmissions), which indicatesthat DBMR-BPI is effectively learning which action to take depending on state values.On the other hand RND and Disagreement tend only to focus on initial MCS indexvalues during the first transmission. We also note a lower BLER target for DBMR-BPI,as evidenced also by the fact that the average number of transmissions is lower forDBMR-BPI compared to the other algorithms. In general, these performances appear to be primarily due to optimized usage ofresource elements, as shown in Figure 12. Figure 12 illustrates a distribution ofnormalized rewards collected during training by the algorithms. While the TBS issimilar across all algorithms, DBMR-BPI clearly focuses on optimizing the number ofresource elements.In Figure 13, the estimate ^^,^ for both rewards functions TBS and RE (top plot) andthe number of times DBMR-BPI selected each reward function without forced trackingor uniform exploration (bottom plot). From the top plot it can be seen that optimizingthe number of resource elements seems to be a harder task due to a lower value of^^,^which explains why DBMR-BPI focuses more on optimizing the number ofresource elements used per packet transmission.REFERENCES: 1. Chi Jin, Akshay Krishnamurthy, Max Simchowitz, and Tiancheng Yu. Reward-free exploration for reinforcement learning. In Proc. of the 37th International Conference on Machine Learning, 2020. 2. Andrew J. Wagenmaker, Yifang Chen, Max Simchowitz, Simon Du, and KevinJamieson. Reward-free RL is no harder than reward-aware RL in linear markov decision processes. In International Conference on Machine Learning.PMLR, 2022 3. Aymen Al Marjani, Aurélien Garivier, and Alexandre Proutiere. Navigating to thebest policy in markov decision processes. In Proc. of Neural InformationProcessing Systems (NeurIPS), 2021 4. E. Kaufmann, P. Ménard, O. Darwiche Domingues, A. Jonsson, E. Leurent, andM. Valko. Adaptive reward-free exploration. In Proc. of Algorithmic Learning Theory, 2021. 5. Alessio Russo, Alexandre Proutiere., Model-free active exploration inreinforcement learning. In Proc. of Neural Information Processing Systems, 2023. 6. WD Cook and RJ Webster. Caratheodory’s theorem. Canadian MathematicalBulletin, 1972 7. Yang, Runzhe, Xingyuan Sun, and Karthik Narasimhan. "A generalizedalgorithm for multi-objective reinforcement learning and policy adaptation." Advances in neural information processing systems 32 (2019).8. Mirco Mutti and Marcello Restelli. An intrinsically-motivated approach forlearning highly exploring and fast mixing policies. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5232–5239, 2020.9. Laskin, M., Yarats, D., Liu, H., Lee, K., Zhan, A., Lu, K., Cang, C., Pinto, L., andAbbeel, P. Urlb: Unsupervised reinforcement learning benchmark. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 10.Mutti, Mirco. "Unsupervised reinforcement learning via state entropy maximization." (2023). PhD Thesis http: / / amsdottorato.unibo.it / 10588 / 1 / mutti_mirco_tesi.pdf 11.Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation. In International Conference on LearningRepresentations, 2018. 12.Deepak Pathak, Dhiraj Gandhi, and Abhinav Gupta. Self-supervised exploration via disagreement. In International conference on machine learning, pages 5062–5071. PMLR, 2019. 13.Hao Liu and Pieter Abbeel. Behavior from the void: Unsupervised active pre- training. Advances in Neural Information Processing Systems, 34:18459–18473, 2021.

Claims

CLAIMS1) A method (100, 500), performed by a first network node (1100), for training areinforcement learning, RL, algorithm to perform a task in a communicationnetwork (1000), the RL algorithm being configured with multiple reward functions wherein each reward function of the multiple reward functions is associated with the task, the communication network comprising the first network node (1100) and a second network node (1200), the method (100, 500) comprising: -transmitting (103, 503) a first message, to the second network node (1200),comprising parameters of the RL algorithm; -receiving (104, 507) a second message, from the second network node(1200), comprising one or more observations, wherein each observation of the one or more observations comprise rewards provided by the multiplereward functions; -updating (105, 508) the parameters of the RL algorithm based on thereceived one or more observations comprising the rewards; -performing (106, 509) the transmitting (103, 503), the receiving (104, 507)and the updating (105, 508) iteratively until a stopping criterion exceeds athreshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm; and -storing (107, 510) the updated parameters of the RL algorithm.2) The method (100, 500) according to claim 1, comprising:- receiving (101, 501), from the second network node (1200), one or morecriteria for performing the task.3) The method (100, 500) according to claim 1 or 2, comprising:- transforming (102, 502) one or more criteria for performing the task intothe multiple reward functions.4) The method (100, 500) according to any of claims 1-3, wherein the parameters ofthe RL algorithm comprise one or more of the following: -a sampling rule for action selection- an average cumulative allocation- a first policy ^∗ ^- a second policy ^^,^- a policy weighting factor ^^- an indication of number of actions for the second network node (1100) toperform based on the sampling rule for action selection ^^, wherein t indicates a value of an iteration of the RL algorithm training.5) The method (100, 500) according to claim 4, wherein the sampling rule for actionselection ^^is computed according towherein^ indicates state and ^, ^′ indicate action.6) The method (100, 500) according to any of claims 4-5, wherein the indication ofnumber of actions for the second network node (1200) to perform based on thesampling rule for action selection ^^comprises a minimum or a maximum number of actions for the second network node (1200) to perform based on the samplingrule for action selection ^^.7) The method (100, 500) according to any of the claims 1-6, wherein eachobservation in the one or more observations comprises: -an action ^^ executed at the ^^^iteration of the RL algorithm training; -a set of rewards ^^ provided for the execution of the action ^^, whereineach reward of the set of rewards ^^ is provided by one reward function ofthe multiple reward functions; and -a new state ^^^^ observed upon executing the action ^^.8) A method (200, 500) performed by a second network node (1200), for training areinforcement learning, RL, algorithm to perform a task in a communication network (1000), the RL algorithm being configured with multiple reward functionswherein each reward function of the multiple reward functions is associated withthe task, the communication network (1000) comprising a first network node (1100)and the second network node (1200), the method (200, 500) comprising:- receiving (202, 503) a first message, from the first network node (1100),comprising parameters of the RL algorithm; -determining (203, 504), based on the received parameters of the RLalgorithm, one or more actions associated with the task; -executing (204, 505) the determined one or more actions associated withthe task; -obtaining (205, 506) one or more observations based on the execution ofthe determined one or more actions, wherein each observation of the one or more observations comprise rewards provided by the multiple reward functions; -transmitting (206, 507) a second message, to the first network node(1100), the second message comprising the one or more observations; and -performing (207, 509) the receiving (202, 503), the determining (203, 504),the executing (204, 505), the obtaining (205, 506) and the transmitting (206, 507) iteratively until a stopping criterion exceeds a threshold, wherein thestopping criterion is based on at least one of the parameters of the RL algorithm.9) The method (200, 500) according to claim 8, comprising:- transmitting (201, 501), to the first network node (1100), one or morecriteria for performing the task. 10)The method (200, 500) according to any of claims 8-9, wherein the parameters of the RL algorithm comprise the following: -a first policy ^∗ ^- a second policy-a policy weighting factor ^^wherein ^ indicates a value of an iteration of the RL algorithm training.11) The method (200, 500) according to claim 10, comprising:- computing a sampling rule for action selection ^^ according to12) The method (200, 500) according to any of claims 8-9, wherein the parametersof the RL algorithm comprise the following: -an average cumulative allocation ^∗ ^- a second policy ^^,^- a policy weighting factor ^^wherein ^ indicates a value of an iteration of the RL algorithm training.13) The method (200, 500) according to claim 12, comprising:- computing a first policy ^^⋆=wherein ^ indicates state and ^, ^′indicate action; and -computing a sampling rule for action selection ^^ according to14) The method (200, 500) according to any of claims 10-13, wherein theparameters of the RL algorithm comprise: -an indication of number of actions for the second network node (1200) toperform based on the sampling rule for action selection ^^.15) The method (200, 500) according to any of the claims 8-14, wherein eachobservation in the one or more observations comprises: -an action ^^ executed at the ^^^iteration of the RL algorithm training; -a set of rewards ^^ provided for the execution of the action ^^, whereineach reward in the set of rewards ^^is provided by one reward function of the multiple reward functions; and -a new state ^^^^ observed upon executing the action ^^.16) The method (200, 500) according to any of claims 8-15, wherein determining oneor more actions associated with the task comprises determining one or moreactions ^^ based on the sampling rule ^^ and a state ^^ observed by the secondnetwork node (1100), ^^ ∼ ^^(⋅ |^^).17) A method (300, 600) performed by a first network node (1100), for executing areinforcement learning, RL, algorithm to perform a task in a communication network (1000), the RL algorithm being configured with multiple reward functionswherein each reward function of the multiple reward functions is associated with the task, the communication network (1000) comprising the first network node(1100) and a second network node (1200), the method (300, 600) comprising:- receiving (301, 601) a request message, from the second network node(1200), comprising a request for a policy corresponding to each rewardfunction of the multiple reward functions; -verifying (302, 602) availability of the policy corresponding to each rewardfunction of the multiple reward functions; and -transmitting (304, 603) a response message, to the second network node(1200), comprising an indication regarding the availability of the policy corresponding to each reward function of the multiple rewardfunctions.18) The method (300, 600) of claim 17, wherein the request message indicates themultiple reward functions.19) The method (300, 600) of claim 17, wherein the request message indicates oneor more criteria for performing the task.20) The method (300, 600) of claim 19, comprising:- transforming the one or more criteria into the multiple rewards functions.21) The method (300, 600) according to any of claims 17-20, wherein the responsemessage comprises available policy corresponding to each reward function ofthe multiple reward functions.22) The method (300, 600) according to any of claims 17-20, comprising:- transmitting (304, 604) available policy corresponding to each rewardfunction of the multiple reward functions.23) A method (400, 600) performed by a second network node (1200), for executinga reinforcement learning, RL, algorithm to perform a task in a communication network (1000), the RL algorithm being configured with multiple reward functionswherein each reward function of the multiple reward functions is associated withthe task, the communication network comprising a first network node (1100) andthe second network node (1200), the method (100, 600) comprising:- transmitting (401, 601) a request message, to the first network node(1100), comprising a request for a policy corresponding to each rewardfunction of the multiple reward functions; and -receiving (402, 603) a response message, from the first network node(1100), comprising an indication regarding availability of the policy corresponding to each reward function of the multiple rewardfunctions.24) The method (400, 600) of claim 23, wherein the request message indicates themultiple reward functions.25) The method (400, 600) of claim 23, wherein the request message indicates oneor more criteria for performing the task.26) The method (400, 600) according to any of claims 23-25, wherein the responsemessage further comprises the policy corresponding to each reward function ofthe multiple reward functions.27) The method (400, 600) according to any of claims 22-24, further comprising:- receiving (403, 604) the policy corresponding to each reward function ofthe multiple reward functions.28) The method (400, 600) according to any of claims 25-26, further comprising:- executing (404, 605) the policy corresponding to each reward function ofthe multiple reward functions.29) The method (100, 200, 300, 400, 500, 600) according to any of claims 1-28,wherein the task in the communication network (1000) comprises a radio access network, RAN, parameter control task.30) The method (100, 200, 300, 400, 500, 600) according to claim 29, wherein thetask in the communication network (1000) comprises downlink link adaptation.31) The method (100, 200, 300, 400, 500, 600) according to claim 29, wherein thetask in the communication network (1000) comprises antenna tilt optimization.32) The method (100, 200, 300, 400, 500, 500) according to any of the claims 1-31,wherein the first network node (1100) comprises:- a RAN node,- an Orchestration and Management, OAM, node,- a Service Management and Orchestration, SMO, node,- a non-real-time RAN Intelligent controller, Non-RT RIC, or- an rApp.33) The method (100, 200, 300, 400, 500, 600) according to any of claims 1-32,wherein the second network node (1200) comprises:- a RAN node,- an rApp, or- an xApp.34) A first network node (1100), for training a reinforcement learning, RL, algorithm toperform a task in a communication network (1000), the RL algorithm beingconfigured with multiple reward functions wherein each reward function of themultiple reward functions is associated with the task, the communication network comprising the first network node (1100) and a second network node (1200), the first network node (1100) being configured to: -transmit (103, 503) a first message, to the second network node (1200),comprising parameters of the RL algorithm; -receive (104, 507) a second message, from the second network node(1200), comprising one or more observations, wherein each observation ofthe one or more observations comprise rewards provided by the multiplereward functions; -update (105, 508) the parameters of the RL algorithm based on the receivedone or more observations comprising the rewards; -perform (106, 509) the transmitting (103, 503), the receiving (104, 507) andthe updating (105, 508) iteratively until a stopping criterion exceeds athreshold, wherein the stopping criterion is based on at least one of the parameters of the RL algorithm; and -store (107, 510) the updated parameters of the RL algorithm.35) The first network node (1100) according to claim 34, being configured to performa method according to any of claims 2-7 and 29-33.36) A second network node (1200), for training a reinforcement learning, RL, algorithmto perform a task in a communication network (1000), the RL algorithm beingconfigured with multiple reward functions wherein each reward function of themultiple reward functions is associated with the task, the communication network (1000) comprising a first network node (1100) and the second network node(1200), the second network node being configured to: -receive (202, 503) a first message, from the first network node (1100),comprising parameters of the RL algorithm; -determine (203, 504), based on the received parameters of the RLalgorithm, one or more actions associated with the task; -execute (204, 505) the determined one or more actions associated withthe task; -obtain (205, 506) one or more observations based on the execution of thedetermined one or more actions, wherein each observation of the one or more observations comprise rewards provided by the multiple reward functions; -transmit (206, 507) a second message, to the first network node (1100),the second message comprising the one or more observations; and -perform (207, 509) the receiving (202, 503), the determining (203, 504), theexecuting (204, 505), the obtaining (205, 506) and the transmitting (206,507) iteratively until a stopping criterion exceeds a threshold, wherein thestopping criterion is based on at least one of the parameters of the RL algorithm.37) The second network node (1200) according to claim 36, being configured toperform a method according to any of claims 9-16 and 29-33.38) A first network node (1100), for executing a reinforcement learning, RL,algorithm to perform a task in a communication network (1000), the RL algorithmbeing configured with multiple reward functions wherein each reward function ofthe multiple reward functions is associated with the task, the communication network (1000) comprising the first network node (1100) and a second networknode (1200), the first network node (1100) being configured to: -receive (301, 601) a request message, from the second network node(1200), comprising a request for an estimated policy corresponding toeach reward function of the multiple reward functions; -verify (302, 602) availability of the estimated policy corresponding to eachreward function of the multiple reward functions; and -transmit (304, 603) a response message, to the second network node(1200), comprising an indication regarding the availability of the estimated policy corresponding to each reward function of the multiple rewardfunctions.39) The first network node (1100) according to claim 38, being configured to performa method according to any of claims 18-22 and 29-33.40) A second network node (1200), for executing a reinforcement learning, RL,algorithm to perform a task in a communication network (1000), the RL algorithmbeing configured with multiple reward functions wherein each reward function ofthe multiple reward functions is associated with the task, the communication network comprising a first network node (1100) and the second network node(1200), the second network node (1200) being configured to:- transmit (401, 601) a request message, to the first network node (1100),comprising a request for an estimated policy corresponding to each rewardfunction of the multiple reward functions; and -receive (402, 603) a response message, from the first network node(1100), comprising an indication regarding availability of the estimated policy corresponding to each reward function of the multiple rewardfunctions.41) The second network node (1200) according to claim 40, being configured toperform a method according to any of claims 24-33.42) A computer program (1101), comprising instructions which when run on aprocessor (1102) of a first network node (1100), causes the first network node (1100) to perform a method according to any of claims 1-7, claims 17-22 orclaims 29-33.43) A computer program product (1104) which comprises a computer readablestorage medium (1103) on which a computer program according to claim 42 isstored.44) A computer program (1201), comprising instructions which when run on aprocessor (1202) of a second network node (1200), causes the second network node (1200) to perform a method according to any of claims 8-16 or claims 23-33.45) A computer program product (1204) which comprises a computer readablestorage medium (1203) on which a computer program according to claim 44 isstored.

Images