Optimizing adaptive online network slicing

Abstract

The present disclosure relates to adaptive online network slicing. The present disclosure provides a method comprising deriving, independently of an allocation of resources of a second slice of a base station, a first action to adapt an allocation of resources of a first slice of the base station to meet a first performance requirement; the method further comprising providing information about the first action. The method further comprises deriving, independently of the allocation of resources of the first slice, a set of second actions to adapt an allocation of resources of the second slice to meet a second performance requirement; the method further comprising providing information about the second actions; the method further comprising jointly optimizing the allocation of resources of the first slice and the allocation of resources of the second slice based on the information about the first action, the allocation of resources of the first slice, the information about the second actions, and the allocation of resources of the second slice to derive a modified first action and a modified second action; the method further comprising instructing the base station to apply the modified first action and the modified second action.

Description

Technical Field

[0001] This disclosure relates to adaptive online network sharding. abbreviation 3GPP Third Generation Partnership Project 5G / 6G (Fifth Generation / Sixth Generation) ADMM Alternating Direction Multiplier Method AI (Artificial Intelligence) AONS Adaptive Online Network Slicing BO Bayesian Optimization CSP (Communication Service Provider) CU-CP Central Unit - Control Plane CU-UP Central Unit - User Plane DL (Deep Learning) DNN (Deep Neural Network) DRL (Deep Reinforcement Learning) DSL Domain-Specific Language DU Distributed Unit E2E end-to-end GBO Global Bayesian Optimization gNB (Next Generation NodeB, Base Station) GP Gaussian process GPR Gaussian process regression ML Machine Learning MN mobile network MNO mobile network operator Near-RT NG-RAN (Next Generation RAN) Non-RT (Non-real-time) NR New Radio O-DU O-RAN Distributed Unit O-RAN Open RAN O-RU O-RAN radio unit QoS (Quality of Service) RAN (Radio Access Network) RIC RAN Intelligent Controller RL (Reinforcement Learning) RRM Radio Resource Management RTRIC Real-time RAN Intelligent Controller SLA (Service Level Agreement) SMO Service Management and Orchestration Background Technology

[0002] The O-RAN architecture, proposed by the O-RAN Consortium, paves the way for decomposed radio access networks coupled with robust software-defined control architectures. While the NG-RAN architecture pioneered the decomposition from monolithic gNBs to CU-CP, CU-UP, and DU, O-RAN further decomposes the DU into two distinct entities—the O-RAN DU (O-DU) and the O-RAN Radio Unit (O-RU).

[0003] O-RAN introduces novel network control functions and assurance frameworks specifically for the access network through a set of external network controllers: near real-time RAN Intelligent Controllers (near RT RICs) and non-real-time RAN Intelligent Controllers (non-RT RICs). Specifically, the near RT RIC operates on a timescale greater than 10 ms and exercises near real-time control over the underlying RAN on the E2 interface through an application called xApp. Key functions of the near RT RIC include policy parsing and enforcement from the non-RT RIC, and statistical data collection. On the other hand, as part of the Service Management and Orchestration (SMO) framework, the non-RT RIC supervises the control and optimization of RAN elements on a non-real-time basis (i.e., greater than 1 s). This control is enabled by an application called rApp, which provides policy-based guidance and enrichment information to the near RT RIC. The SMO (and therefore the non-RT RIC) interacts with both the RAN and the near RT RIC using the O1 interface. Alternatively, the non-RT RIC can interact with the near RT RIC using the A1 interface. Summary of the Invention

[0004] The purpose of this invention is to improve upon existing technologies.

[0005] According to a first aspect, an apparatus is provided, comprising: a first component for deriving a set of first actions, independent of the current allocation of resources of a second slice of a base station, to adapt the current allocation of resources of the first slice of the base station to meet a first performance requirement on the first slice, the second slice being different from the first slice; a first component for providing information about the set of first actions; a second component for deriving a set of second actions, independent of the current allocation of resources of the first slice of the base station, to adapt the current allocation of resources of the second slice of the base station to meet a second performance requirement on the second slice; a second component for providing information about the set of second actions; a component for jointly optimizing the allocation of resources of the first slice and the allocation of resources of the second slice based on the information about the set of first actions, the current allocation of resources of the first slice, the information about the set of second actions, and the current allocation of resources of the second slice, to derive a modified set of first actions to adapt to the current allocation of resources of the first slice and a modified set of second actions to adapt to the current allocation of resources of the second slice; and a component for instructing a base station to apply the modified set of first actions to the current allocation of resources of the first slice and to apply the modified set of second actions to the current allocation of resources of the second slice.

[0006] The apparatus may further include at least one of the following: a first component for preventing the base station from applying the set of first actions to the current allocation of resources of the first slice; or a second component for preventing the base station from applying the set of second actions to the current allocation of resources of the second slice.

[0007] The component used to derive the set of the first actions can be configured to derive the set of the first actions by considering at least one of the following: constraints of the first slice without considering the second slice, or slice-specific costs of resource allocation of the first slice.

[0008] The component used to derive the set of the second action can be configured to derive the set of the second action by considering at least one of the following: constraints of the second slice without considering the first slice, or slice-specific costs of resource allocation of the second slice.

[0009] The component used for joint optimization can be configured to jointly optimize the allocation of resources in the first slice and the allocation of resources in the second slice, taking into account at least one of the total capacity of base station resources or the total cost of the operator for the base station.

[0010] The resources of the first slice and the resources of the second slice can be the radio resources of the base station.

[0011] The first component used for derivation can be a first learning agent. The second component used for derivation can be a second learning agent.

[0012] The first component and the second component for derivation can be configured to: derive a set of first actions and a set of second actions as subordinate parts of the alternating direction multiplier method, and the component for joint optimization can be configured to: derive a set of modified first actions and a set of modified second actions as the dominant part of the alternating direction multiplier method. According to a second aspect, a method is provided, the method comprising: deriving a set of first actions, independent of the current allocation of resources of a second slice of a base station, to adapt the current allocation of resources of the first slice of the base station to meet a first performance requirement on the first slice, the second slice being different from the first slice; providing information about the set of first actions; deriving a set of second actions, independent of the current allocation of resources of the first slice of the base station, to adapt the current allocation of resources of the second slice of the base station to meet a second performance requirement on the second slice; providing information about the set of second actions; jointly optimizing the allocation of resources of the first slice and the allocation of resources of the second slice based on the information about the set of first actions, the current allocation of resources of the first slice, the information about the set of second actions, and the current allocation of resources of the second slice, to derive a modified set of first actions to adapt to the current allocation of resources of the first slice and a modified set of second actions to adapt to the current allocation of resources of the second slice; instructing the base station to apply the modified set of first actions to the current allocation of resources of the first slice and to apply the modified set of second actions to the current allocation of resources of the second slice.

[0013] The method may further include at least one of the following: prohibiting the base station from applying the set of first actions to the current allocation of resources in the first slice; or prohibiting the base station from applying the set of second actions to the current allocation of resources in the second slice.

[0014] The set of first actions can be derived by considering at least one of the following: constraints of the first slice without regard to the second slice, or slice-specific costs of resource allocation of the first slice.

[0015] The set of second actions can be derived by considering at least one of the following: constraints of the second slice without regard to the first slice, or slice-specific costs of resource allocation in the second slice.

[0016] The allocation of resources in the first slice and the allocation of resources in the second slice are jointly optimized, taking into account at least one of the total capacity of base station resources or the total cost of the operator for the base station.

[0017] The resources of the first slice and the resources of the second slice can be the radio resources of the base station.

[0018] The set of first actions can be derived from the first learning agent. The set of second actions can be derived from the second learning agent.

[0019] The set of the first action and the set of the second action can be derived from the subordinate part of the alternating direction multiplier method, and the modified set of the first action and the modified set of the second action can be derived from the dominant part of the alternating direction multiplier method.

[0020] The second approach could be a method for optimizing adaptive network slicing.

[0021] According to a third aspect, a computer program product including an instruction set is provided, which, when executed on a device, is configured to cause the device to perform the method according to the second aspect. The computer program product may be embodied in a computer-readable medium or directly loaded into a computer.

[0022] According to some example embodiments, at least one of the following advantages can be achieved: • Performance can be optimized at both the per-slice level and the system level; This solution can handle strong time variables; • The solution is scalable; This solution can be implemented on top of existing O-RAN implementations. Attached Figure Description

[0023] Further details, features, objects, and advantages will be apparent from the following detailed description of preferred exemplary embodiments, taken in conjunction with the accompanying drawings, wherein:

[0024] Figure 1 An AONS system deployed within an O-RAN network according to some example embodiments is shown;

[0025] Figure 2 The workflow is shown according to some example embodiments;

[0026] Figure 3 The design of AONS xApp according to some example embodiments is shown;

[0027] Figure 4 The design of the AONS rApp according to some example embodiments is shown;

[0028] Figure 5 A distributed adaptive online learning algorithm according to some example embodiments is shown;

[0029] Figure 6 Pseudocode for an adaptive online learning algorithm according to some example embodiments is shown;

[0030] Figure 7 An apparatus according to an example embodiment is shown;

[0031] Figure 8 A method according to an example embodiment is shown;

[0032] Figure 9 An apparatus according to an example embodiment is shown. Detailed Implementation

[0033] Certain exemplary embodiments will be described in detail below with reference to the accompanying drawings, wherein features of the exemplary embodiments may be freely combined with each other unless otherwise described. However, it should be clearly understood that the description of certain exemplary embodiments is given by way of example only and is in no way intended to be construed as limiting this disclosure to the details disclosed.

[0034] Furthermore, it should be understood that the apparatus is configured to perform the corresponding method, although in some cases only the apparatus or only the method is described.

[0035] Several use cases have been proposed for O-RAN, among which network slicing is one. Network slicing provides multiple logical networks—slices—to be provisioned over a common physical infrastructure, allowing networks to be delivered "per service," where each slice can have a different Service Level Agreement (SLA) than the others. While network slicing spans various domains of the network—access, transport, core, edge, and central cloud—in the context of O-RAN, network slicing primarily refers to RAN slices. For example, different slices can be used for different types of services, such as V2X, IoT, enterprise, or fixed wireless access (FWA) services. The importance of network slicing lies in improving network resource utilization and enabling on-demand customization of network services, thereby supporting more innovative applications and business models. Therefore, communication service providers (CSPs) consider RAN slice resource management strategies when seeking to optimize radio resource utilization and improve network monetization capabilities.

[0036] Similarly, machine learning (ML)-based radio resource management (RRM) techniques can be adopted for resource-efficient slicing. In particular, focusing on sub-second near-RT operations, near-RT RIC has emerged as a key asset for training and deploying xApp-based ML models for robust slice-aware RRM with several state-of-the-art solutions across industry and academia.

[0037] Any resource management solution should consider the dynamic nature of network slicing. This dynamic nature has several aspects: for example, services within different slices can change rapidly, but the number of slices in the network, the number of users within a slice, and channel conditions and mobility patterns are also time-varying. Therefore, changes in slice dynamics will be considered for resource allocation and SLA fulfillment. Within the context of slice dynamics, changes in the number of slices in the system can be considered in more detail. Such changes can be triggered by changes in network policies and as a result of user mobility. For example, the provision of new network services by a network operator may lead to the addition of new slices to the network, while deconfiguration will result in the removal of slices. Furthermore, RAN slices are often heterogeneous; that is, for example, each slice may have different latency, throughput, and reliability requirements.

[0038] Traditional RRM frameworks typically use a uniform optimization algorithm for all slices, which fails to adapt well to the dynamics of individual slice users. For example, if a Bayesian optimization algorithm is used for all slices, its surrogate model represents the overall situation of all slices, resulting in overgeneralization and a loss of the ability to accurately fit individual slices. Furthermore, many algorithms employ offline methods, which have the drawback of not being able to respond to dynamic network changes in real time. The network environment is time-varying, and current optimization results may be applicable to previous network states but may no longer be applicable to current network conditions. This lag can lead to inappropriate resource allocation, impacting network performance and user experience. Additionally, offline methods struggle to handle sudden surges in traffic and emergencies, lacking the ability to adjust in real time.

[0039] A hierarchical framework, as illustrated in some example embodiments, may be a more suitable approach. This framework can combine machine learning (ML) and mathematical optimization. ML's advantages in optimization lie in its adaptability and ability to handle complex nonlinear problems, learning from historical data and predicting future trends. Mathematical optimization's advantages lie in its robust theoretical foundation, providing deterministic solutions and performance guarantees. In this hierarchical framework, online learning using ML can handle intra-slice optimization (such as user behavior prediction and resource demand estimation), while mathematical optimization methods can handle resource allocation conflicts between slices (inter-slice optimization). This mathematical framework can include ADMM (Alternating Direction Multipliers), which, for example, can effectively decompose large-scale optimization problems into smaller subproblems. Other suitable approaches may include robust optimization that can handle uncertainty; distributed optimization suitable for large-scale networks; and multi-objective optimization that can balance multiple performance metrics. This hybrid approach combines the advantages of both methods, providing a more flexible, efficient, and reliable network slice management solution.

[0040] However, using online ML learning for intra-slice optimization is not mandatory; mathematical methods can be used instead. Similarly, using mathematical methods for inter-slice optimization is not mandatory; online ML learning can be used instead.

[0041] Some example implementations address the challenges of designing adaptive online network slicing systems for the O-RAN ecosystem. Specifically, they address how to effectively coordinate multiple xApps and their associated strategies across different network slices under the dynamic characteristics of a given network environment. This system faces three key technical challenges: • Slicing strategies need to be learned and adapted online because they involve various rules and optimization results that change over time. Second, network slicing is unstable and involves frequent connections and disconnections, which complicates consistency management. Third, system-wide constraints (such as resource priorities and capacity limits) are constantly changing.

[0042] Some example implementations provide a coordinator that can adapt to these dynamic conditions while minimizing costs, improving resource utilization, and maintaining system constraints. This is particularly challenging because the number of slices, constraints, and policies are variable, requiring a flexible and robust orchestration solution.

[0043] Some example implementations provide an Adaptive Online Network Slicing (AONS) system for large-scale, practical network slice management. Unlike traditional network slice orchestration frameworks that rely on static configuration, AONS systems employ online learning to dynamically sense network state with multiple agents and coordinators to adjust slice-specific actions. This approach enhances the system's adaptability and scalability in time-varying network environments.

[0044] Specifically, some example embodiments deploy an adaptive online orchestration framework that includes multiple learning agents and a coordinator. These components achieve different optimization objectives; the learning agents handle intra-slice constraints, such as variable requirements for network performance within a slice, while the coordinator handles system-wide issues like capacity constraints on radio resources. Instead of using a single model to manage all slices, the interaction between the agent and coordinator for each slice allows for more precise allocation of network resources, thus improving resource utilization.

[0045] In some example implementations, learning agents are used to manage in-slice optimization challenges. These agents are primarily implemented in xApps near-RT RICs. Various policy options may be available for these agents, including empirical rules, ML or deep learning (DL) optimization, and mathematically based optimization. The agent takes network state as input and periodically acquires network performance data to determine the optimal resource allocation strategy that satisfies slice constraints.

[0046] Furthermore, some example implementations employ a coordinator to handle inter-slice issues. This coordinator can reside within the rApp and can leverage distributed optimization frameworks such as the ADMM method. Communication between the coordinator and the agent can occur via the O-RANO1 interface, with auxiliary variables facilitating iterative action updates that instruct resource allocation for each slice.

[0047] Some aspects of some example embodiments can be summarized as follows: • An end-to-end network slicing orchestration system that can adapt to network dynamics more effectively than existing methods. • A distributed resource orchestration framework that can adapt to time-varying networks. • A method that optimally addresses constraints within slices and coordination between slices. • Adaptive online network slicing orchestration system, which is built to be fully compatible with the O-RAN framework and tailored for dynamic network environments.

[0048] The AONS system is applicable not only to the context of O-RAN, but also to all cellular networks, and does not exclude any specific deployment.

[0049] The overall architecture of the adaptive online network slicing system, followed by its operational workflow, key components, algorithms, and implementation aspects, are described below.

[0050] Overall System Architecture

[0051] According to some example embodiments, such as Figure 1As shown, the Adaptive Online Network Slicing System (AONS) consists of several modules: a coordinator typically deployed within the rApp (i.e., AONS rApp) and learning agents typically deployed within the xApp (i.e., AONS xApp). The AONS system deploys an interface between the coordinator and the learning agents. Within each slice, one or more xApp-level learning agents are responsible for solving slice-specific challenges, such as meeting performance metrics requirements for the corresponding slice (e.g., latency, frames per second, throughput). These agents utilize techniques such as AI-based training, regression models, and mathematical frameworks to implement diverse strategies to solve the problem. The rApp-level coordinator is responsible for managing system-wide constraints, including radio resource capacity and mobile operator costs. It typically utilizes distributed algorithms, such as the Alternating Directional Multiplier Method (ADMM), to facilitate interaction with the agents. Notably, the agents autonomously avoid direct information exchange or consideration of inter-agent constraints during problem-solving. The coordinator centrally processes the optimization results from the agents, which may involve iterative practice and refinement of the process.

[0052] Operational Workflow

[0053] Figure 2 The operational workflow of a system according to some example embodiments is illustrated, describing general network slicing orchestration, and optimizations and collaboration between the AONS xApp agent and the AONS rApp coordinator. Figure 2 As shown, this process can be summarized as follows: • S1: The learning agents use historical data, such as action-performance pairs, to train and update local models. These agent models serve the various agents. • S2: The agent makes decisions based on the latest model and its policy, combining empirical rules, ML / DL optimization, and mathematical solutions. Decisions consist of a set of actions, where one action may modify some parameters or not modify (maintain) some parameters. Each slice's policy is independent of the states and policies of other slices, providing the optimal action for the corresponding slice that satisfies its specific constraints. • S3: Instead of instructing the base station to perform actions based on its derived decisions, the agent relays its decisions (action sets) to the coordinator, for example, via the O1 interface. • S4: Considering system-level constraints such as radio resource capacity, the coordinator receives decisions from the agent and uses one or more distributed algorithms to achieve an optimal solution for the system under these constraints. This optimization may potentially involve multiple iterations, such as ADMM. •S5: The near-RT RIC then receives the action set for the slice from the coordinator and publishes these actions to network elements (e.g., DU, gNB, etc.). • S6: Near RT RIC also monitors slice data, focusing on performance metrics such as latency, frames per second, and throughput. Based on the performance results of actions within the network, a set of real-time action-performance pairs are generated and fed back to the first step in a loop to update the agent model.

[0054] Through the above process, the system can dynamically orchestrate network resources to enhance resource utilization while minimizing CSP operating costs. The xApp-level agent strategy continuously learns and updates online to address slice-specific issues, while the rApp-level coordinator orchestrates global solutions for system-wide problems.

[0055] Key components and interactions

[0056] The AONS system comprises two main components—AONS xApp and AONS rApp—which interact with each other in a layered manner. This architecture facilitates efficient resource allocation, dynamic adaptation, and optimization within network slicing. Detailed descriptions of each component are provided below, highlighting their roles, functions, and interactions within the overall system.

[0057] AONS xApp:

[0058] xApp can be combined with learning agents that perform online learning to solve slice-specific problems. For example... Figure 3 As shown, the learning agent receives a set of states (input states), including the number of resources allocated to the slice, priority values, and other scheduling algorithm parameters. The agent then makes decisions based on a strategy derived from local data and the learned model. The agent may also consider an in-slice cost model. The decisions include a set of actions reported to the coordinator (AONS rApp). The set of actions derived by the AONS xApp (learning agent) can be considered a proposal to the coordinator. The xApp does not instruct network elements (base stations) to execute the proposed set of actions.

[0059] Figure 3 The following (logical) sub-components of AONS xApp are shown: • State Set: The state set belongs to the state space for each xApp strategy. This state space covers slice-related information, such as the number of resources allocated, priority value, the number of resources in other slices, and the aggregate priority value of other slices. • Action Set: The action set belongs to the action space. The action space for each xApp includes variables such as the number of resources allocated, priority values, weights, and slice-specific scheduling algorithm parameters. • Cost Model: From the perspective of a mobile network operator (MNO), the operating cost of running a slice depends primarily on resource usage. This cost model focuses mainly on total resource consumption while considering slice-specific constraints, which may include time-varying requirements for certain metrics. • Strategy: Strategies can include empirical rules, machine learning (ML) / deep learning (DL) optimization algorithms, or mathematical optimization techniques. The flexibility of strategy design enables tailored solutions for different slicing requirements.

[0060] AONS rApp:

[0061] The coordinator (AONS rApp) is designed to iteratively resolve inter-slice and system-wide issues. For example... Figure 4 As shown, the coordinator takes the results from the previous (i-1)th iteration (i.e., the optimization variables and auxiliary variables) as input. Since the coordinator aims to solve the inter-slice cost problem, the inter-slice cost model (= system-wide cost model) is also input. Simultaneously, the coordinator considers global constraints and returns the optimization results of the next round (i-th iteration) to the optimization variables. In other words, the coordinator receives a set of actions from the xApp agent and combines them with auxiliary variables, which are typically introduced by a distributed optimization algorithm (such as the alternating direction multiplier method or the Lagrange method). As a result, the coordinator determines the action set. This action set is provided to the xApp of the slice, instructing the corresponding network elements (e.g., gNBs) to apply these actions.

[0062] like Figure 4 As shown, rApp can contain the following sub-components: • Optimization variables: The main optimization variables are the action sets received from each slice. These actions represent the optimal decisions made by the agents in solving their respective problems, and the coordinator can refine and iterate these decisions. • Auxiliary Variables: The coordinator uses auxiliary variables to help refine and adjust optimization variables. These auxiliary variables are typically derived from distributed optimization algorithms, facilitating iterative updates of the optimization variables. Auxiliary variables help decompose complex optimization problems into simpler subproblems, which often converge to a stable point or can be solved in parallel. This approach reduces the complexity of solving the original problem and is suitable for large-scale optimization problems. For example, ... Figure 6 As shown in line 10, the ADMM method introduces an auxiliary variable zi to help optimize the main optimization variable. • Cost Model: The coordinator's goal is to minimize the total cost of the MNO, which differs from the agent-level cost model. This system-wide cost model can take into account the radio resource capacity constraints and the dynamic nature of slices (i.e., their connection and disconnection), which therefore affects the associated agents.

[0063] Layered xApp-rApp operation and interaction:

[0064] The AONS system provides a two-tiered optimization approach. xApp focuses on intra-slice optimization, while rApp handles inter-slice optimization. In detail: • Intra-slice optimization: xApp horizontal agents optimize intra-slice problems by considering inputs (such as the amount of resources allocated and priority values) without regard to information about other slices. The output is a set of actions that satisfy given conditions. Various optimization strategies can be employed, including AI-based methods (e.g., DNN, DRL), empirical rules, or mathematical models. This optimization problem is characterized by time-varying constraints and the potential for slice disconnection or reconnection, requiring stable and adaptive optimization strategies. • Inter-slice optimization: The rApp horizontal coordinator addresses the inter-slice problem by refining the optimization results from the agents. Because agents optimize independently based on local data and objectives, their resource allocation can be suboptimal from a system-wide perspective, potentially violating network capacity constraints. The coordinator uses auxiliary variables to refine these results to update the action set, ensuring system-wide feasibility and optimality. • Constraints: Some example implementations can adapt to any type of constraint, from explicit mathematical formulas—both linear and nonlinear—to more complex implicit constraints that do not conform to standard mathematical expressions. For the latter, where the constraints are unknown but queryable, fitting techniques (such as Gaussian processes (GP)) can be used to model these unknown constraints. These methods excel at modeling and approximating unknown constraints. The fitted model is then incorporated into the optimization objective as a penalty term.

[0065] In some example implementations, information exchange between the coordinator and the agent can be facilitated through the O1 interface conforming to the O-RAN architecture.

[0066] This comprehensive design ensures that the AONS system can effectively manage resource allocation, adapt to dynamic network conditions, and optimize performance across multiple slices. The combination of local agent optimization and global coordination provides a robust and flexible solution. This solution can also be adapted to the challenges of next-generation network slicing. The interfaces between components can be standardized.

[0067] Distributed adaptive online learning algorithm

[0068] Additional technical details are described regarding a distributed adaptive online learning algorithm deployed across AONS xApp and rApp according to some example embodiments. For example... Figure 5As shown, the algorithm employs a distributed optimization approach, which decomposes the problem into multiple subproblems. For example, we use the ADMM technique to decouple the problem into a dominant problem and several subordinate problems. Within this framework, the demand for radio resources for each network slice constitutes a subordinate problem, while the dominant problem focuses on optimizing auxiliary variables to coordinate the solutions to the subordinate problems.

[0069] Within this algorithm, each learning agent (at AONS xApp) locally handles its slice-specific optimizations, searching for an optimal solution within a defined scope. Simultaneously, a coordinator (at AONS rApp) ensures that these per-slice local optima can work in coordination across the entire system to achieve an overall optimal result. The optimization challenge is processed iteratively, resolving dependent and dominant problems until the cost of the CSP converges to an efficient solution. This decomposition and coordination approach not only enhances the efficiency of problem-solving but also improves the algorithm's scalability and flexibility when handling large-scale network systems.

[0070] After the problem is decoupled through a distributed optimization framework, existing optimization tools (such as MATLAB's CVX toolbox) can be utilized to solve the dominant problem. For the subordinate problems, which may involve unknown, dynamic, and time-varying constraints, the learning agent can employ an agent model to learn these network characteristics online—for example, using Gaussian process regression (GPR). By iteratively learning the relationship between wireless resource quantity and network characteristics, the agent continuously improves the accuracy of its model. Therefore, each per-slice optimization challenge is treated as a black-box problem and solved using black-box optimization methods (such as Bayesian optimization). This approach aims to minimize the cost of each slice under given constraints, thereby optimizing the overall system efficiency and cost-effectiveness.

[0071] Figure 5 The process is shown below: 0. Begin. 1. Initially, a warm start is performed on a surrogate model with a small amount of historical data (such as resource-performance pairs) to quickly initiate the model for the first iteration of the distributed optimization algorithm. Here, the ADMM (Alternating Direction Multiplier) algorithm can be used. 2. Learning agents can query the system status and train agent models accordingly. 3. Subsequently, for each slice problem, the learning agent solves its corresponding subordinate problem, which may involve unknown or complex expressions, potentially using deep neural networks or optimization algorithms such as DRL (Deep Reinforcement Learning), DNN (Deep Neural Network), or Bayesian optimization. 4. The coordinator handles the dominant problem, essentially coordinating solutions across different slices. Solutions may involve tools like those in the CVX toolbox, updating auxiliary variables within the ADMM algorithm framework to coordinate the global solution. 5. Calculate the cost of CSP; if the cost is stable and less than a predetermined threshold, terminate the iteration. Otherwise, based on the latest solution obtained from step 4, query the performance of the corresponding system, and repeatedly practice and update the proxy model for the next iteration. 6. Repeat the above process (2.-5.) until the requirements are met and the resources to be allocated to each slice are output. 7. End.

[0072] Figure 6 Example pseudocode for an algorithm according to some example embodiments is shown. As shown, the algorithm input includes... U h , H , Q i and p ,in U h Indicates the price of wireless resources. H This represents the total amount of radio resources available to the CSP. Q i Let represent the network performance requirements of the user in the i-th slice, and p These are positive constants related to the Lagrange equation. Output x i This represents the amount of radio resources allocated to each slice, where w i Indicates weight, s i This represents the state of each slice. During the execution of the algorithm, additional variables (such as...) z i and y i ) are introduced as auxiliary variables to decompose the original cost optimization problem using the ADMM method. L represents the Lagrangian function used for decoupling the problem.

[0073] For the surrogate model, Gaussian Process Regression (GPR) is used. GPR is a sample-efficient, nonparametric Bayesian regression method based on Gaussian processes. To address the dependency problem per slice, we utilize Bayesian Optimization (BO), a global optimization method suitable for optimizing black-box functions. This approach is particularly appropriate given the high cost of evaluation and lack of analytical expressions in dependency problem constraints. For the dominance problem, the CVX Toolbox is used. The CVX Toolbox is a toolbox specifically designed for convex optimization problems. It provides a method for describing and solving various convex optimization problems using a domain-specific language (DSL), which is very convenient for solving the dominance problem.

[0074] Implementation and Verification within the O-RAN Ecosystem

[0075] To verify the performance benefits claimed by the AONS system according to the example embodiment, the framework was deployed on an over-the-air experimental test platform. This test platform included: O-RAN compliant base stations, a 5G core network, and near-RT and non-RT RICs. Multiple network slices were deployed, each hosting users running multiple applications such as browsers, video streaming, and speed tests, creating a comprehensive end-to-end network environment.

[0076] Each slice contains learning agents capable of running heterogeneous or identical policies, implemented as part of the AONS xApp. These policies include Bayesian optimization, DRL agents, and DNNs. Given the diversity of applications within a slice, each application requires unique network performance characteristics. However, the agents periodically receive near-RT network performance metrics as input to the xApp. These metrics then inform the policy, allowing for dynamic adjustments to the output to optimize actions.

[0077] The rApp coordinator refines the actions derived from each xApp under the constraint of the total available radio resources. Specifically, the coordinator uses the ADMM technique, utilizing dual and auxiliary variables to update the actions. These refined actions are then passed to the agent via the O1 interface for the next round of optimization. The interaction between the coordinator and the agent is iterative, updating actions to satisfy agent-specific constraints while also aiming to minimize the cost of CSP.

[0078] Furthermore, unstable connections were simulated during system operation by disconnecting and reconnecting slices. This tested not only the system's ability to handle interruptions but also its adaptability to dynamic conditions such as the addition of new slices. These scenarios are essential for evaluating the robustness and scalability of network slicing solutions for AONS systems within continuously evolving O-RAN environments.

[0079] The performance of the AONS system was compared with several state-of-the-art solutions (Global Bayesian Optimization (GBO), ATLAS, and Exhaustive Search (ExSearch)). The results show that the AONS system achieves the lowest total cost while also achieving the highest normalized performance even in dynamic environments, where the number of slices and the number of users per slice change throughout the evaluation process.

[0080] Figure 7 An apparatus according to an example embodiment is shown. The apparatus may be an optimizer (such as an AONS system) or a component thereof. Figure 8 A method according to an example embodiment is shown. Figure 7 The device can perform Figure 8 The method, but not limited to this method. Figure 8 The method can be derived from Figure 7 The device performs the action, but is not limited to the device performing the action.

[0081] The device includes a first component 110 for derivation, a first component 120 for providing, a second component 130 for derivation, a second component 140 for providing, a component 150 for joint optimization, and a component 160 for instruction. The first component 110 for derivation, the first component 120 for providing, the second component 130 for derivation, the second component 140 for providing, the component 150 for joint optimization, and the component 160 for instruction can respectively be a first derivation component, a first providing component, a second derivation component, a second providing component, a joint optimization component, and an instruction component. The first component 110 for derivation, the first component 120 for providing, the second component 130 for derivation, the second component 140 for providing, the component 150 for joint optimization, and the component 160 for instruction can respectively be a first derivator, a first provider, a second derivator, a second provider, an optimizer, and an instruction generator. The first component 110 for derivation, the first component 120 for provisioning, the second component 130 for derivation, the second component 140 for provisioning, the component 150 for joint optimization, and the component 160 for instructions can be respectively a first derivation processor, a first provisioning processor, a second derivation processor, a second provisioning processor, a joint optimization processor, and an instruction processor.

[0082] The first component 110, used for derivation, derives a set of first actions (S110). This set of first actions is used to adapt the current allocation of resources in the first slice of the base station to meet the first performance requirements on the first slice. The first component 110, used for derivation, also derives a set of first actions independent of the current allocation of resources in the second slice of the base station. The second slice is different from the first slice.

[0083] The first component 120 for providing information provides information about a set of first actions (S120). For example, the first component 120 for providing information may provide information about a set of first actions to the component 150 for joint optimization. The information about the set of first actions may indicate the first action of that set of first actions.

[0084] The second component 130 used for derivation derives a set of second actions (S130). This set of second actions is used to adapt the current allocation of resources in the second slice of the base station to meet the second performance requirements on the second slice. The second component 130 used for derivation derives a set of second actions independent of the current allocation of resources in the first slice of the base station.

[0085] The second component 140 for providing information provides information about the set of second actions (S140). For example, the second component 140 for providing information may provide information about the set of second actions to the component 150 for joint optimization. The information about the set of second actions may indicate the second action of the set of second actions.

[0086] The component 150 for joint optimization jointly optimizes the allocation of resources in the first slice and the allocation of resources in the second slice based on information about the set of first actions, the current allocation of resources in the first slice, information about the set of second actions, and the current allocation of resources in the second slice, in order to derive a modified set of first actions to adapt to the current allocation of resources in the first slice, and a modified set of second actions to adapt to the current allocation of resources in the second slice (S150).

[0087] The component 160 for instructions instructs the base station to apply a set of modified first actions to the current allocation of resources in the first slice, and to apply a set of modified second actions to the current allocation of resources in the second slice (S160).

[0088] Figure 9 An apparatus according to an example embodiment is shown. The apparatus includes: at least one processor 810 and at least one memory 820 storing instructions, wherein, when executed by the at least one processor 810, the apparatus causes to at least perform the following... Figure 8 And related descriptive methods.

[0089] "Optimization" does not necessarily mean finding the absolute optimal solution. It can include finding a solution that is better than a certain threshold, or finding a solution that does not exceed the threshold effort required to find a solution.

[0090] Some example implementations can be implemented for base stations based on slicing. These implementations can be used for base stations in any generation of 3GPP or non-3GPP communication networks.

[0091] A message can be sent from one entity to another in one or more messages. Each of these messages can include additional (different) message entries.

[0092] The names of network elements, network functions, protocols, and methods are based on current standards or existing proposals. These names are not restrictive. For example, in other versions or technologies, the names of corresponding entities may be different from or the same as those in this specification.

[0093] Unless otherwise stated or made clear from the context, the statement that two entities are “different” means that they perform different functions. This does not necessarily mean that they are based on different hardware. That is, each entity described in this specification may be based on different hardware or may be based on the same hardware in whole or in part. Nor does it necessarily mean that they are based on different software. That is, each entity described herein may be based on different software or may be based on the same software in whole or in part. Each entity described in this specification may be deployed in the cloud.

[0094] Based on the above description, it should be obvious that, for example, the exemplary embodiments provide a slice management system or elements thereof, or a resource allocation optimizer or elements thereof, a device embodying the system, a method for controlling and / or operating the system, and a plurality of computer programs for controlling and / or operating the system, and a medium carrying such a plurality of computer programs and forming a plurality of computer program products.

[0095] By way of non-limiting example, any implementation of the blocks, apparatus, systems, techniques, or methods described above includes implementations as hardware, software, firmware, special-purpose circuitry or logic, general-purpose hardware or controllers or other computing devices, or combinations thereof. Each entity described in this specification can be represented in the cloud.

[0096] It should be understood that the above description represents what are currently considered preferred exemplary embodiments. However, it should be noted that the description of preferred exemplary embodiments is given by way of example only, and various modifications may be made without departing from the scope of this disclosure as defined by the appended claims.

[0097] Unless otherwise specified, the terms "first X" and "second X" include options where "first X" is the same as "second X," and options where "first X" is different from "second X." As used herein, phrases such as "at least one of the following: " and "at least one of " and similar wording, where the list of two or more elements is connected by "and" or "or," mean at least any one of the elements, or at least any two or more of the elements. Unless otherwise specified, for example, with the use of "otherwise" or "or in the alternative," the term "or" means a non-exclusive "or."

Claims

1. An apparatus for network slicing, comprising: At least one processor and at least one memory, the at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to at least: The set of first actions is derived from the current allocation of resources in the second slice, which is independent of the base station, to adapt to the current allocation of resources in the first slice of the base station to meet the first performance requirements on the first slice, wherein the second slice is different from the first slice; Provide information about the set of actions; A set of second actions is derived independently of the current allocation of resources in the first slice of the base station to adapt to the current allocation of resources in the second slice of the base station to meet the second performance requirements on the second slice. Provide information about the set of actions; Based on the information about the set of first actions, the current allocation of resources in the first slice, the information about the set of second actions, and the current allocation of resources in the second slice, the allocation of resources in the first slice and the allocation of resources in the second slice are jointly optimized to derive a modified set of first actions to adapt to the current allocation of resources in the first slice and a modified set of second actions to adapt to the current allocation of resources in the second slice; The instruction specifies that the base station applies the set of modified first actions to the current allocation of resources in the first slice, and applies the set of modified second actions to the current allocation of resources in the second slice.

2. The apparatus of claim 1, wherein the at least one processor and the at least one memory storing instructions, the instructions, when executed by the at least one processor, further cause the apparatus to: The base station is prohibited from applying the set of first actions to the current allocation of resources in the first slice; or The base station is prohibited from applying the set of second actions to the current allocation of resources in the second slice.

3. The apparatus according to any one of claims 1 and 2, wherein at least one of the following: The set of methods for deriving the first action considers at least one of the following: Without considering the constraints of the first slice, or The slice-specific cost of the allocated resources of the first slice; or The set of methods for deriving the second action considers at least one of the following: Without considering the constraints of the first slice, or The slice-specific cost of the allocation of the resources in the second slice; or The joint optimization considers at least one of the total capacity of the base station's resources or the total cost to the operator for the base station, to jointly optimize the allocation of the resources of the first slice and the allocation of the resources of the second slice.

4. The apparatus according to any one of claims 1 to 2, wherein the resources of the first slice and the resources of the second slice are radio resources of the base station.

5. The apparatus according to any one of claims 1 to 2, wherein at least one of the following: The set of actions to derive the first action is executed by the first learning agent; or The set of actions for deriving the second action is executed by the second learning agent.

6. The apparatus according to any one of claims 1 to 2, wherein the set of derivation of the first action and the set of derivation of the second action are derivations as subordinate parts of the alternating direction multiplier method, and the joint optimization is a dominant part of the alternating direction multiplier method to derive the modified set of the first action and the modified set of the second action.

7. A method for network slicing, the method comprising: The set of first actions is derived from the current allocation of resources in the second slice, which is independent of the base station, to adapt to the current allocation of resources in the first slice of the base station to meet the first performance requirements on the first slice, wherein the second slice is different from the first slice; Provide information about the set of actions; A set of second actions is derived independently of the current allocation of resources in the first slice of the base station to adapt to the current allocation of resources in the second slice of the base station to meet the second performance requirements on the second slice. Provide information about the set of actions; Based on the information about the set of first actions, the current allocation of resources in the first slice, the information about the set of second actions, and the current allocation of resources in the second slice, the allocation of resources in the first slice and the allocation of resources in the second slice are jointly optimized to derive a modified set of first actions to adapt to the current allocation of resources in the first slice and a modified set of second actions to adapt to the current allocation of resources in the second slice; The instruction specifies that the base station applies the set of modified first actions to the current allocation of resources in the first slice, and applies the set of modified second actions to the current allocation of resources in the second slice.

8. The method of claim 7, further comprising at least one of the following: The base station is prohibited from applying the set of first actions to the current allocation of resources in the first slice; or The base station is prohibited from applying the set of second actions to the current allocation of resources in the second slice.

9. The method according to any one of claims 7 or 8, wherein at least one of the following: The set of the first action is derived considering at least one of the following: Without considering the constraints of the first slice, or The slice-specific cost of the allocated resources of the first slice; or The set of second actions is derived considering at least one of the following: Without considering the constraints of the first slice, or The slice-specific cost of the allocation of the resources in the second slice; or The allocation of resources in the first slice and the allocation of resources in the second slice are jointly optimized, taking into account at least one of the total capacity of the resources of the base station or the total cost of the operator for the base station.

10. The method according to any one of claims 7 to 8, wherein the resources of the first slice and the resources of the second slice are radio resources of the base station.

11. The method according to any one of claims 7 to 8, wherein at least one of the following: The set of first actions is derived by the first learning agent; or The set of the second actions is derived by the second learning agent.

12. The method according to any one of claims 7 to 8, wherein the set of first actions and the set of second actions are derived as subordinate parts of the alternating direction multiplier method, and the modified set of first actions and the modified set of second actions are derived by the dominant part of the alternating direction multiplier method.

13. A computer program product comprising an instruction set, which, when executed on a device, is configured to cause the device to perform the method according to any one of claims 7 to 8.

14. The computer program product of claim 13, wherein the computer program product is embodied in a computer-readable medium or can be directly loaded into a computer.

Images