WO2023011371A1

WO2023011371A1 - Method and system for configuring a threshold value for a handover parameter of a wireless communication system

Info

Publication number: WO2023011371A1
Application number: PCT/CN2022/109191
Authority: WO
Inventors: Mehrtash MEHRABI; Walid Masoudimansour; Yingxue Zhang
Original assignee: Huawei Technologies Co., Ltd.
Priority date: 2021-07-31
Filing date: 2022-07-29
Publication date: 2023-02-09

Abstract

Methods and systems for determining a threshold value for a handover (HO) parameter of a wireless communication network comprising a center cell and a plurality of neighbor cells using a prediction model are provided. The prediction model is generated by selecting a subset of the plurality of neighbor cells, forming an augmented set of cells by using a feature transformation on the center cell to map the center cell to a latent space, and by augmenting the subset of the plurality of neighboring cells with a set of additional cells closest to the center cell in the latent space, dividing the augmented set of cells into groups, applying a permutation invariant function to each of the groups to generate an output for each of the groups, aggregating the outputs, and determining the threshold value for the HO parameter at least partially based on the aggregated outputs.

Description

METHOD AND SYSTEM FOR CONFIGURING A THRESHOLD VALUE FOR A HANDOVER PARAMETER OF A WIRELESS COMMUNICATION SYSTEM

Cross-Reference to Related Applications

This application claims priority to, and the benefit of, U.S. Provisional Application No. 63/228,087 filed on July 31, 2021, which is incorporated by reference herein in its entirety.

Technical Field

The present disclosure relates to wireless communications, and in particular to a method and system for configuring a threshold value for a handover parameter of a wireless communication system.

Background

In wireless communication networks, user equipments (UEs) establish radio connections with a service station (SS) of a wireless communication network. The service station (of which the antenna is often observable on top of high towers to maximize coverage) is called a base transceiver station (BTS) and handles speech encoding, encryption, multiplexing (time division multiple access (TDMA) , code division multiple access (CDMA) , frequency division multiple access (FDMA) , orthogonal frequency division multiplexing (OFDM) ) , and modulation/demodulation of the radio signals. Each BTS has a limited coverage area due to limited signal power (especially of a user equipment (UE) , since the UE also has to be able to send a signal to the BTS) . Also, the number of channels that can be dedicated to the UE is limited. Therefore, to solve these problems, a geographical area (such as a city) is often divided into several areas (hence the name “cellular network” ) , and a BTS serves this area while it is connected to other BTSes. This hierarchical structure makes it possible to cover (virtually) an unlimited geographical area and serve a very large number of users. With the recent developments in wireless mobile networks and new technologies, details in the network have changed. For example, wireless communication networks have moved from circuit-switching to an all-IP evolved packet core, the service stations are now called eNodeB, gNodeB, etc. The main concept of a cellular network, however, is still the same. Geographical areas are divided into smaller areas, with each being covered by one service station.

The main challenge with a network that uses such technique of dividing an area into several partitions is the handling of mobile UEs that move from an area covered by one SS to an area covered by another SS. Consider a UE served by one SS in a cell, and assume the UE starts moving toward the boundary of that cell. The signal can start to become weaker, and the channel quality may not be high enough for an acceptable service. In this case, the UE must be switched (handed over) to the new neighboring SS to maintain a proper level of service. This handover (HO) process must be done such that the UE does not experience noticeable interruption in service, and can continue all their sessions seamlessly.

There are many parameters that control the behavior of a wireless communication network. The HO parameters are among the most important of all as these HO parameters control the handling of edge UEs (UEs at the edge or boundary of a cell) . Proper configuration of these HO parameters contributes to providing smooth and uninterrupted service to the UE by SSs. This matter found even more significant importance with the rapid increase in the number of UEs in the past decade and also the type of the required service (demand for higher bandwidth, faster UEs such as self-driving cars which cross borders more often, etc. ) .

Table 1 lists some of the most important HO parameters that affect the HO process in a wireless communication network.

Event	Trigger HO when …
A1	the serving cell becomes better than a threshold
A2	the serving cell becomes worse than a threshold
A3	a neighboring cell becomes better than the serving cell by an offset
A4	a neighboring cell becomes better than a threshold

Table 1: Examples of HO parameters

Each of the HO parameters listed in Table 1 have to be met for a specific time (called time to trigger, TTT) for the HO event to trigger. This avoids acting upon brief changes that may be caused due to fluctuations of the signal in the environment. These HO parameters are configured in each SS, and improper values of these parameters can cause interruptions and delays in the service. For example, choosing a low/high value for one parameter may cause frequent HO events to trigger, translating to frequent interruptions in the service to a UE, while choosing high/low values for that parameter may not trigger the HO process at the proper time which can yield bad signal quality. Therefore, configuring these HO parameters is of utmost importance for the proper functioning of the wireless communication network.

The performance of a cellular network relies heavily on its parameter configurations, such as access control, handover, resource management parameters. As the number of mobile users grows rapidly, optimizing parameter configurations of cellular network is becoming increasingly important. Due to the dynamic nature of the environment, any employed algorithm for parameter configuration should be capable of tracking the non-stationary environment change (i.e., fluctuations of user numbers, network load, etc. ) , and configuring the parameters appropriately over time. Also, due to the diverse characteristics of cells across the network, the best parameter configuration for one cell may not be optimal for another. Consequently, developing an algorithm that can adapt to the temporal dynamics and cell diversity in real networks is important for parameter configuration.

A typical HO process works as follows. Consider a UE is moving between a cell serviced by a first service station (SS1) to a cell serviced by a second service station (SS2) . A reference signal is sent by an SS and a UE measures the signal strength of the reference signal and computes a reference signal received power (RSRP) . The UE periodically verifies if the value of the RSRP is higher than a threshold value for the RSRP of neighboring SSs. The HO process is triggered if the value of the RSRP of SS1 becomes lower than the value of the RSRP of one of the neighboring SSs (this is particular to the type of HO event as some HO events trigger in comparison to an absolute predetermined threshold instead of comparison with the value of the RSRP of neighboring SSs) plus the hysteresis parameter of that neighboring SS for a specified time (TTT) . The hysteresis parameter is used to remove sudden and small fluctuations of RSRP in the environment and acts as a damper for small changes.

Traditionally, HO parameters have been tuned by human experts. These experts use statistical models to analyze the wireless communication network (i.e., the radio access network, or RAN) . The current cellular network deployments are highly dependent on human experience where an expert designs rules or analytical models based on domain knowledge and assumptions of the network dynamics. This approach is far from optimal. First, the human-designed rules/models only consider a limited number of network states (e.g., user distribution, channel quality, etc. ) and parameters (e.g., HO parameters) , and cannot capture the complex relation between network states, parameter configurations and network performance. Second, the assumptions of the network dynamics, based on which the rules/models are developed, are often simplified ad only use a limited number of measurements from the wireless communication network without considering the non-stationary changes in real environments, which degrades their performance.

The dynamic nature of a wireless communication network requires methods for computing the threshold values of HO parameters that can respond faster to the changes in the environment. Thus, researchers have started using iterative try and test methods to resolve this problem. However, with the introduction of fuzzy logic, tools such as fuzzy systems and dynamic programming have showed promising results in HO parameter optimization.

Finally, these rules/models may not be able to deal with the cell diversities in the network which makes them sub-optimal. Recently, data-driven approaches based on machine learning (ML) have been extensively used for parameter configuration and network management in cellular networks. It has been shown that multi-layer perceptron (MLP) can be considered as a universal function approximator. Thus, in environments such as cellular networks where there is lack of an accurate analytical model and the network is highly dynamic, neural-network-based methods can be used to achieve performance metric prediction. ML models can utilize high dimensional information and approximate complex functions to fully describe the relation between network states, parameter configurations and the network performance, which cannot be achieved by human experts.

One of the key challenges in the network parameter optimization problem is the complex spatial and temporal dependencies in the cellular network. The parameter configuration of one cell not only affects its own performance, but also affects its neighbors’. Therefore, there are strong interactions between neighboring cells. When the network is heterogeneous, e.g., when cells have different frequencies, bandwidths and characteristics, these interactions become extremely complex. Furthermore, the network states are often temporally correlated, which should also be considered during optimization.

Improvements in HO parameter optimization are therefore desirable to optimize the overall performance of interconnected and interacting cells.

Summary

The present disclosure describes systems and methods which provide one or more efficient techniques to perform

In accordance with a first aspect of the present disclosure, there is provided a computer-implemented method for determining a threshold value for a handover (HO) parameter for a wireless communication network using a prediction model, the prediction model generated by: selecting a subset of a plurality of neighbor cells of a center cell in a cluster of cells of a wireless communication network; forming an augmented set of neighbor cells by using a feature transformation on the cluster of cells to map the cluster of cells to a latent space, and augmenting the subset of the plurality of neighbor cells with a set of additional neighbor cells closest to the center cell in the latent space; dividing the augmented set of neighbor cells into groups; applying a permutation invariant function to each of the groups to generate an output for each of the groups; aggregating the outputs; and determining the threshold value for the HO parameter at least partially based on the aggregated outputs.

In some or all examples of the first aspect, the HO parameter can be a threshold power below which an HO is triggered.

In some or all examples of the first aspect, the subset of the plurality of neighbor cells can represent the neighbor cells with which the center cell has handovers exceeding a threshold.

In some or all examples of the first aspect, the subset of the plurality of neighbor cells can represent a fixed number of the neighbor cells with which the center cell has the highest number of handovers.

In some or all examples of the first aspect, the set of additional cells can be equal in number to the number of cells in the subset of the plurality of neighbor cells.

In some or all examples of the first aspect, the permutation invariant function can include a set of learnable weights for each neighbor cell in the group.

In some or all examples of the first aspect, the method can further comprise: grouping samples collected in a time interval t into K groups by temporal order; determining an average network state for the samples in each of the k groups; processing the average network state for the samples in the k groups arranged in temporal order using a recurrent neural network to generate a temporal feature model; and determining the threshold value for the HO parameter at least partially based on the temporal feature model.

In some or all examples of the first aspect, the prediction model can be trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, the throughput ratio being used as training labels for training the prediction model.

In some or all examples of the first aspect, the prediction model can be trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, a throughput of the center cell and the groups of neighbor cells being used as training labels for training the prediction model.

In some or all examples of the first aspect, the predicted threshold value can be formulated as the output of a non-linear transformation function of the state vector and a change in the threshold for the previous day.

In some or all examples of the first aspect, the state vector can be a throughput ratio state vector, and the prediction model can be trained by concatenating temporal features of the center cell and a throughput of the center cell to form a throughput state vector.

In some or all examples of the first aspect, during training of the prediction model,

and

can be minimized, wherein LOSS ₁ is a throughput ratio at time t for cell v,

is a throughput ratio of a throughput for cell v at time t+1 relative to an average throughput for the augmented set of cells,

is a throughput ratio of a throughput for cell v at time t relative to an average throughput for the augmented set of cells,

is the throughput for cell v at time t+1,

is the throughput for cell v at time t, λ ₁ and λ ₂ are hyperparameters chosen for normalization, and θ ₁ and θ ₂ represent all trainable parameters in the prediction model.

In accordance with a second aspect of the present disclosure, there is provided a computing system for configuring a threshold value for a handover (HO) parameter of a wireless communication network using a prediction model, the computing system comprising: a processor configured to generate the prediction model by: selecting a subset of a plurality of neighbor cells of a center cell in a cluster of cells of a wireless communication network; forming an augmented set of neighbor cells by using a feature transformation on the cluster of cells to map the cluster of cells to a latent space, and augmenting the subset of the plurality of neighboring cells with a set of additional neighbor cells closest to the center cell in the latent space; dividing the augmented set of neighbor cells into groups; applying a permutation invariant function to each of the groups to generate an output for each of the groups; aggregating the outputs; and determining the threshold value for the HO parameter at least partially based on the aggregated outputs.

In some or all examples of the second aspect, the HO parameter can be a threshold power below which an HO is triggered.

In some or all examples of the second aspect, the subset of the plurality of neighbor cells can represent the neighbor cells with which the center cell has handovers exceeding a threshold.

In some or all examples of the second aspect, the subset of the plurality of neighbor cells can represent a fixed number of the neighbor cells with which the center cell has the highest number of handovers.

In some or all examples of the second aspect, the set of additional cells can be equal in number to the number of cells in the subset of the plurality of neighbor cells.

In some or all examples of the second aspect, the permutation invariant function can include a set of learnable weights for each neighbor cell in the group.

In some or all examples of the second aspect, the processor can be configured to generate the prediction model by: grouping samples collected in a time interval t into K groups by temporal order; determining an average network state for the samples in each of the k groups; processing the average network state for the samples in the k groups arranged in temporal order using a recurrent neural network to generate a temporal feature model; and determining the threshold value for the HO parameter at least partially based on the temporal feature model.

In some or all examples of the second aspect, the prediction model can be trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, the throughput ratio being used as training labels for training the prediction model.

In some or all examples of the second aspect, the prediction model can be trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, a throughput of the center cell and the groups of neighbor cells being used as training labels for training the prediction model.

In some or all examples of the second aspect, the predicted threshold value can be formulated as the output of a non-linear transformation function of the state vector and a change in the threshold for the previous day.

In some or all examples of the second aspect, the state vector can be a throughput ratio state vector, and the prediction model can be trained by concatenating temporal features of the center cell and a throughput of the center cell to form a throughput state vector.

In some or all examples of the second aspect, during training of the prediction model,

and

can be minimized, wherein LOSS ₁ is a throughput ratio at time t for cell v,

is the throughput for cell v at time t+1,

In accordance with a third aspect of the present disclosure, there is provided a non-transitory machine-readable medium having tangibly stored thereon executable instructions for execution by one or more processors, wherein the executable instructions, in response to execution by the one or more processors, cause the one or more processors to configure a threshold value for a handover (HO) parameter of a wireless communication network using a prediction model, the prediction model generated by: selecting a subset of a plurality of neighbor cells of a center cell in a cluster of cells of a wireless communication network; forming an augmented set of neighbor cells by using a feature transformation on the cluster of cells to map the cluster of cells to a latent space, and augmenting the subset of the plurality of neighboring cells with a set of additional neighbor cells closest to the center cell in the latent space; dividing the augmented set of neighbor cells into groups; applying a permutation invariant function to each of the groups to generate an output for each of the groups; aggregating the outputs; and determining the threshold value for the HO parameter at least partially based on the aggregated outputs.

In some or all examples of the third aspect, the HO parameter can be a threshold power below which an HO is triggered.

In some or all examples of the third aspect, the subset of the plurality of neighbor cells can represent the neighbor cells with which the center cell has handovers exceeding a threshold.

In some or all examples of the third aspect, the subset of the plurality of neighbor cells can represent a fixed number of the neighbor cells with which the center cell has the highest number of handovers.

In some or all examples of the third aspect, the set of additional cells can be equal in number to the number of cells in the subset of the plurality of neighbor cells.

In some or all examples of the third aspect, the permutation invariant function can include a set of learnable weights for each neighbor cell in the group.

In some or all examples of the third aspect, the executable instructions, when executed by the processor, can cause the processor to generate the prediction model by: grouping samples collected in a time interval t into K groups by temporal order; determining an average network state for the samples in each of the k groups; processing the average network state for the samples in the k groups arranged in temporal order using a recurrent neural network to generate a temporal feature model; and determining the threshold value for the HO parameter at least partially based on the temporal feature model.

In some or all examples of the third aspect, the prediction model can be trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, the throughput ratio being used as training labels for training the prediction model.

In some or all examples of the third aspect, the prediction model can be trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, a throughput of the center cell and the groups of neighbor cells being used as training labels for training the prediction model.

In some or all examples of the third aspect, the predicted threshold value can be formulated as the output of a non-linear transformation function of the state vector and a change in the threshold for the previous day.

In some or all examples of the third aspect, the state vector can be a throughput ratio state vector, and the prediction model can be trained by concatenating temporal features of the center cell and a throughput of the center cell to form a throughput state vector.

In some or all examples of the third aspect, the executable instructions, when executed by the processor, can cause the processor, during training of the prediction model, to minimize

and

wherein LOSS ₁ is a throughput ratio at time t for cell v,

is the throughput for cell v at time t+1,

Other aspects and features of the present disclosure will become apparent to those of ordinary skill in the art upon review of the following description of specific implementations of the application in conjunction with the accompanying figures.

Brief Description of the Drawings

FIG. 1A is a schematic diagram illustrating graph modelling of a cell cluster of a wireless communication network in accordance with example embodiments described herein.

FIG. 1B is a schematic diagram of a graph generated for the cell cluster of FIG. 1A.

FIG. 1C shows a message passing framework applied to the graph of FIG. 1B.

FIG. 2 is a flowchart of a general method of modeling and configuring a threshold value for a handover parameter of a wireless communication system in accordance with example embodiments described herein.

FIG. 3 is a schematic diagram showing network augmentation performed during a method for configuration of handover parameters in wireless communication systems in accordance with example embodiments described herein.

FIG. 4 is a schematic diagram showing the processing of neighbor groups automatically formed for an exemplary cell topology during the method in accordance with some example embodiments described herein.

FIGS. 5A and 5B are schematic diagrams of a spectrum of neighbor aggregating that can be performed for a cell topology.

FIG. 6 is a schematic diagram illustrating a prediction model employing

as weights in accordance with example embodiments described herein.

FIG. 7A is a flowchart of the process of configuring a threshold value for a handover parameter using trained sub-models for throughput prediction and throughput ratio prediction.

FIG. 7B is a flowchart of the internal workflow of a simulator in accordance with example embodiments described herein.

FIG. 8 is a set of graphs illustrating the training and test accuracy of different models trained with the same dataset in terms of MSE.

FIG. 9 is a schematic diagram of the process of action recommendation by the trained model and the simulator.

FIG. 10 is a graph of a performance comparison of different models initialized with random actions.

FIG. 11 is a graph showing the impact of different optimization objectives for the parameter configuration in the proposed TAG-GCN model.

FIG. 12 is a graph showing a performance comparison of the described TAG-GCN model under different types of action initialization for day t ₂.

FIG. 13 is a graph showing the impact of the described TAG-GCN model on the load balancing measured as the average throughput ratio across the network.

FIG. 14 is a schematic diagram showing various physical and logical components of a computing system for configuration of a threshold value for a handover parameter in a wireless communication network in accordance with example embodiments described herein.

FIG. 15 shows the overall process of modeling a cell cluster and generating a prediction for it in accordance with an embodiment.

Detailed Description of Example Embodiments

The present disclosure is made with reference to the accompanying drawings, in which embodiments are shown. However, many different embodiments may be used, and thus the description should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete. Wherever possible, the same reference numbers are used in the drawings and the following description to refer to the same elements, and prime notation is used to indicate similar elements, operations or steps in alternative embodiments. Separate boxes or illustrated separation of functional elements of illustrated systems and devices does not necessarily require physical separation of such functions, as communication between such elements may occur by way of messaging, function calls, shared memory space, and so on, without any such physical separation. As such, functions need not be implemented in physically or logically separated platforms, although such functions are illustrated separately for ease of explanation herein. Different devices may have different designs, such that although some devices implement some functions in fixed function hardware, other devices may implement such functions in a programmable processor with code obtained from a machine-readable medium. Lastly, elements referred to in the singular may be plural and vice versa, except wherein indicated otherwise either explicitly or inherently by context.

A model is proposed herein in accordance with some embodiments to precisely imitate the cellular network environment. This model is then used to configure network parameters. For each center cell surrounded by some neighbors, a novel method is defined to consider the adjacent cells and differentiate their impact on the center cell. This model helps to capture the spatial interactions and the network heterogeneity. Different time instants of the network are considered to capture the temporal dependencies. Finally, a multi-objective exploration strategy is introduced to balance the center cell’s own performance and that of its neighbors to configure the parameters.

A novel method in an embodiment is proposed to use the impact from the neighbors of each cell in a distinguishable way to capture the spatial dependencies of the network. Different time-instants in the model are considered to better reflect the temporal dependencies of the network. A multi-objective optimization strategy on the imitated model is introduced to consider several performance metrics to configure the network parameters.

UEs travelling at higher speeds, such as autonomous vehicles, need effective and efficient handover procedures to guarantee minimum service delay or interruption. Also the increase in the number of SSs (due to increase of the number of UEs) causes frequent service area crossings in UEs which, again, requires proper handling techniques.

The adjustment of handover parameters can significantly affect the network throughput in several ways. First, it can help balance the traffic load in the network. Take the A2 threshold parameter in Table 1 (that is, the threshold signal strength for the serving cell, measured in dBm, below which a handover is triggered) as an example. During the HO process in cellular networks, in order to guarantee an acceptable service quality, a UE must keep monitoring the RSRP of the serving cell. As soon as the RSRP value becomes less than this pre-defined A2 threshold, the UE then starts to report measurements to its serving cell and prepares for handover. When this threshold is set higher, more UEs in the serving cell will be triggered for handover, and therefore it spreads the serving cell’s load to its neighbors, resulting in the change of throughput of the serving cell and its neighbors.

Apart from the effect of load balancing, the handover parameters also impact a cell’s throughput via other mechanisms. For example, small values of the A2 threshold cause a very poor condition for edge UEs and lead to repeated connection loss. On the other hand, the handover process is called too frequently with a large value of this threshold, which requires massive amount of bandwidth for measurement reporting and causes a drop in UE’s data throughput.

Hence, when optimizing this parameter, cell interactions must be considered and an accurate optimization process is of high importance for the network overall performance. This requires an accurate model to perfectly mimic the environment and an optimization strategy to achieve the highest possible performance.

Due to its significant importance, handover parameter optimization has been an interesting topic for researchers. To assess the effectiveness of their proposed algorithms on optimizing handover parameters, researchers use different measures such as handover failure rate (HOF) , handover frequency, ping-pong rate (PP) , network throughput, and load balancing.

While optimizing a problem at hand in a real network, some of these measures may conflict. Therefore, it is crucial to consider a combination of these quantities to find a solution which satisfies different requirements of the network.

There are different parameters that correspond to handover in a network. These parameters generally are grouped in three categories. RSRP and RSRQ parameters which are indicators of the signal strength and quality of the serving station, respectively, Hysteresis parameters which act as a tolerance margin to avoid PP effect, and time to trigger (TTT) parameters that are set such that short period violation of handover conditions are ignored to avoid a PP effect.

Traditionally, different methods have been used for managing handovers in a wireless network. Fuzzy system handover algorithms, for example, have been used in different designed techniques the design of an effective set of fuzzy rules based on different measures of the QoS in the network. Such techniques are accurate and stable due to deterministic rules, however designing proper rules for such complex problem is not an easy task. Moreover, the set of rules for optimal functioning of the network will become increasingly complex and therefore not manageable.

In the past decade, more advanced ML techniques (e.g., reinforcement learning) have been used in mobility management and handover optimization and proved to be very effective. These models are a significantly efficient tool for the task of handover optimization due to the fact that they can model non-linear functions and adapt to highly dynamic environments. Many studies suggest to use RL methods with fuzzy systems for this purpose. Wu et al., “Dynamic fuzzy q-learning for handover parameters optimization in 5g multi-tier networks, ” in 2015 International Conference on Wireless Communications and Signal Processing (WCSP) , 2015, disclose a method that tries to properly adjust the handover margin for improving call drop ratio (CDR) . Klein et al., “Fuzzy q-learning for mobility robustness optimization in wireless networks, ” in 2013 IEEE Globecom Workshops (GC Wkshps) , 2013, use the HOF rate as its measure for improvement while the input parameters are TTT and hysteresis. Although measures of HOF rate and CDR are important in terms of user experience, the actual throughput that is delivered to the users is also a very important measure. This issue is tackled partially, for example, by

et al., “Fuzzy rule-based reinforcement learning for load balancing techniques in enterprise LTE femtocells, ” IEEE Transactions on Vehicular Technology, vol. 62, no. 5, 2013 where the authors adjust the transmission power (TXP) and the hysteresis in small BS to improve the balance of the load in the network. Load balancing is also considered by Dinh et al., “Joint implementation of several lte-son functions, ” in 2013 IEEE Globecom Workshops (GC Wkshps) , 2013, as an optimization objective where the hysteresis and TTT parameters are simultaneously adjusted, while a heuristic algorithm optimizes the handover offset according to load measurements in the cell.

et al., “Load balancing and handover joint optimization in LTE networks using fuzzy logic and reinforcement learning, ” Computer Networks, vol. 76, pp. 112–125, 2015, introduce an algorithm that considers load balancing and handover optimization jointly as measures.

Deep reinforcement learning (DRL) is another technique that has been used to solve the handover optimization problem. Cao et al., “Aif: An artificial intelligence framework for smart wireless network management, ” IEEE Communications Letters, vol. 22, no. 2, pp. 400–403, 2018, propose a framework based on DRL where actions are flexible and can be chosen by the user, and the objective of the optimization is the throughput and the handover count. Hence, the parameter value is not optimized directly and instead, they learn decision policy for UEs, and let UE to decide which cell to connect directly. Wang et al., “Handover control in wireless systems via asynchronous multiuser deep reinforcement learning, ” IEEE Internet of Things Journal, vol. 5, no. 6, pp. 4296–4307, 2018, propose a framework in which the UEs first are clustered based on their usage pattern and then the handover process is optimized in each cluster by a DRL method. The input to the algorithm is the RSRQ and the objectives are throughput and handover rate. Another newly introduced approach to the problem of handover optimization is the Contextual Bandit model which is used in Chuai et al., “A collaborative learning based approach for parameter configuration of cellular networks, ” in IEEE INFOCOM 2019 -IEEE Conference on Computer Communications, 2019, pp. 1396–1404, to optimize network parameters. The authors have used the throughput of the cells in their tests. Transfer learning is also used in the framework to transfer the knowledge from other cells to the subject cell.

Moreover, the use of Graph Neural Networks (GNN) and Graph Convolutional Networks (GCN) has also yielded significantly well-designed models to predict the network traffic and optimize the corresponding parameters. For example, Zhang et al., “An AI-based optimization of handover strategy in non-terrestrial networks, ” in 2020 ITU Kaleidoscope: Industry-Driven Digital Transformation (ITU K) , 2020, introduce a novel handover strategy based on GCNs. The handover process is modeled as a directed graph by which the user tries to predict its future signal strength. In the final step, a GCN is used to extract the underlying regularity of the best handover strategies of different users. Other works such as Zhao et al., “Cellular network traffic prediction incorporating handover: A graph convolutional approach, ” in 2020 17th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON) , 2020, introduce novel methods of network traffic prediction which combined with a greedy search or action configuration method can be used for handover parameter optimization.

Threshold values for HO parameters are used to trigger handovers in wireless communication networks. Consider a network with N cells, and form N clusters each composed of one of the network cells as its center cell along with its neighboring cells. Note that the neighboring cells are defined as the cells with which the center cell has handover events. The target HO parameter is denoted by A. In one example, in case of the HO parameters noted in Table 1, the target HO parameter has a threshold value with which the RSRP of the reference signal received from SS is compared with and the UE decides whether to trigger an HO based on this comparison; that is, the A2 threshold. According to the 3GPP standard, “3gpp ts36.311” , Evolved Universal Terrestrial Radio Access (E-Ultra) ; Radio Resource Control (RRC) ; Protocol specification, 2016, an A2 event is triggered when the received power at user u from cell n, P ^u, n, satisfies

P ^u, n+H _ys＜Thresh, (1)

where H _ys is the hysteresis parameter to avoid frequent handovers and Thresh is the A2 threshold being optimized.

FIG. 1A shows an exemplary cluster in a wireless communication network 20 including a set of SSes 24, each defining a cell 28. A center SS 28’ is positioned in a center cell 28’. A plurality of UEs 32 are shown distributed throughout the wireless communication network 20. Some of the UEs 32 are shown proximal to the borders between cells 28, such as UE 32’. The UE 32’ monitors the received power from cell n to check the handover criteria.

Methods and systems for configuring a threshold value for HO parameters which trigger a HO event are provided. The methods and systems described herein can include a modeling phase and an HO parameter value configuration phase. The modeling phase includes applying a new aggregation method to collect useful information coming from each adjacent cell to improve the learning capability of a model which predicts load balancing and the throughput of a wireless communication network. The modeling phase also includes using the temporal information (different time intervals) of a center cell to better learn the trend in the features of the center cells in one day. The HO parameter value configuration phase (otherwise referred to as an action configuration phase) includes recommending optimal values for specific HO parameters by sequentially optimizing multiple objectives for joint optimization for the HO parameter configuration using the above learned model. In other words, HO parameter value configuration phase includes finding the best possible solutions for one objective and optimizing the next objective over those solutions.

A model which predicts load balancing and the throughput of a wireless communication network is trained. Based on the predictions of the load balancing and throughput output by the model, optimal values for specific HO parameters of the wireless communication network are recommended.

Methods and systems in accordance with some exemplary embodiments use a dataset for training the model using measurements of a state of a cell of a wireless communication network made in a real wireless communication network. The measurements can include, for example, the antenna transmission power, throughput, user number, physical resource block (PRB) usage ratio, the amount of data traffic, and the transmission bandwidth. These measurements are made periodically during the time in which the HO parameters are fixed (e.g. every hour for one day) , and accumulated in a three-dimensional (3-D) matrix, where each element is specified by a tuple (cell, date-time, feature) .

The training dataset is used to train the model for predicting the load balancing and the throughput, and then, an optimization technique generates a value for a targeted HO parameter such that both performance metrics (namely, throughput and throughput ratio) are satisfied partially. These generated values for the HO parameters of each cell yield a higher performance in the network.

An online optimization process is considered. In real practice, network operators are often conservative and only allow a limited number of experiments. Assume the optimization period consists of T days, and the value of A2 threshold can be adjusted once for each cell at the beginning of each day. For day t, let D _t be the total bits transmitted by all the cells, and T _t be the total transmission time. It is desirable to maximize the accumulated network throughput of the optimization period; i.e.,

Maximizing the overall network throughput by jointly optimizing the A2 threshold of all cells can be challenging. The problem becomes even more complicated as the network size increases, which makes a centralized solution not scalable. The adjustment of the A2 threshold of one cell only affects its local neighborhood.

Methods and systems in accordance with some example embodiments described herein model any wireless communication network as a graph. A graph is defined as an ordered pair of sets G= (V, E) in which each edge e∈E connects two nodes in the set V (this connection can be directional on non-directional) . Graphs are used extensively in any network to model the interaction/connection/relation between the members of a set. In wireless communication networks, a graph model of the wireless communication networks consists of the cells as nodes. The connection between the nodes (cells) represents an edge between any two nodes (cells) that have a common border. The edges are representative of potential HO events. That is, two nodes in the representing graph of a wireless communication network are connected with an edge if an HO event can happen or happened between them.

FIG. 1B shows a sub-graph 36 of the exemplary cluster of the wireless communication network 20 of FIG. 1A, wherein the cells 28 are represented as nodes, and the borders between cells 28 are represented as edges 40 between the nodes. In the graph 36, the edges 40 between two cells 28 are represented if and only if there is an HO between the two cells 28.

The models and methods in accordance with some embodiments described herein convert the centralized problem into a local decision problem. That is, each cell only looks at its local performance metrics and chooses its own parameter configuration value.

The adjustment of the A2 threshold affects the network throughput via two means: better resource utilization by load balancing, and improved cell throughput with less connection loss and measurement reporting. Consequently, in order to configure the A2 threshold parameter, these two metrics may be considered in the local decision problem.

The throughput of cell i at day t is highly dependent on its A2 threshold a and denoted as

The load balancing factor in the i-th cluster with center cell i at day t with the A2 threshold a is identified as the ratio of the center cell throughput to the average throughput of its neighboring cells, denoted by

and formulated as follows:

where

is the average throughput of the neighbors of cell i and, assuming the set of all neighbors of cell i at day t denoted by N _t (i) , it can be formulated as:

The throughput ratio (rather than traffic/user ratio) is used since different cells have different capacities. This value approaches 1 when loads of different cells match their capacities.

The proposed method aims at maximizing the throughput of each individual cell while keeping the load balancing factor as close as possible to 1 by choosing the threshold A from a limited number of possible values.

The goal is to optimize the two important network performance metrics, namely, throughput ratio

and throughput

for each cell i∈ [1, N _t] , where N _t is the total number of cells at day t, at the same time. In order to meet these requirements, the optimization problem for tuning the A2 threshold for cell i is proposed as follows:

where A is the set of all possible values for the A2 threshold in the cellular network.

The challenges of solving the above problem are many. First, since the network performance functions are hugely complex, dynamic and unknown, getting accurate values for

and

is hard. Instead, a data-driven approach is adopted where reward models to estimate the performance metrics are learned. Second, in real-world practice, only a limited amount of experiment budget is allowed by network operators. Therefore, enough diverse historical data (state, action pairs) to train our data-driven learning model may not be available, and hence a data augmentation technique in the form of neighbor cell augmentation is used in the methods and models described herein to enrich the features from each cell. Third, the handover parameter aims to balance the load between adjacent cells. Thus, it is important to model the information coming from the adjacent cells to achieve accurate reward modeling.

Discussed herein is an exemplary data-driven solution to learn an accurate reward model to imitate how the environment will respond to given networks’ states and handover parameter (A2 threshold) configurations, which is intertwined with the term action. The goal is to have a reward model that can accurately predict the cell performance metrics such as throughput and throughput ratio, which will be a determining factor during the action configuration stage.

There are three key components of the method and model in accordance with some embodiments disclosed herein. First, to better capture the dependency between each cell and its neighboring cells, a novel method for neighboring cell feature aggregation is introduced. Second, a temporal feature aggregation step with recurrent neural networks (RNN) to model the temporal correlation from the historical sequence of the network states is proposed. Lastly, the overall training process is elaborated, considering the impact from the neighboring cells, the temporal correlation in the network states and the action to be optimized.

Since the HO parameter tuning parameter and the learning of the network model is heavily impacted by the features of the center cell as well as its neighboring cells, it is desired to capture the neighboring cell information during the modeling process. Recently, message-passing neural network (MPNNs) frameworks in the form of graph neural networks (GNN) have been introduced and demonstrated to be effective in modeling real world applications with structured information where the dependencies in the data set is modeled in the form of a graph. In each layer of a graph neural network, each node’s representation includes its features, as well as its neighboring nodes’ features. The graph neural network framework is suitable to handle the dependency between the center cell and the neighboring cells in cellular networks. We will first elaborate on how we transfer the cellular network modeling problem into a learning problem on graph. Then we will delve into the challenge of using the message passing framework in cellular networks.

Now with reference to FIG. 2, a method 100 of configuring parameters in a wireless communication network that is provided in accordance with particular embodiments will be described. The method 100 commences with the selection of a set of neighbors with which handovers are likely (110) .

In FIG. 1B, the sub-graph 36 G _t= (V _t, ε _t) for day t corresponds to the cluster of the wireless communication network 20 of FIG. 1A. Each nodev∈V _t represents one cell and is associated with a feature vector

including the statistical properties of node v measured in day t. The statistical properties could include the antenna transmission power, physical resource block (PRB) , usage ratio, the amount of data traffic, transmission bandwidth, etc. These features serve as the node attributes. The edge set ε _t encodes the interactions between cells based on the handover events between pairs of cells. Based on historical data, in accordance with some embodiments, if any pair of cells has an average number of handover events above a threshold τ, an edge is provided between those two cells. The neighboring set for node v is denoted as

In other embodiments, other manners of selecting the neighboring set with which handovers are likely can be employed. For example, a number of the neighboring cells with each which the central cell has the highest number of handovers can be selected as the neighboring set.

The successful application of graph neural networks mostly relies on a homogeneous graph structure in nature. For example, in social networks or recommendation systems, each edge represents a positive correlation between the adjacent nodes. The message passing framework applies the same transformation function to every neighbor and aggregate across them to obtain the neighbor representation, as shown in FIG. 1C. It is generally understood how to process a heterogeneous graph with multiple relation types. Prior knowledge of the given relation type between each pair of nodes is used, and the meta-path design is used to distinguish the information coming from different types of neighbors.

When using a graph to represent the interactions between cells in the cellular networks, the interaction types are not known. Due to the heterogeneous nature of the cellular network, the relationships between the neighboring cells can be complex. Concretely, there might be an implicit M latent relationship types R= {r ₁, r ₂, …, r _M} that can be learned to better handle the complex interactions in the cellular networks. Intuitively, assuming each cell is represented by its states such as PRB usage, channel quality indicator (CQI) , the amount of traffic, and available bandwidth, after dividing the neighbors into different groups, it is desired that the neighbors in each group have similar network states with respect to the center cell to better capture the rich information from neighboring cells in a distinguishable way. A novel graph convolutional network (GCN) approach called auto-grouping GCN (AG-GCN) is proposed to characterize this special property of cellular networks when handling the interactions between the neighboring cells.

Referring again to FIG. 2, an additional set of neighbors in latent representation space is used to augment the original set of neighbors selected at 110 (120) .

In cellular network modeling, only a limited amount of experiment budget is allowed by network operators. There may not be sufficient diverse historical data (state-action pairs) to train a data-driven model. In constructing the graph based on the handover events, there are cells that may have very limited number of neighboring cells. Thus, a data augmentation technique in the form of neighbor cell augmentation is employed to enrich the feature of each cell. The neighbor cell augmentation is based on the similarity between cells in the latent space.

A feature transformation function f (·) : R ^d →R ^l which maps the input node feature

to a latent space

where

is defined. In order to capture the long-range dependencies and similarity in the cellular network, an additional neighborhood in the latent representation space is added based on its similarity in Euclidean distance. For each node v∈V _t, the augmented neighborhood

is formed, where

and

are the neighbors of node v in the original graph and in the latent space, respectively. This augmented set may be referred to alternatively as the “set of neighbors” . The neighbors in the latent space are selected based on their Euclidean distance to the center cell. The n nearest nodes in the latent space are selected to create neighborhood

for cell v, where the number of nodes selected based on the feature similarity will be equal to the neighborhood size in the original graph

effectively doubling the number of neighbors. In other embodiments, the number of nearest nodes n can differ from the number of nodes in the original graph.

FIG. 3 shows the original graph 36 of neighbors being augmented by the latent space neighbor graph 44 to form an augmented neighborhood 48 in accordance with methods and systems in example embodiments. The sub-graph 36 of FIG. 1B is shown together with a latent space neighbor sub-graph 44 of all the nodes of the wireless communication network (or a portion thereof) . Then, the nearest n neighbors, N _s (v) , from the latent space neighbor sub-graph 44 are used to augment the set N _g (v) to arrive at a neighborhood augmentation sub-graph 48 of set

Referring again to FIG. 2, the augmented set of neighbors is divided into groups (130) .

Advantageously, aggregating the information of the neighbor cells ( “neighbors’ information” ) in the center cell provides a better prediction of load balance of the wireless communication network and the throughput measures. The neighbors’ information is grouped using a grouping function, which allows utilization of the neighbors’ information without losing of too much of it, while keeping the model complexity low. Methods and systems described herein in accordance with some embodiments use two sets of neighbors for increasing the effective range of the aggregation adding extra neighbors from an embedding space to the graph space.

Once the augmented neighborhood set is obtained, the neighbors in the augmented neighborhood N _t (v) are divided into different groups by the geometric operator γ. Considering node v and its neighbor node u ∈ N _t (v) , the relation between them at day t is denoted as

Intuitively, it is desired to assign an interaction type (edge membership) between each pair of center and neighbor cell based on their latent representations. For each r _i∈R, the neighborhood feature set for group r _i at time t is defined as:

Then, each group is processed using an MLP and the results are aggregated using a GCN (140) .

Since the order within each neighbor group should not impact the output of the representation, a permutation invariant (mean pooling) function π (·) is applied to the neighbors within each group, a multi-layer perceptron (MLP) module is trained, and they are aggregated separately. FIG. 4 shows an example of the aggregation process from the augmented sub-graph 48 of FIG. 3, where l = 2 and |R| = 4 and the representation after the permutation invariant function π (·) is shown by black dashed arrows ended to nodes 1, 2, 3, and 4. Then, for each group r _i∈R, a non-linear transformation/activation function is further applied as:

where

is a learnable weight matrix for the neighbors in group r _i of node v at day t, and σ (·) is a non-linear function; e.g., tanh. Then for each node, it is desired to aggregate the transformed neighborhood features from their different groups of neighbors in a distinguishable way where

for r∈R are further aggregated by a concatenation operation:

where [; ] represents concatenation.

Note that, in forming the groups, if no members are assigned to group N _t (v) , all zeros or an average of all the neighbor values can be passed into the corresponding MLP model.

Auto-group neighbor aggregation applies a trade-off between the amount of information that is lost in case of aggregating all the neighbors as one whole group, and the complexity of the MLP model in case of using the information from every single neighbor.

FIGS. 5A and 5B illustrate a spectrum of the number of groups R. If R is too small, as is shown in FIG. 5A, there is too much information loss and less space/time complexity. At the other end of the spectrum, if R is too large, as is shown in FIG. 5B, there is no information loss and too much space/time complexity, thus resulting in significant resources to process each group N _t (v) .

A similar argument about the information-complexity trade-off is used to group the data of the center cell based on sampling time. The data from each cell is divided into K (set by user) subsets. The main motivation is that after aggregating the neighbors with AG-GCN, if all the temporal data from a specified period in which the parameters are fixed are aggregated using a permutation invariant function, some information is lost, for example regarding the similarity in the usage patterns. To better understand this, note that, for example, the traffic of a cell is completely different in different times of the day (commute hour, noon, afternoon, etc. ) and this method allows capturing the periodicities lying within daily data. Hence, it is reasonable to group the samples taken in a specific time period for each cell based on such similarities to avoid information loss.

Referring again to FIG. 2, temporal features are modelled (150) .

Well capturing the trend in the states of each cell within a day can benefit the prediction of the performance metric for the following day. Additional temporal features for each center cell are used to extract the changing dynamic pattern of its states within each day to further improve the reward model performance. It is assumed that the samples of the center cell v at day t can be divide into K groups (e.g., busy hour, non-busy hour, etc. ) by the temporal order. For all the samples in each group k, the average network state for each group of samples is denoted as

An RNN layer is used to capture this temporal dependency of the features from different groups by feeding all the network states as an input sequence

The output of the temporal feature modeling phase can be written as:

where δ is the set of trainable parameters of the bidirectional RNN layer.

Then the full model is compiled (160) .

The main purpose of the model is to estimate the real network’s response, and predict the throughput ratio and throughput of the center cell for the next day based on the observed network states in the current day. These performance metrics are not only affected by the current day’s states, but also highly correlated with the action chosen to configure for the next day. Thus, the model is also fed by the actions of the next day. Furthermore, the throughput ratio and throughput of the next day are highly dependent on the previous values of throughput and throughput ratio of the cells and clusters. Hence, the throughput ratio of the current day, i.e.,

is considered in the prediction process.

To make the final prediction, the learned representation of the neighborhood by the AG-GCN aggregation, the temporal features of the center cell, and the throughput ratio of the current day, i.e.

are concatenated to form the state vector of cell v, and defined as:

FIG. 6 shows the prediction model employing

as weights.

Since the final representation should be sensitive to the input action chosen, the throughput ratio and throughput of the next day for cell v are formulated as the output of a non-linear transformation function of state and action:

The parameter selected for optimization is the A2 threshold as previously discussed. In order to properly use the A2 threshold for the prediction, the change in this parameter is compared to the previous day as the action

where

and

are the A2 thresholds for cell v at day t+1 and t, respectively. The reason for this design choice has twofold. First, the original action space of A2 is large, but the range of the change of action can be smaller by controlling the adjustment steps, making it easier for the model to learn and conduct the decision making step. Besides, the delta action directly reflects the change in the cell coverage/loads, so they are more sensitive to the performance metrics.

Finally, the same model is built for the throughput

where the states are

To form the training objective, data of T + 1 consecutive days is considered and the pairs (t, t+1) , t∈ {1, 2, …, T} are formed to predict the throughput ratio and throughput of the center cell in day t+1. The model is finally trained by minimizing the followi ng loss functions

where λ ₁ and λ ₂ are the hyperparameters chosen for the regularization. θ ₁ and θ ₂ represent all the trainable parameters in the models. T is the total number of pairs of data used for the training. N _t is the total number of cells available at day t.

The trained model is now able to mimic the real network and predict both throughput ratio and throughput of each center cell the coming day and can be used to check the impact of actions towards the performance metrics being considered. The best action is the action that can optimize the throughput ratio and throughput jointly or in other words maximize the optimization problem in equation (5) .

Referring again to FIG. 2, once the model is compiled, the HO parameters can be configured (170) .

The algorithm for action configuration based on the trained model is shown below and illustrated in FIG. 7A, with one action per day recommended for each cell. The goal is to find the best action for cell i at day t, i.e.,

that can optimize the problem in equation (5) .

Algorithm 1: TAG-GCN for Action Configuration

Result: The final actions for all cells at day t

Initialization: v=1,

while v≤N _t do

Set the input vectors of the TAG-GCN model for cell v and its neighbors u as

and

Form the predicated functions

and

based on the trained TAG-GCN model;

for

do

end

The threshold values for the HO parameters of the wireless communication network are configured based on the two prediction models. A sequential method is used in which a candidate set of solutions is computed using one of the two prediction models, and then, the output of the other model is optimized within that candidate set of actions. This allows considering two objectives at the same time and making a compromise even when the objectives are conflicting.

As discussed in the earlier sections, the main objectives to consider in the action configuration process are load balancing, identified by the throughput ratio, and the cell throughput. In general, when dealing with a multi-objective problem, different objectives are often conflicting, and it may be challenging to optimize them simultaneously. Here, the problem in equation (5) is broken into two sub-problems that are solved sequentially. A subset of actions which optimize the throughput ratio with respect to its predicted value, i.e.,

for cell i at day t, is chosen first, and then, the action that optimizes

is selected from the above subset.

Specifically, the throughput ratio is optimized and the set of best c values for

denoted

is found, such as

Then, the goal is to achieve the maximum possible throughput for cell v at day t and this is through

is then the final recommended action for cell v at day t. This procedure for all the N _t cells of the network at day t is presented in Algorithm 1 above.

Experiments were conducted on a large-scale cellular network simulator constructed from real-world data.

In the following experiments, principal component analysis (PCA) was used as the mapping function f for the graph embedding in the AG-GCN neighbor aggregation to transform the graph space to a 2-dimensional space. The relationship operator γ for the auto-grouping is given by Table 1. The permutation-invariant function π applied on each group of neighbor is average in our experiment.

TABLE 1: The relationship operator γ

The experiments were conducted on a proprietary cellular network simulator. Here, the methodologies of the simulator construction are briefly introduced.

The simulator is constructed from a dataset that is collected from a real metropolitan cellular network containing around 1500 cells. The collection period starts from October 17 to October 31, 2019, where for each day the data of cells are sampled hourly. Each data sample contains information such as the cell ID, sample time, configuration of cell parameters, and measurements of the cell states (e.g., the number of total users within the cell, the number of active users, the cell average CQI, the cell traffic load, etc. ) and performance indicators (e.g., the average cell throughput, the edge user throughput, etc. ) . The neighbor relation information between cells are also collected and the hourly average handover counts between neighbor cells are recorded.

The dataset is collected under the default network configuration and the value of A2 threshold is not changed during the collection period. To evaluate the model and the action configuration strategy, the simulator needs to simulate the network performance given arbitrary A2 threshold values of the cells. More precisely, let

represent cell i’s states at hour t under the default A2 threshold configuration, and

be the configured A2 threshold value for cell i at hour t, the simulator will output

where

is cell i’s throughput, and

is the observed states of cell i under the new configurations at t.

Fig. 7B gives a simple sketch on the internal workflow of the simulator. During simulation, when

are configured at hour t, the cell states

are read from the collected data, and inputted to the simulator. The cell coverage C ⁱ for each i is first computed by a built-in function that considers cell i’s frequency, bandwidth, and

Then cell i’s traffic load (including the number of users, and the amount of data bits for transmission) at hour t is redistributed among itself and its neighbors based on the change of C ⁱ and C ^j with j∈N _t (i) . After each cell’s traffic load is updated, the throughput

is predicted based on cell i’s traffic load. The load-throughput prediction model for each cell is pre-trained from the collected historical data and stored when the simulator is initialized. The value

is further adjusted by a factor that considers the throughput loss due to measurement reporting or connection loss. Finally, the simulator outputs

As each cell’s traffic load (part of the cell states) is modified, the updated states

are also outputted.

In order to evaluate the prediction accuracy of our model, the simulator was used to generate the dataset with a random policy. On each day, the A2 threshold for each cell is randomly selected around the default action -100 dBm within the range [-105, -95] . This allows for the training of all the models with the same dataset to provide a fair comparison of their accuracy.

As samples generated by the simulator are hourly, the daily samples were aggregated for each cell in each day as described above. In order to evaluate the learning accuracy of the model in predicting cell throughput and throughput ratio, the model was trained with samples for seven days *t ₁, …, t ₇+. The pairs (t ₁, t ₂) , …, (t ₆, t ₇) were generated, and the model was trained with these sample pairs. At day t _i, where i < 7, the samples (t ₁, t ₂) , …, (t _i, t _i+1) were shuffled and 80%of them were used in training and the rest for testing. The mean square error (MSE) was reported to measure the reward model performance.

In order to show the effectiveness of the model described herein, it was compared with some other models. In Table 2, benchmarks and the properties of each model are shown. The first model is MLP, where only the features of the center cells were used, and the neighboring cells’ features were ignored. In the GCN model, the neighbors information is aggregated in only one group without an AG strategy. The AG-GCN model uses the auto-grouping scheme described in this document, however, it ignores the temporal dependencies of the data. Finally, TAG-GCN is the most complete model that uses both auto-grouping of the neighbors and the temporal patterns in the data. In FIG. 8, the prediction accuracy of these models trained by the generated dataset was compared. As seen, the best model accuracy for both train and test sets is achieved by TAG-GCN meaning that the neighbor graph aggregation and temporal features described herein have a considerable impact on the learning capability of the model.

TABLE 2: Properties of Different Models

In the following experiments, the models presented in Table II were used to recommend the actions. The actions in day t ₁, i.e., October 17, had been set to the default action which is -100 dBm. Unless otherwise mentioned, the action for the second day, i.e., October 18, was initialized by a set of random actions around the default action which is -100 dBm in the range of [-105, -95] . The model was trained iteratively in each day and was used to recommend actions for the next day. The action recommendation process is depicted in FIG. 9, where the states of the cells at day t _i are given to the trained model and given the states and predicted performance metrics of the network at day t _i+1, the action for day t _i is adjusted for each cell. Finally, the exact network performance measurements for day t _i+1 are computed by the cellular network simulator and used for the next day.

In addition to the result achieved by the actions recommended by the models, three baseline performance bounds achieved by the default A2 threshold, the expert rule, and the optimal values of the simulator were used. As previously noted, the default A2 threshold value is -100 dBm and this is used as the lower bound in the following experiments. The optimal actions in the simulator were obtained by a brute-force search and the upper performance bound was introduced. The expert rule-based action recommendation is a simple rule presented in Ye et al., “User association for load balancing in heterogeneous cellular networks” , IEEE Transactions on Wireless Communications, vol. 12, no. 6, pp. 2706–2716, 2013. The performance achieved by the expert rule was closer to the best performance when compared to the default action. However, there was still a significant gap that the method described herein in accordance with some embodiments reduces.

In the following experiments, the performance of the network under the actions recommended by different models was compared in terms of the network throughput computed as the ratio of total bits transmitted over all cells to the total transmission time of all cells and the difference to the baseline was found, which is the performance achieved by the default A2 threshold value.

In FIG. 10, the overall performance of the cellular network was compared in terms of the network throughput as presented above, for all the models listed in Table II. Several experiments were run, each of which use the same set of random actions on October 18 for all the models, to compare the performance. The performance achieved through the expert rule action recommendation, default action, and the best action of the simulator is also shown. As it can be seen, the TAG-GCN model can achieve the best performance and outperforming other methods. It should be noted that the TAG-GCN model has the lowest variance in the achieved performance showing its stability in the parameter configuration in the network. As expected, all the models can beat the experts rule algorithm which is highly dependent on the human experience and unable to compensate for performance degradation caused by bad random initialization on the first day.

In FIG. 11, an ablation study on the effectiveness of the proposed action recommendation strategy is shown. Multiple experiments were run with TAG-GCN, all initialized with random actions at day t2 but follow different action configuration strategies. As described herein, the novel action configuration strategy optimizes the load balancing in each cluster and also keeps the throughput of each cell as high as possible. Two more approaches for action configuration are considered, one considering only the cell throughput in the optimization problem, and one considering only the load-balancing. As shown, the proposed TAG-GCN model with both load-balancing and throughput can outperform others and at the same time have the lowest variance in the achieved result.

FIG. 12 presents an ablation study on the effectiveness of different initialization schemes on the TAG-GCN model for day t ₂, i.e., October 18. In addition to the random initialization, two more action initialization schemes including the expert rule and the negative slope are shown. The expert rule initialization is as provided in Ye et al., “User association for load balancing in heterogeneous cellular networks, ” IEEE Transactions on Wireless Communications, vol. 12, no. 6, pp. 2706–2716, 2013, where at day t, the actions are set for each cell v∈N _t based on the load balancing metric as

and where

and

are the weights set by the human experts. The negative slope procedure for the action initialization is defined based on the throughput ratio of the first day, i.e., t _i, of each cell v, i.e.,

as

where

and

is the number of cells in the dataset at day t ₁. For controlling the variation of actions, the maximum range of change is defined as

It should be also noted that [·] is the rounding operation. It is obvious from FIG. 12 that, regardless of the type of the initialization, the model can compensate for performance degradation in the first days and converge to almost the same performance in the final days of the experiment.

In FIG. 13, the impact of the proposed model and action configuration solution on the load balancing of the network is shown. The average throughput ratio of different clusters is shown for all days of the experiment. As seen, towards the final days of the experiment, the average throughput ratio keeps decreasing until it converges to a value close to one which shows the effectiveness of the proposed algorithm.

The model disclosed herein in accordance with exemplary embodiments leverage the states of the neighbors in each cluster by following the graph aggregation method described herein. The model considers the temporal features of the cells to learn the trend of features for parameter configuration. These two main properties of the model allow for the use of the underlying temporal and spatial dependencies in the network to configure the parameters. As an example, the A2 threshold parameter, which controls the handovers between the cells in the cellular network and affects the balance of the traffic between the cells, can be configured. Two objectives that are highly affected by the handover parameter configuration, namely, load balancing, and throughput of the cells, are considered. The effectiveness of this neighbor graph aggregation and temporal model is shown to increase the learning capability of the model and improve the overall performance of the network.

FIG. 14 shows various physical and logical components of an exemplary computing system 200 for training and using a model for configuring handover parameters for a wireless communication network in accordance with an embodiment of the present disclosure. Although an example embodiment of the computing system 200 is shown and discussed below, other embodiments may be used to implement examples disclosed herein, which may include components different from those shown. Although FIG. 14 shows a single instance of each component of the computing system 200, there may be multiple instances of each component shown. The example computing system 200 may be part of, or connected to, a component in a wireless communication network on which parameters for triggering handover actions are stored.

The computing system 200 includes one or more processors 204, such as a central processing unit, a microprocessor, an application-specific integrated circuit (ASIC) , a field-programmable gate array (FPGA) , a dedicated logic circuitry, a tensor processing unit, a neural processing unit, a dedicated artificial intelligence processing unit, or combinations thereof. The one or more processors 204 may collectively be referred to as a processor 204. The computing system 200 may include a display 208 for outputting data and/or information in some applications, but may not in some other applications.

The computing system 200 includes one or more memories 212 (collectively referred to as “memory 212” ) , which may include a volatile or non-volatile memory (e.g., a flash memory, a random access memory (RAM) , and/or a read-only memory (ROM) ) . The non-transitory memory 212 may store machine-executable instructions for execution by the processor 204. A set of machine-executable instructions 216 defining a handover parameter configuration system and a model builder for the same (described herein) is shown stored in the memory 212, which may be executed by the processor 204 to perform the steps of the methods for configuring handover parameters in a wireless communication network described herein. The memory 212 may include other machine-executable instructions for execution by the processor 204, such as machine-executable instructions for implementing an operating system and other applications or functions.

The memory 212 stores the training database 220 that includes the action data used to train the model for configuring handover parameters as described herein.

The memory 208 may also store other data, information, rules, policies, and machine-executable instructions described herein, including a model builder module 228 for building a model 232 of the wireless communication network or a cluster thereof. A parameter configurator module 236 then uses the model 232 to set handover parameters 238 for the wireless communication network.

In some examples, the computing system 200 may also include one or more electronic storage units (not shown) , such as a solid state drive, a hard disk drive, a magnetic disk drive and/or an optical disk drive. In some examples, one or more datasets and/or modules may be provided by an external memory (e.g., an external drive in wired or wireless communication with the computing system 200) or may be provided by a transitory or non-transitory computer-readable medium. Examples of non-transitory computer readable media include a RAM, a ROM, an erasable programmable ROM (EPROM) , an electrically erasable programmable ROM (EEPROM) , a flash memory, a CD-ROM, or other portable memory storage. The storage units and/or external memory may be used in conjunction with memory 212 to implement data storage, retrieval, and caching functions of the computing system 200.

The components of the computing system 200 may communicate with each other via a bus, for example. In some embodiments, the computing system 200 is a distributed computing system and may include multiple computing devices in communication with each other over a network, as well as optionally one or more additional components. The various operations described herein may be performed by different computing devices of a distributed system in some embodiments. In some embodiments, the computing system 200 is a virtual machine provided by a cloud computing platform.

Although the components for both training and using the model 236 are shown as part of the computing system 200, it will be understood that separate computing devices can be used for training and using the audio-visual transformation network 20 for generating visual images from audio data.

The steps (also referred to as operations) in the flowcharts and drawings described herein are for purposes of example only. There may be many variations to these steps/operations without departing from the teachings of the present disclosure. For instance, the steps may be performed in a differing order, or steps may be added, deleted, or modified, as appropriate.

In other embodiments, the same approach described herein can be employed for other modalities.

FIG. 15 shows the flow of information from graph structure to final prediction in accordance with an embodiment. This flow path is used to form the training pipeline of two models for predicting the throughput

and the throughput ratio

for τ _in = {1, 2, … , t-1} . In the neighbor augmentation module, the number of selected neighbors in the Euclidean space is determined by the number of neighbors in the graph space. The auto-grouping module constructs M groups of neighbors by using the augmented neighbors set (demo example for l = 2 and |R| = 4) . Empty groups are filled with zeros or the average value from all the neighbors.

General

Through the descriptions of the preceding embodiments, the present invention may be implemented by using hardware only, or by using software and a necessary universal hardware platform, or by a combination of hardware and software. The coding of software for carrying out the above-described methods described is within the scope of a person of ordinary skill in the art having regard to the present disclosure. Based on such understandings, the technical solution of the present invention may be embodied in the form of a software product. The software product may be stored in a non-volatile or non-transitory storage medium, which can be an optical storage medium, flash drive or hard disk. The software product includes a number of instructions that enable a computing device (personal computer, server, or network device) to execute the methods provided in the embodiments of the present disclosure.

All values and sub-ranges within disclosed ranges are also disclosed. Also, although the systems, devices and processes disclosed and shown herein may comprise a specific plurality of elements, the systems, devices and assemblies may be modified to comprise additional or fewer of such elements. Although several example embodiments are described herein, modifications, adaptations, and other implementations are possible. For example, substitutions, additions, or modifications may be made to the elements illustrated in the drawings, and the example methods described herein may be modified by substituting, reordering, or adding steps to the disclosed methods.

Features from one or more of the above-described embodiments may be selected to create alternate embodiments comprised of a sub-combination of features which may not be explicitly described above. In addition, features from one or more of the above-described embodiments may be selected and combined to create alternate embodiments comprised of a combination of features which may not be explicitly described above. Features suitable for such combinations and sub-combinations would be readily apparent to persons skilled in the art upon review of the present disclosure as a whole.

In addition, numerous specific details are set forth to provide a thorough understanding of the example embodiments described herein. It will, however, be understood by those of ordinary skill in the art that the example embodiments described herein may be practiced without these specific details. Furthermore, well-known methods, procedures, and elements have not been described in detail so as not to obscure the example embodiments described herein. The subject matter described herein and in the recited claims intends to cover and embrace all suitable changes in technology.

Although the present invention and its advantages have been described in detail, it should be understood that various changes, substitutions and alterations can be made herein without departing from the invention as defined by the appended claims.

The present invention may be embodied in other specific forms without departing from the subject matter of the claims. The described example embodiments are to be considered in all respects as being only illustrative and not restrictive. The present disclosure intends to cover and embrace all suitable changes in technology. The scope of the present disclosure is, therefore, described by the appended claims rather than by the foregoing description. The scope of the claims should not be limited by the embodiments set forth in the examples, but should be given the broadest interpretation consistent with the description as a whole.

Claims

A computer-implemented method for determining a threshold value for a handover (HO) parameter of a wireless communication network using a prediction model, the prediction model generated by:

selecting a subset of a plurality of neighbor cells of a center cell in a cluster of cells of a wireless communication network;

forming an augmented set of neighbor cells by using a feature transformation on the cluster of cells to map the cluster of cells to a latent space, and augmenting the subset of the plurality of neighbor cells with a set of additional neighbor cells closest to the center cell in the latent space;

dividing the augmented set of neighbor cells into groups;

applying a permutation invariant function to each of the groups to generate an output for each of the groups;

aggregating the outputs; and

determining the threshold value for the HO parameter at least partially based on the aggregated outputs.
The method of claim 1, wherein the HO parameter is a threshold power below which an HO is triggered.
The method of claim 1 or claim 2, wherein the subset of the plurality of neighbor cells represents the neighbor cells with which the center cell has handovers exceeding a threshold.
The method of claim 1 or claim 2, wherein the subset of the plurality of neighbor cells represents a fixed number of the neighbor cells with which the center cell has the highest number of handovers.
The method of any one of the previous claims, wherein the set of additional cells is equal in number to the number of cells in the subset of the plurality of neighbor cells.
The method of any one of the previous claims, wherein the permutation invariant function includes a set of learnable weights for each neighbor cell in the group.
The method of any one of the previous claims, further comprising:

grouping samples collected in a time interval t into K groups by temporal order;

determining an average network state for the samples in each of the k groups;

processing the average network state for the samples in the k groups arranged in temporal order using a recurrent neural network to generate a temporal feature model; and

determining the threshold value for the HO parameter at least partially based on the temporal feature model.
The method of claim 7, wherein the prediction model is trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, the throughput ratio being used as training labels for training the prediction model.
The method of any one of claims 1 to 8, wherein the prediction model is trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, a throughput of the center cell and the groups of neighbor cells being used as training labels for training the prediction model.
The method of any one of claims 8 and 9, wherein the predicted threshold value is formulated as the output of a non-linear transformation function of the state vector and a change in the threshold for the previous day.
The method of claim 8, wherein the state vector is a throughput ratio state vector, and wherein the prediction model is trained by concatenating temporal features of the center cell and a throughput of the center cell to form a throughput state vector.
The method of claim 10 or 11, wherein, during training of the prediction model,

and

are minimized, wherein LOSS ₁ is a throughput ratio at time t for cell v,
is a throughput ratio of a throughput for cell v at time t+1 relative to an average throughput for the augmented set of cells,
is a throughput ratio of a throughput for cell v at time t relative to an average throughput for the augmented set of cells,
is the throughput for cell v at time t+1,
is the throughput for cell v at time t, λ ₁ and λ ₂ are hyperparameters chosen for normalization, and θ ₁ and θ ₂ represent all trainable parameters in the prediction model.
A computing system for determining a threshold value for a handover (HO) parameter of a wireless communication network using a prediction model, the computing system comprising:

a processor configured to generate the prediction model by:

selecting a subset of a plurality of neighbor cells of a center cell in a cluster of cells of a wireless communication network;

forming an augmented set of neighbor cells by using a feature transformation on the cluster of cells to map the cluster of cells to a latent space, and augmenting the subset of the plurality of neighboring cells with a set of additional neighbor cells closest to the center cell in the latent space;

dividing the augmented set of neighbor cells into groups;

applying a permutation invariant function to each of the groups to generate an output for each of the groups;

aggregating the outputs; and

determining the threshold value for the HO parameter at least partially based on the aggregated outputs.
The computing system of claim 13, wherein the HO parameter is a threshold power below which an HO is triggered.
The computing system of claim 13 or claim 14, wherein the subset of the plurality of neighbor cells represents the neighbor cells with which the center cell has handovers exceeding a threshold.
The computing system of claim 13 or claim 14, wherein the subset of the plurality of neighbor cells represents a fixed number of the neighbor cells with which the center cell has the highest number of handovers.
The computing system of any one of claims 13 to 16, wherein the set of additional cells is equal in number to the number of cells in the subset of the plurality of neighbor cells.
The computing system of any one of claims 13 to 17, wherein the permutation invariant function includes a set of learnable weights for each neighbor cell in the group.
The computing system of any one of claims 13 to 18, wherein the processor is configured to generate the prediction model by:

grouping samples collected in a time interval t into K groups by temporal order;

determining an average network state for the samples in each of the k groups;

processing the average network state for the samples in the k groups arranged in temporal order using a recurrent neural network to generate a temporal feature model; and

determining the threshold value for the HO parameter at least partially based on the temporal feature model.
The computing system of claim 19, wherein the prediction model is trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, the throughput ratio being used as training labels for training the prediction model.
The computing system of any one of claims 13 to 20, wherein the prediction model is trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, a throughput of the center cell and the groups of neighbor cells being used as training labels for training the prediction model.
The computing system of any one of claims 20 and 21, wherein the predicted threshold value is formulated as the output of a non-linear transformation function of the state vector and a change in the threshold for the previous day.
The computing system of claim 20, wherein the state vector is a throughput ratio state vector, and wherein the prediction model is trained by concatenating temporal features of the center cell and a throughput of the center cell to form a throughput state vector.
The computing system of claim 22 or 23, wherein, during training of the prediction model,

and

are minimized, wherein LOSS ₁ is a throughput ratio at time t for cell v,
is a throughput ratio of a throughput for cell v at time t+1 relative to an average throughput for the augmented set of cells,
is a throughput ratio of a throughput for cell v at time t relative to an average throughput for the augmented set of cells,
is the throughput for cell v at time t+1,
is the throughput for cell v at time t, λ ₁ and λ ₂ are hyperparameters chosen for normalization, and θ ₁ and θ ₂ represent all trainable parameters in the prediction model.
A non-transitory machine-readable medium having tangibly stored thereon executable instructions for execution by one or more processors, wherein the executable instructions, in response to execution by the one or more processors, cause the one or more processors to determine a threshold value for a handover (HO) parameter of a wireless communication network using a prediction model, the prediction model generated by:

selecting a subset of a plurality of neighbor cells of a center cell in a cluster of cells of a wireless communication network;

forming an augmented set of neighbor cells by using a feature transformation on the cluster of cells to map the cluster of cells to a latent space, and augmenting the subset of the plurality of neighboring cells with a set of additional neighbor cells closest to the center cell in the latent space;

dividing the augmented set of neighbor cells into groups;

applying a permutation invariant function to each of the groups to generate an output for each of the groups;

aggregating the outputs; and

determining the threshold value for the HO parameter at least partially based on the aggregated outputs.
The non-transitory machine-readable medium of claim 25, wherein the HO parameter is a threshold power below which an HO is triggered.
The non-transitory machine-readable medium of claim 25 or claim 26, wherein the subset of the plurality of neighbor cells represents the neighbor cells with which the center cell has handovers exceeding a threshold.
The non-transitory machine-readable medium of claim 25 or claim 26, wherein the subset of the plurality of neighbor cells represents a fixed number of the neighbor cells with which the center cell has the highest number of handovers.
The non-transitory machine-readable medium of any one of claims 25 to 28, wherein the set of additional cells is equal in number to the number of cells in the subset of the plurality of neighbor cells.
The non-transitory machine-readable medium of any one of claims 25 to 29, wherein the permutation invariant function includes a set of learnable weights for each neighbor cell in the group.
The non-transitory machine-readable medium of any one of claims 25 to 30, wherein the executable instructions, in response to execution by the one or more processors, cause the one or more processors to:

group samples collected in a time interval t into K groups by temporal order;

determine an average network state for the samples in each of the k groups;

process the average network state for the samples in the k groups arranged in temporal order using a recurrent neural network to generate a temporal feature model; and

determine the threshold value for the HO parameter at least partially based on the temporal feature model.
The non-transitory machine-readable medium of claim 31, wherein the prediction model is trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, the throughput ratio being used as training labels for training the prediction model.
The non-transitory machine-readable medium of any one of claims 25 to 32, wherein the prediction model is trained by concatenating temporal features of the center cell, a throughput ratio of a throughput of the center cell to an average throughput of the augmented set of neighbor cells, and extracted information from the groups of neighbor cells to form a state vector, a throughput of the center cell and the groups of neighbor cells being used as training labels for training the prediction model.
The non-transitory machine-readable medium of any one of claims 32 and 33, wherein the predicted threshold value is formulated as the output of a non- linear transformation function of the state vector and a change in the threshold for the previous day.
The non-transitory machine-readable medium of claim 32, wherein the state vector is a throughput ratio state vector, and wherein the prediction model is trained by concatenating temporal features of the center cell and a throughput of the center cell to form a throughput state vector.
The non-transitory machine-readable medium of claim 34 or 35, wherein, during training of the prediction model,

and

are minimized, wherein LOSS ₁ is a throughput ratio at time t for cell v,
is a throughput ratio of a throughput for cell v at time t+1 relative to an average throughput for the augmented set of cells,
is a throughput ratio of a throughput for cell v at time t relative to an average throughput for the augmented set of cells,
is the throughput for cell v at time t+1,
is the throughput for cell v at time t, λ ₁ and λ ₂ are hyperparameters chosen for normalization, and θ ₁ and θ ₂ represent all trainable parameters in the prediction model.