Learning device, control device, and learning method

Abstract

This learning device for training a model of a neural network for executing base station sleep control in a mobile network includes a learning unit that, with a weight of the model trained for a specific base station serving as an initial value, re-learns all weights of the model using additional data in another base station, thereby tuning the model for the other base station.

Description

Learning device, control device, and learning method

[0001] This invention relates to power control technology in mobile networks.

[0002] With the widespread use of video streaming on mobile devices, there is a growing demand for high-quality communication on cellular networks (mobile networks). Meanwhile, telecommunications carriers need to minimize power consumption to reduce environmental impact and operational costs.

[0003] More than half of the total power consumption of a mobile network is consumed by base stations. Therefore, sleep control, which reduces power consumption by temporarily putting the base station antennas to sleep, has been widely explored.

[0004] Base stations are generally designed based on peak traffic volume. Therefore, traffic utilization tends to be lower in certain areas or time zones. As a result, several sleep control methods have been proposed that utilize this tendency to maximize overall throughput while minimizing power consumption. However, these methods do not take into account the constraint of maintaining a certain level of quality.

[0005] Therefore, Non-Patent Documents 1 and 2 propose a sleep control method that minimizes power consumption while satisfying constraints such as maintaining throughput above a certain value as much as possible. In this method, deep reinforcement learning is used to generate an agent (a neural network model) that determines whether or not to put the antennas belonging to each base station into sleep mode.

[0006] Rie Tagyo, Hideaki Kinsho, Akihiro Shiozu, and Kazuhisa Yamagishi, "A Study on Base Station Sleep Control with Communication Quality Constraints for Power Consumption Reduction," IEICE 2024 General Conference, B-14-25, Mar, 2024.

[0007] Applying conventional methods, such as those disclosed in Non-Patent Documents 1 and 2, to actual mobile networks requires training an agent for each of the multiple base stations. However, there are hundreds of thousands of base stations in total. Therefore, simply training an agent for all of these base stations would incur a huge training cost.

[0008] This invention has been made in view of the above points, and aims to provide a technology for significantly reducing the learning cost in sleep control for mobile networks, while maintaining the compliance rate with quality constraints and power reduction efficiency at the same level as conventional methods.

[0009] According to the disclosed technology, a learning device is provided for learning a neural network model for performing base station sleep control in a mobile network, the learning device comprising a learning unit that tunes the model for another base station by retraining all the weights of the model using additional data from another base station, with the weights of the model learned for a specific base station as initial values.

[0010] According to the disclosed technology, it is possible to significantly reduce the learning cost in sleep control for mobile networks while maintaining the same level of compliance with quality constraints and power reduction efficiency as conventional methods.

[0011] It is a diagram showing an example of the system configuration in the embodiment of the present invention. It is a flowchart for explaining the operation during learning. It is a diagram showing an example of data stored in the learning parameter DB 102. It is a diagram showing an example of data stored in the agent DB 301. It is a diagram showing an example of data stored in the base station / antenna data DB 302. It is a diagram showing an example of the learning algorithm in the agent learning module 103. It is a flowchart for explaining the operation during application. It is a diagram showing experimental results. It is a diagram showing experimental results. It is a diagram showing experimental results. It is a diagram showing experimental results. It is a diagram showing an example of the hardware configuration of the device.

[0012] Hereinafter, embodiments (this embodiment) of the present invention and numerical experiment examples will be described with reference to the drawings. The embodiments described below are merely examples, and the embodiments to which the present invention is applied are not limited to the following embodiments. Specifically, for example, the control target of the sleep control in this embodiment may be the entire base station or each antenna included in the base station. Also, the agent described below is a model of a neural network. The "agent" may be referred to as a "model".

[0013] (Outline of the embodiment) In this embodiment, attention is paid to the fact that the characteristics of the learned agents are somewhat similar between different base stations. The agent learning unit 103 described later tunes the agent learned at a specific base station using information from different base stations, thereby generating an agent that maintains the same control accuracy as the conventional one. Since a desired agent is generated by tuning the already learned agent, the learning cost can be significantly reduced compared to the prior art.

[0014] (Example of device configuration) FIG. 1 shows an example of the configuration of the system in this embodiment. This system (which may also be called a control device) includes an agent learning unit 100, a control unit 200, a data holding unit 300, and a data collection unit 400. Also, FIG. 1 shows the base station / antenna 500 that is the control target. "Base station / antenna" means a base station or an antenna.

[0015] The agent learning unit 100 includes a learning parameter setting module 101, a learning parameter database 102, and an agent learning module 103. The control unit 200 includes an agent application module 201 and a control module 202.

[0016] The data storage unit 300 includes an agent database 301 and a base station / antenna data database 302. The data collection unit 400 includes a data collection module 401.

[0017] Note that "the agent learning unit 100, the control unit 200, the data storage unit 300, and the data collection unit 400" may constitute one device (computer), or one or more functional units may constitute one device.

[0018] For example, the agent learning unit 100, the control unit 200, the data storage unit 300, and the data collection unit 400 may each be one device. In this case, the agent learning unit 100, the control unit 200, the data storage unit 300, and the data collection unit 400 may be referred to as an agent learning device 100, a control device 200, a data storage device 300, and a data collection device 400, respectively. The agent learning device may also be referred to as a learning device.

[0019] Hereinafter, the operations of each unit shown in FIG. 1 will be described by dividing them into two cases: the operation during learning and the operation during application. Note that "during application" may be replaced with "during inference", "during control", or "during determination".

[0020] (During learning) The operation during learning will be described according to the procedure of the flowchart in FIG. 2.

[0021] <S1 (Step 1)> Learning starts when a learning update trigger is fired. This trigger may be fired when a new base station or a new antenna is introduced, or may be fired periodically or in response to some event. Further, this learning update trigger may include information indicating a base station to be updated or information indicating an antenna to be updated, or based on the learning update trigger, all base stations may be updated at once.

[0022] <S2> When the learning update trigger is fired, the learning parameter setting module 101 obtains the parameters to be used for learning from the learning parameter DB 102.

[0023] Figure 3 shows an example of the data stored in the learning parameter DB 102. As shown in Figure 3, the learning parameter DB 102 stores the number of layers, the size of the hidden layer, the learning rate, the number of learning episodes, the number of learning steps, the learning data utilization rate, the initial weight values, and the convergence threshold for the reward function.

[0024] It should be noted that the data exemplified here is merely an example, and the data stored in the learning parameter DB102 is not limited to the parameters shown in Figure 3.

[0025] In the learning parameter DB102, the column for the learning data utilization rate shall contain a real value greater than 0 and less than or equal to 1, and the column for the initial weight value shall contain the name of the agent ID in the agent DB301 or null.

[0026] <S3> If the column for the initial weight value is not null, the learning parameter setting module 101 queries the agent DB 301 and obtains the agent corresponding to the agent ID. This agent is a trained agent that has been trained at a specific base station.

[0027] An example of data stored in Agent DB 301 is shown in Figure 4. In the example in Figure 4, trained agents are stored in Agent DB 301 as binary data in pth format for each agent ID. Note that the data format is not limited to pth format.

[0028] The learning parameter setting module 101 passes the learning parameters obtained in S2 and the agent obtained in S3 to the agent learning module 103.

[0029] <S4> The agent learning module 103 uses the learning parameters passed from the learning parameter setting module 101 to retrain the agent passed from the learning parameter setting module 101, for example, using the reinforcement learning method disclosed in Non-Patent Documents 1 and 2. Here, all weights in the agent are retrained. The trained agent learned by the agent learning module 103 is stored in the agent DB 301. The learning in the agent learning module 103 may also be called tuning. The agent learning module 103 may also be called the learning unit.

[0030] The reinforcement learning method used in this embodiment is not limited to the methods disclosed in Non-Patent Documents 1 and 2.

[0031] If a non-null value is entered as the initial weight value for the agent, that value will be treated as the initial value for that agent. The base station from which that initial value was obtained is different from the base station targeted for learning in S4.

[0032] Furthermore, during training, data such as throughput distribution and resource block utilization stored in the base station / antenna data DB302 may also be used. An example of data stored in the base station / antenna data DB302 is shown in Figure 5.

[0033] In Figure 5, the cells to be slept are stored as an array, with 0 representing cells that will not sleep and 1 representing cells that will sleep. Furthermore, the sleep states are represented in alphabetical order according to the antenna names. You may also think of "cells" as "antennas."

[0034] The throughput distribution represents the probability density of throughput in each bin under the given base station ID at the given time. The division of the bins can be arbitrary; for example, they may be [0.0, 1.0), [1.0, 2.0), [2.0, 3.0], [3.0, 4.0], [4.0, 5.0]. However, the unit in this example is Mbps.

[0035] Note that the throughput distribution may be the distribution of the resource block utilization rate. At this time, the binning method may be, for example, [0.0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), [0.8, 1.0]. However, it is assumed that the resource block is normalized within the range of [0.0, 1.0] at this time. Also, the bin width does not have to be equally spaced.

[0036] The specific learning method shall conform to, for example, the learning methods disclosed in Non-Patent Documents 1 and 2, but is not limited thereto. As long as it is a reinforcement learning method using rewards, learning methods other than those disclosed in Non-Patent Documents 1 and 2 may be used.

[0037] (Example of specific learning method) Here, an example of the learning method will be described. In this example, in a mobile network (referred to as an environment) that is a control target, consider the problem of selecting the activation / stopping of a plurality of antennas installed at one base station that provide different frequency bands. Under the jurisdiction of the base station, there is one or more UEs (User Equipment) that communicate with the base station. The UE may also be referred to as a terminal.

[0038] Let the antenna set be A ｃａｐａｃｉｔｙ = {1, 2,..., A ａｌｌ}, among which, from the perspective of maintaining connectivity, the antenna set that cannot be stopped is A ｃｏｖｅｒａｇｅ = {1, 2,..., A ｃｏｖｅｒａｇｅ}, and the antenna set that can be stopped is A ｃａｐａｃｉｔｙ = {1, 2,..., A ｃａｐａｃｉｔｙ}. This means that the coverage area where the antenna set A ｃａｐａｃｉｔｙ can provide services is covered by the coverage areas of at least one or more antennas in the antenna set A ｃｏｖｅｒａｇｅ .

[0039] Let the antenna set that has not been stopped be A ｏｎ (∈ A ａｌｌ ). In sleep control, the antenna set A ｏｎ is determined for each control time. When the antenna in A ｃａｐａｃｉｔｙ stops, the UE connected to the stopped antenna is A ｏｎOf these, the handover will be to one of the antennas that has an overlapping coverage area with the antenna that has stopped transmitting.

[0040] The number of resource blocks (RBs) each antenna possesses is B. ｔｏｔａｌ The RB utilization rate of antenna i at time t is ρ ｉ Let (t) ∈ [0, 1].

[0041] Power consumption P of all antennas at time t ｃ This can be expressed by the following formula.

[0042] Here, P ｉ represents the maximum power consumption of antenna i, and q ｉ ∈(0,1) represents the proportion of power consumed fixedly out of the maximum power consumption. However, different power consumption models may be used.

[0043] Antenna set A ａｌｌ Let N = {1, 2, ..., N} be the set of UEs that provide the service. The power consumption P is calculated as follows: If all antennas are not shut down (i.e., no antennas are shut down) at time t when traffic demand is met. ａ (t) is expressed by the following formula.

[0044] Here, ^ρ ｉ (t) is the RB utilization rate of antenna i at time t if all antennas were not shut down. The power reduction P(t) at time t is P(t) = 1 - P ｃ (t) / P ａ It is represented as (t).

[0045] Let Q(t) be the communication quality index at time t. Q(t) is a function consisting of one or more quality indexes that the network operator should consider. However, it may be set so that a larger Q(t) indicates better communication quality, and may be expressed as a function represented by a weighted linear sum of multiple indexes, for example.

[0046] Here, the quality metric is set as average throughput, and the throughput of UE at time t is set to T. ｎ If we let (t), the average throughput Q(t) at time t can be expressed by the following formula.

[0047] The set quality objective is that the throughput of each UE is T ｔａｒｇｅｔ This concludes the analysis. However, other indicators such as delay may be added or replaced as quality indicators, and statistics other than the mean may be used.

[0048] The agent learning module 103 performs action a through reinforcement learning. ｔ We train an agent that outputs [this]. The parameters (which can also be called weights) of the agent (a neural network model) correspond to the policy π.

[0049] Action a at time t ｔ A set of antennas capable of shutting down the signal ｃａｐａｃｉｔｙ Let this be a vector representing the state of the antenna at time t, either activated or deactivated. For example, if activated is 1 and deactivated is 0, then action a ｔ is, A ｃａｐａｃｉｔｙ This will be a binary vector of the number of elements. Note that if an antenna transitions to the same state, such as being in the activated state at time t and also being in the activated state at time t+1, the operation will not be executed.

[0050] State s at time t ｔ Let be a vector representing the utilization rate of all antennas and the combination of activation / deactivation of all antennas. Also, the reward r at time t ｔ The power reduction amount P(t) and average throughput Q(t) are weighted parameters w １ The summation is performed using the following. The reward may also be called the reward function. Also, all UE ｎ In T ｎ ≧T ｔａｒｇｅｔ The constraint is that the following must be satisfied. Based on these, the control device 100 performs action a at time t+1. ｔ＋１ Learn the policies for defining them. The problem we want to solve can be expressed by the following formula.

[0051] Here, γ∈(0,1) refers to the discount rate for future rewards, E ｓ，π [*] represents the expected value regarding the state and policy. s.t. is the constraint mentioned above, and the reward rｔ As mentioned above, the power reduction amount P(t) and the average throughput Q(t) are weighted by the parameter w １ It is the sum of the two values ​​used.

[0052] The agent learning module 103 learns a policy π to find the solution to the above problem. The learning environment (i.e., the controlled object) may be an actual base station or a base station on a simulator.

[0053] Specifically, the agent learning module 103 executes multiple learning steps for each learning episode.

[0054] In each learning step, the agent learning module 103 selects an action a from the agent with the current policy π. ｔ Sample the behavior a ｔ The agent learning module 103 performs action a ｔ From the environment in which it was executed, state s ｔ＋１ Observe and reward r ｔ The agent learning module 103 calculates the transition (s ｔ , a ｔ ,r ｔ ,s ｔ＋１ ) is stored in replay buffer D.

[0055] After executing a predetermined number of learning steps, the agent learning module 103 updates the policy π (i.e., the agent's parameters) using the transition data stored in the replay buffer D.

[0056] The above "a predetermined number of learning steps and policy π update" constitutes one episode. The agent learning module 103 executes a predetermined number of episodes.

[0057] (Method for determining the number of learning steps and learning episodes) In this embodiment, there are two examples (Example 1 and Example 2) of how to determine the number of learning steps and learning episodes, and each will be explained below.

[0058] Example 1: In Example 1, the method for determining the number of training episodes and training steps is to use a specified training data utilization rate. In this method, the training data utilization rate is multiplied by either the training episode count or the training step count, or both. This makes it possible to maintain a certain level of control quality by utilizing already trained weight information, while maintaining equivalent control accuracy with a smaller amount of data than before.

[0059] For example, in the data example shown in Figure 3, the number of learning episodes is 1000 × 0.5 = 500, and the number of learning steps is 30 × 0.5 = 15.

[0060] Example 2: In Example 2, the method for determining the number of training episodes and training steps is to use the convergence threshold of the reward function. In the method in Example 2, the reward value obtained by the most recent agent is recorded after each training episode is completed. When the episode progresses, if the degree of change is within the convergence threshold, training is considered complete, and the agent at that time is output.

[0061] The processing procedure of the agent learning module 103 in Example 2 will be explained with reference to the flowchart in Figure 6.

[0062] In S101, the agent learning module 103 uses the previous reward function r ｌａｔｅｓｔ and the current reward function r ｎｏｗ Initialize the previous reward function r. ｌａｔｅｓｔ Let r be inf (infinity), and the current reward function r ｎｏｗ Set this to 0. Then, for example, perform the learning using the method described in the specific example above.

[0063] The agent learning module 103 executes S102 each time an episode is completed. In S102, the agent learning module 103 executes "|r ｎｏｗ -r ｌａｔｅｓｔThe system determines whether the value has reached <ε, or whether the number of training episodes has reached the specified number of episodes. ε is a predetermined convergence threshold. If the result is No, the system proceeds to S103; if the result is Yes, the process terminates.

[0064] In S103, the agent learning module 103, ｌａｔｅｓｔ r ｎｏｗ Update with . Furthermore, the agent learning module 103 performs additional learning for one episode, r ｎｏｗ Update.

[0065] The procedure in Example 2 allows for training agent weights using a large amount of data when the characteristics of the source base station and the target base station are significantly different, and training agent weights using a small amount of data when the characteristics of the source base station and the target base station are similar.

[0066] (During Application) The operation during application will be explained according to the steps in the flowchart in Figure 7. Here, it is assumed that the agent learning module 103 modifies the sleep control status for a base station using an agent tuned for that base station.

[0067] <S11> When changing the sleep control status, the control module 202 sends an update request to the agent application module 201. This update request may be sent for some base stations or for all base stations.

[0068] <S12> Upon receiving the update request, the agent application module 201 retrieves the agent data (e.g., neural network weight parameters) and base station / antenna data of the base station that is the target of the update request from the agent DB 301 and base station / antenna data DB 302. The agent application module 201 applies the base station / antenna data to the agent to obtain a list of appropriate sleep target antennas and returns the result to the control module 202.

[0069] <S13> The control module 202 performs sleep control of the actual base station and antenna based on the received list. The data acquisition module 401 acquires the control results from the base station / antenna 500 and stores the control results (antenna data, etc.) in the base station / antenna data DB 302. Note that the "agent application module 201 and control module" may also be called the "control unit".

[0070] (Example of numerical experiment) An example of an experiment in which the technology according to this embodiment (configuration shown in Figure 1) was implemented in a simulation environment is shown. In this experiment, the number of macrocells was set to 1, and the number of capacity cells was varied from 2 to 4. Note that the macrocells are as described above. ｃｏｖｅｒａｇｅ It corresponds to the antenna, and the capacity cell is A ｃａｐａｃｉｔｙ This corresponds to the antenna. It determines whether to put only the capacity cells to sleep, while keeping the macrocells asleep.

[0071] The macrocell had a radius of 1000m, and the capacity cell had a radius of 900m, resulting in a sector-shaped coverage area with a central angle of 120°.

[0072] The maximum power consumption of each antenna was set to 100W, and the power consumed regardless of whether the antenna was ON or OFF was set to 60% of the total. The maximum number of resource blocks that each capacity antenna had was set to 100, and the parameter w shown in Non-Patent Documents 1 and 2 was set to w. １ As for gamma, lol １ The values ​​were set to 0.7 and γ = 0.9. The neural network had 3 layers, with a hidden layer size of 256, a batch size of 256, a replay buffer size of 10^5, and a learning rate of 3.0 * 10^-4.

[0073] First, to demonstrate the existence of the aforementioned challenges, Figures 8 to 10 show the results when two throughput target values ​​were arbitrarily selected from 1 Mbps, 3 Mbps, and 5 Mbps during training and application. The number of episodes and steps in these cases were 20,000 and 20, respectively.

[0074] Figures 8, 9, and 10 represent the results when the target throughput value during application is changed to 1 Mbps, 3 Mbps, and 5 Mbps. The vertical axis in each figure represents the constraint violation rate and takes a real value between 0.0 and 1.0. The horizontal axis represents each condition. The conditions are expressed in the notation (number of antennas) - (target throughput value during learning) - (method). There are two types of (methods): "safe7" and "safeES". The former is the method shown in Non-Patent Documents 1 and 2, and the latter is the optimal solution found by brute force search.

[0075] As can be seen from Figures 8 to 10, constraint violations occur when "3-1000-safe7" is evaluated with a throughput target of 3 Mbps, and when "4-1000-safe7" is evaluated with a throughput target of 5 Mbps.

[0076] Next, we focused on the case evaluated under the "4-1000-safe7" condition with a throughput target of 5 Mbps, and performed additional training with data at a throughput target of 5 Mbps. The method for determining the number of episodes for additional training followed the method described in Example 1 above, and the training data utilization rate was evaluated in two patterns: 0.5 and 0.1. The results are shown in Figure 11.

[0077] In Figure 11, "safe" shows the result of training and evaluation with a throughput target of 5 Mbps, "safe_0" shows the result of evaluation under the condition "4-1000-safe7", "safe_01" shows the result when the weights from training under "4-1000-safe7" are used as initial values ​​and 10% additional training data is used for evaluation, "safe_05" shows the result when 50% additional training data is used in the same way as safe_01, and "safeES" shows the optimal solution. From these results, it can be seen that constraint violations can be reduced to almost 0% by adding a small amount of training data.

[0078] (Example Hardware Configuration) Any of the devices described in this embodiment (control device, agent learning device 100, control device 200, data holding device 300, data acquisition device 400, etc.) can be realized, for example, by having a computer execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0079] In other words, the device can be realized by using hardware resources such as the CPU and memory built into a computer to execute a program corresponding to the processing performed by the device. The program can be recorded on a computer-readable recording medium (such as portable memory), saved, and distributed. It can also be provided via a network, such as the Internet or email.

[0080] Figure 12 shows an example of the hardware configuration of the computer described above. The computer in Figure 12 has a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., all of which are interconnected by bus B. The computer may also be equipped with a GPU.

[0081] The program that enables processing on the computer is provided on a recording medium 1001, such as a CD-ROM or memory card. When the recording medium 1001 containing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001; it may also be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files and data.

[0082] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when a program startup command is received. The CPU 1004 implements the functions related to the memory device 1003 according to the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) etc., based on a program. The input device 1007 consists of a keyboard and mouse, buttons, or a touch panel, etc., and is used to input various operation commands. The output device 1008 outputs the calculation results.

[0083] (Summary of Embodiments and Effects) As described above, the technology described in this embodiment makes it possible to significantly reduce the learning cost in sleep control for mobile networks while maintaining the compliance rate with quality constraints and power reduction efficiency at the same level as conventional methods.

[0084] The following additional information is disclosed regarding the embodiments described above.

[0085] <Notes> (Note 1) A learning device for learning a neural network model for performing base station sleep control in a mobile network, comprising a learning unit that tunes the model for the other base station by retraining all the weights of the model using additional data from another base station, with the weights of the model learned for a specific base station as initial values. (Note 2) The learning device according to Note 1, wherein the learning unit learns the model using a specified proportion of additional data from the other base station, or learns the model using additional data from the other base station until the value of the reward function converges within a specified range. (Note 3) A control device comprising a control unit that performs sleep control for the other base station using the model learned by the learning unit according to Note 1 or 2. (Note 4) A learning method performed by a learning device for learning a neural network model for performing base station sleep control in a mobile network, comprising a learning step of tuning the model for the other base station by retraining all the weights of the model using additional data from another base station, with the weights of the model learned for a specific base station as initial values.

[0086] Although this embodiment has been described above, the present invention is not limited to this specific embodiment, and various modifications and changes are possible within the scope of the gist of the invention as described in the claims.

[0087] 100 Agent learning unit 101 Learning parameter setting module 102 Learning parameter DB 103 Agent learning module 200 Control unit 201 Agent application module 202 Control module 300 Data holding unit 301 Agent DB 302 Base station / antenna data DB 400 Data acquisition unit 401 Data acquisition module 500 Base station / antenna 1000 Drive device 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims

1. A learning device for learning a neural network model for performing base station sleep control in a mobile network, comprising a learning unit that tunes the model for another base station by retraining all the weights of the model using additional data from another base station, with the weights of the model learned for a specific base station as initial values.

2. The learning device according to claim 1, wherein the learning unit learns the model using a specified proportion of additional data from the other base station, or learns the model using additional data from the other base station until the value of the reward function converges to a specified range.

3. A control device comprising a control unit that performs sleep control for the other base station using the model learned by the learning unit according to claim 1 or 2.

4. A learning method performed by a learning device for learning a neural network model for performing base station sleep control in a mobile network, comprising a learning step of tuning the model for another base station by retraining all the weights of the model using additional data from another base station, with the weights of the model learned for a specific base station as initial values.

Images