Baseline-based automatic edge computing network resource allocation method

Abstract

The invention belongs to the field of mobile edge computing, and discloses a baseline-based automatic edge computing network resource allocation method. Comprising the following steps: acquiring channel gains between each wireless device and each hybrid access point; inputting the channel gain into a discrete network branch prediction unloading decision; inputting the channel gain and the unloading decision into a first continuous network branch and a second continuous network branch which are parallel, and predicting an energy distribution proportion and an unloading time distribution proportion; taking a parallel branch after historical iteration training as a baseline, and outputting a historical prediction result according to the channel gain of the current time slot; respectively calculating the maximum calculation rate according to the current time slot and the historical prediction result, and updating the discrete network branches; and calculating the maximum calculation rate of the current time slot according to the prediction result of the current time slot, and updating the first continuous network branch and the second continuous network branch. According to the invention, under the condition that the distributed renewable energy source has a wireless equipment function, a relatively high overall maximum calculation rate of the edge calculation network is achieved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of mobile edge computing, and particularly relates to a baseline-based automatic edge computing network resource allocation method. BACKGROUND

[0002] In recent years, the Internet of Things (IoT) has developed rapidly, and the scale of IoT devices has expanded rapidly. Related devices are continuously increasing in scenarios such as autonomous driving, smart home, and smart city. At the same time, artificial intelligence (AI) technology is also developing, making IoT devices more and more intelligent. Large-scale devices and highly intelligent technology have generated a large amount of computing data and real-time demand. Generally, IoT devices are limited by battery capacity and other factors and cannot complete a large number of high-performance computing tasks. Therefore, new technical paths are needed to improve computing power and energy self-sufficiency.

[0003] Mobile edge computing (MEC) technology has become a potential solution to this challenge. Deploying edge servers near terminal devices can assist or even replace terminal devices to complete task computing, which can significantly reduce latency and significantly improve real-time computing capability. For terminal devices with limited battery, distributed power based on renewable energy can provide continuous and stable clean energy supply for terminal devices, providing reliable energy guarantee for local computing and data offloading of terminal devices. For limited resources, how to optimize resources reasonably to maximize computing performance has become an effective way to break through the computing power of IoT devices. Currently, there are various algorithms to solve the problems existing in general IoT devices, such as dynamic programming-based algorithms to maximize computing completion rate and heuristic-based algorithms to minimize service latency. However, such algorithms often work well on small samples. If the size of the edge computing network expands, the search space will grow exponentially, and the algorithm performance will decrease, and the algorithm time consumption will increase significantly. There are also deep reinforcement learning (such as deep Q network) to minimize latency and energy consumption. Ordinary deep reinforcement learning (DRL) algorithms are only used to optimize offloading decisions and cannot integrate other resource optimization such as time allocation. Therefore, it is necessary to cooperate with other mathematical optimization algorithms. This way needs to design mathematical optimization algorithms for specific problems, and is only effective for convex optimization problems, and is difficult to apply to non-convex optimization problems. SUMMARY

[0004] The application aims to provide a baseline-based automatic edge computing network resource allocation method, which can achieve a higher maximum computing rate of the overall edge computing network in the case of distributed renewable energy for wireless device functions.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the application is:

[0006] A baseline-based automatic edge computing network resource allocation method, the edge computing network has a hybrid access point and a wireless device, comprising the following steps:

[0007] After the start of each time slot, the channel gain between each wireless device and each hybrid access point is obtained to form a channel set of the current time slot;

[0008] The channel set of the current time slot is input into a parallel discrete network branch to predict the offloading decision of each wireless device to obtain an offloading decision set of the current time slot;

[0009] The channel set and the offloading decision set of the current time slot are input into a parallel first continuous network branch to predict the energy allocation ratio of each wireless device;

[0010] The channel set and the offloading decision set of the current time slot are input into a parallel second continuous network branch to predict the offloading time allocation ratio of each wireless device;

[0011] After the historical iteration training, the parallel discrete network branch, the first continuous network branch and the second continuous network branch are taken as a baseline, and the historical offloading decision, energy allocation ratio and offloading time allocation ratio are output according to the channel set of the current time slot;

[0012] According to the offloading decision, energy allocation ratio and offloading time allocation ratio of the current time slot and the history, the maximum computing rate is calculated respectively to update the discrete network branch by forming a discrete loss function;

[0013] According to the offloading decision, energy allocation ratio and offloading time allocation ratio of the current time slot, the maximum computing rate of the current time slot is calculated to update the first continuous network branch and the second continuous network branch by forming a continuous loss function, and the training is repeated until all rounds.

[0014] Several optional modes are also provided below, but not as additional limitations to the above overall scheme, but only as further supplements or preferences, and each optional mode can be combined with the above overall scheme alone, and can also be combined between multiple optional modes, without technical or logical contradiction.

[0015] As preferred, the discrete network branch comprises a discrete network and a normalization layer, the discrete network comprises an input layer, a multi-scale pyramid network and an output layer;

[0016] The channel set of the current time slot is input into the input layer of the discrete network, and a first matrix is output by the output layer after passing through the multi-scale pyramid network, and a probability matrix is obtained after the first matrix is subjected to SoftMax normalization operation of the normalization layer, and the offloading decision of each wireless device is output by random offloading sampling according to the probability distribution in the probability matrix.

[0017] As preferred, the first continuous network branch comprises a first continuous network and a normalization layer, the first continuous network comprises an input layer, a multi-scale pyramid network and an output layer;

[0018] The channel set and the offloading decision set of the current time slot are input into the input layer of the first continuous network to form a two-channel matrix, and a second matrix is output by the output layer after passing through the multi-scale pyramid network, and the energy allocation ratio obtained by local calculation of all wireless devices is obtained after the second matrix is subjected to SoftMax normalization operation of the normalization layer.

[0019] As preferred, the second continuous network branch comprises a second continuous network and a normalization layer, the second continuous network comprises an input layer, a multi-scale pyramid network and an output layer;

[0020] The channel set and the offloading decision set of the current time slot are input into the input layer of the second continuous network to form a two-channel matrix, and a third matrix is output by the output layer after passing through the multi-scale pyramid network, and the offloading time allocation ratio of each wireless device is obtained by the SoftMax normalization operation of the element value corresponding to the wireless device offloaded to the same hybrid access point in the third matrix by the normalization layer according to the offloading decision set of the current time slot.

[0021] As preferred, the multi-scale pyramid network comprises a convolution pyramid layer, a pooling pyramid layer and a full connection layer;

[0022] The convolution pyramid layer comprises dilated convolutions with different convolution kernel sizes, the pooling pyramid layer comprises pooling pyramids, and each pooling pyramid performs downsampling operations on different pooling regions, the input data obtains feature maps after passing through the convolution pyramid layer, each feature map passes through a pooling pyramid, and finally feature maps are obtained, the feature maps are unfolded into one-dimensional vectors for splicing, and the spliced features obtain the output data of the multi-scale pyramid network after passing through the full connection layer.

[0023] Preferably, the baseline is obtained in the following way:

[0024] After each round of training iterations, the discrete network branches, the first continuous network branch, and the second continuous network branch are stored in a cache queue. If the number of cached branches exceeds the size of the cache queue, the earliest cached element is removed from the cache queue according to the first-in-first-out principle, and the first element in the cache queue is taken as the baseline for training.

[0025] Preferably, the step of forming a discrete loss function to update the discrete network branch includes:

[0026] Calculate the maximum computing rate of the current time slot based on the offloading decision, energy allocation ratio, and offloading time allocation ratio of the current time slot;

[0027] Calculate the historical maximum computing rate based on historical unloading decisions, energy allocation ratios, and unloading time allocation ratios;

[0028] The discrete loss function is calculated based on the maximum computation rate of the previous time slot and the historical maximum computation rate as follows:

[0029]

[0030] In the formula, For discrete loss functions, The probability matrix is ​​the nth List, The absolute value of the difference in maximum computing speed. This represents the maximum computation rate for the current time slot. This represents the highest computing speed in history. For decision parameters, if ,but ;otherwise .

[0031] Preferably, the continuous loss function is the reciprocal of the logarithm of the maximum computation rate of the current time slot.

[0032] The baseline-based automated edge computing network resource allocation method provided by this invention has the following advantages compared with the prior art:

[0033] Firstly, the application proposes a baseline-based optimization scheme effect evaluation method. Unlike the current supervised learning, which uses a basic algorithm as a baseline to measure the percentage point of performance improvement of the improved algorithm, the current baseline in most supervised learning is a static value, which only serves as a result comparison and does not participate in algorithm training and updating. The baseline proposed in the application is based on its own network and dynamically adjusts the corresponding baseline at different training rounds, which can evaluate the effect of the network output optimization strategy under the current state, the baseline value participates in the calculation of the loss function, provides a new measurement standard for unsupervised problems, guides the network parameter update, and makes the model converge quickly and improve the model performance.

[0034] Secondly, the application constructs a multi-scale pyramid network structure, which can convert input data of different dimensions into a unified feature vector by combining convolution pyramid and pooling pyramid. It can be applied to the case of changing number of nodes in edge computing network scene without re-designing network parameters for model training. In order to reduce the problem of too large decision space, the current common multi-agent algorithm divides the entire edge computing network into multiple sub-networks, and each agent is responsible for one sub-network. However, this method can only see part of the overall network, and the observation information is processed by the master node and the decision of the master node is executed. In this mode, the information obtained by the master network is heavily dependent on the sub-network, and the deviation of any sub-network will affect the judgment of the master network. It is difficult for the master network to obtain the original global information of the network, and it is easy to cause decision error. The parallel multi-network architecture adopted by the application can receive global information in each sub-network, and can independently calculate loss and update parameters, so that the network training process is more flexible, and the network can be easily debugged, thereby improving the model training efficiency and improving the final prediction effect of the model.

[0035] Finally, the application adopts a parallel multi-network architecture design, and the discrete network can predict the offloading decision in the optimization problem, and the parallel continuous network can predict the energy and time allocation. This method can conveniently adjust the number of networks and expand it to other optimization scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 It is an edge computing network structure schematic diagram;

[0037] Figure 2 It is a flowchart of the automatic edge computing network resource allocation method based on baseline of the application;

[0038] Figure 3 It is a structure schematic diagram of three parallel neural networks of the application;

[0039] Figure 4 It is a structure schematic diagram of the multi-scale pyramid network of the application. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0042] like Figure 1 As shown, the edge computing network in this embodiment has A hybrid access point and The system comprises several wireless devices, each with a hybrid access point equipped with an edge server and power lines, and powered by renewable energy sources. This embodiment employs Time Division Multiple Access (TDMA) technology for data communication between the wireless devices (terminal devices) and the edge server. When multiple terminal devices simultaneously offload data to the same edge server, different time slots are allocated to different wireless devices. During data offloading, a partial offloading mode is used, meaning the data on the wireless device can be arbitrarily divided. During data offloading, any proportion of data can be selected to be offloaded to the edge server, and each wireless device can only select one edge server for data offloading.

[0043] In one embodiment, the entire edge computing network scenario comprises the following components: 3 hybrid access points and 5 wireless devices; [Let the hybrid access point set...] and wireless device collection The entire system time is divided into multiple very short time slots. It is the length of a time slot. Within a time slot, the energy obtained by each wireless device from the renewable energy source is denoted as... , Indicates the first The renewable energy acquisition rate of a wireless device is typically measured in energy per unit time (e.g., joules per second). The wireless device and the first The channel gain between the hybrid access points is denoted as . .

[0044] During local task computation, each wireless device generates data that needs to be computed in each time slot. Of the wireless devices, the proportion of energy harvested used for local computing is [percentage missing]. Therefore, the remaining energy used for offloading to the hybrid access point is... ; Indicates the first The CPU (Central Processing Unit) computing speed of a wireless device, measured in cycles per second; This represents the number of loops required for the CPU to process 1 bit of data; therefore, in The total computational load of the data within a given time period can be expressed as: The CPU requires the following energy: The CPU consumes the following energy: Then you can get ; So, for the first The computing speed of a wireless device's local computation can be expressed as:

[0045]

[0046] in, For the first The computation rate of local computation by each wireless device, a constant. , This is the CPU's dynamic power consumption factor.

[0047] For task offloading computation, the wireless device needs to first offload the data to the hybrid access point, and then the hybrid access point performs the computation; for the first... For an individual wireless device, assuming it offloads data to a hybrid access point The percentage of time required for uninstallation is marked as follows. Therefore, the unloading time of all devices and the time for energy acquisition should not exceed the duration of the time slot. , can be represented as:

[0048]

[0049] If the The wireless device offloads the data to the first... On a hybrid access point If the bandwidth is represented by the offload rate, then the offload rate can be expressed as:

[0050]

[0051] in, Indicates the first The wireless device offloads the data to the first... offloading rate on a hybrid access point, constant , denotes the additional transmission data in the offloading process, denotes the Gaussian noise.

[0052] As Figure 2 shown, the baseline-based automatic edge computing network resource allocation method of the embodiment is used to solve the resource allocation problem in the edge computing network. The method aims to propose a baseline-based resource allocation strategy evaluation method in the scenario of data communication between wireless device terminals and edge servers. A multi-scale pyramid structure is used to design the network architecture for offloading decision resource allocation prediction. The network architecture can have a unified output for different input dimensions without modifying the network parameters, and can adapt to the problem of dynamic changes of nodes in the edge computing network. At the same time, multiple networks are used in parallel for prediction to avoid using complex optimization algorithms and achieve flexible, simple and efficient optimization effect. Specifically, the method includes the following steps:

[0053] Step 1, after the start of each time slot, the channel gains between each wireless device and each hybrid access point are obtained to form a channel set of the current time slot. The channel set denotes the channel gains between the 5 wireless devices and all hybrid access points, wherein denotes the channel gain of the i-th wireless device.

[0054] Step 2, input the channel set of the current time slot into the parallel discrete network branch to predict the offloading decision of each wireless device, and obtain an offloading decision set of the current time slot.

[0055] As Figure 3 shown, the discrete network branch of the embodiment includes a discrete network and a normalization layer, wherein the discrete network includes an input layer, a multi-scale pyramid network and an output layer. The channel set of the current time slot is input into the input layer of the discrete network, and after passing through the multi-scale pyramid network, a first matrix is output by the output layer. After the SoftMax normalization operation of the normalization layer, a probability matrix is obtained, and the offloading decision of each wireless device is output according to the probability distribution in the probability matrix.

[0056] Taking the embodiment as an example, the input layer of the discrete network is the channel set denotes the channel gains between the 5 wireless devices and all hybrid access points; the multi-scale pyramid network is connected after the input layer, and the output is the first matrix of ; the first matrix ​​SoftMax normalization operation is performed to obtain a probability matrix ; according to the probability distribution in the probability matrix , random offloading sampling is performed (for example, the probabilities of device offloading to three hybrid access points are , then there is a 0.1 probability of sampling to the hybrid access point 1, a 0.2 probability of sampling to the hybrid access point 2, and a 0.7 probability of sampling to the hybrid access point 3), to obtain an offloading decision set , wherein represents the offloading decision of the i-th wireless device, wherein represents that the i-th wireless device offloads data to the j-th hybrid access point, and vice versa represents no offloading. Step 3, input the channel set and the offloading decision set of the current time slot into a parallel first continuous network branch to predict the energy allocation ratio of each wireless device. The first continuous network branch of the embodiment includes a first continuous network and a normalization layer, wherein the first continuous network includes an input layer, a multi-scale pyramid network, and an output layer; the channel set and the offloading decision set of the current time slot are input into the input layer of the first continuous network as a two-channel matrix, and after passing through the multi-scale pyramid network, a second matrix is output by the output layer. After SoftMax normalization operation of the normalization layer, the energy allocation ratio of all wireless devices obtained by local calculation is obtained.

[0057] Step 4, input the channel set and the offloading decision set of the current time slot into a parallel second continuous network branch to predict the offloading time allocation ratio of each wireless device.

[0058] The second continuous network branch of the embodiment includes a second continuous network and a normalization layer, wherein the second continuous network includes an input layer, a multi-scale pyramid network, and an output layer; the channel set and the offloading decision set of the current time slot are input into the input layer of the second continuous network as a two-channel matrix, and after passing through the multi-scale pyramid network, a third matrix is output by the output layer. According to the offloading decision set of the current time slot, the normalization layer performs SoftMax normalization operation on the element values corresponding to the wireless devices offloaded to the same hybrid access point in the third matrix to obtain the offloading time allocation ratio of each wireless device.

[0059] The embodiment of the present application utilizes parallel continuous network design to obtain energy and time allocation, and the input layer of the continuous network is a two-channel matrix composed of the channel set

[0060] and the offloading decision set .

[0061] The first continuous network branch of the embodiment includes a first continuous network and a normalization layer, wherein the first continuous network includes an input layer, a multi-scale pyramid network, and an output layer; the channel set and the offloading decision set of the current time slot are input into the input layer of the first continuous network as a two-channel matrix, and after passing through the multi-scale pyramid network, a second matrix is output by the output layer. After SoftMax normalization operation of the normalization layer, the energy allocation ratio of all wireless devices obtained by local calculation is obtained. ​​; the input layer is connected to a multi-scale pyramid network; the first continuous network outputs a second matrix , the second matrix is subjected to a Sigmoid normalization operation to obtain a set of energy allocation ratios obtained by local calculation of all wireless devices ; the second continuous network outputs a third matrix , according to the set of offloading decisions , the corresponding values of the wireless devices offloaded to the same hybrid access point are subjected to a SoftMax normalization operation to obtain a set of offloading time allocation ratios , wherein represents the th hybrid access point, used for association with , specifically the hybrid access point corresponding to the th wireless device.

[0062] For example, there are 5 wireless devices, and the numbers of the hybrid access points to which the wireless devices are offloaded are , then it is indicated that the 1st and 5th wireless devices are offloaded to the same hybrid access point, and the 2nd and 4th wireless devices are offloaded to the same hybrid access point. When performing the SoftMax normalization operation, the data corresponding to the 1st and 5th wireless devices in the third matrix are subjected to a SoftMax normalization operation together, the data corresponding to the 2nd and 4th wireless devices are subjected to a SoftMax normalization operation together, and the data corresponding to the 3rd wireless device is subjected to a SoftMax normalization operation alone, thereby obtaining the offloading time allocation ratios.

[0063] As shown in Figure 4 , the multi-scale pyramid network used in the three parallel branches of the embodiment includes a convolution pyramid layer, a pooling pyramid layer, and a fully connected layer. The convolution pyramid layer includes dilated convolutions with different convolution kernel sizes, the pooling pyramid layer includes pooling pyramids, and each pooling pyramid performs downsampling operations on different pooling regions. After the input data passes through the convolution pyramid layer, it obtains feature maps, each feature map passes through a pooling pyramid, and finally feature maps are obtained. The feature maps are unfolded into one-dimensional vectors and spliced, and the spliced features pass through the fully connected layer to obtain the output data of the multi-scale pyramid network.

[0064] For example, the embodiment provides a multi-scale pyramid network structure design, which constructs 3 dilated convolutions with different convolution kernel sizes, and the convolution kernel sizes are , using 3 convolution kernels to perform convolution operation on the input data, and the sizes of the convolution kernels are , , , 3 feature maps are obtained , then for each feature map, down-sampling is performed on 2 different pooling regions (down-sampling of different pooling regions can be realized by setting different pooling kernel sizes and / or strides, and the pooling layer or the pooling convolution can be directly sampled), and finally 6 feature maps are obtained, all the feature maps are unfolded into one-dimensional vectors for splicing, followed by a fully connected layer and an output layer to output data.

[0065] Step 5, taking the parallel discrete network branch, the first continuous network branch and the second continuous network branch after the historical iteration training as the baseline, according to the channel set of the current time slot, outputting the historical offloading decision, energy allocation ratio and offloading time allocation ratio.

[0066] In order to break the defects of the static baseline in the prior art, in this embodiment, the parallel discrete network branch, the first continuous network branch and the second continuous network branch after each round of iteration training are stored in a cache queue, if the number of caches exceeds the size of the cache queue, the earliest cached element in the cache queue is removed according to the first-in first-out principle, and the first element (denoted as the parallel network model after the historical iteration training) in the cache queue is taken as the baseline to participate in training.

[0067] For example, if the queue size is set to 100 in this embodiment, if the number of cached networks exceeds 100, the earliest model is removed according to the first-in first-out principle of the queue; the number of model training rounds is denoted as , if , the baseline is set to 0, and the historical maximum calculation rate is set in the subsequent loss calculation; if , the first set of network parameters in the cache queue is taken to predict the output of the channel in the current state, and the baseline network output result , is the historical offloading decision set, is the historical energy allocation ratio set, is the historical offloading time allocation ratio set.

[0068] Step 6, according to the current time slot and the historical offloading decision, energy allocation ratio and offloading time allocation ratio, the maximum calculation rate is calculated respectively to form a discrete loss function to update the discrete network branch.

[0069] In this embodiment, the maximum calculation rate is provided as follows:

[0070]

[0071] Therefore, in the calculation of the discrete loss function, the maximum calculation rate of the current time slot is calculated according to the offloading decision, the energy allocation ratio and the offloading time allocation ratio of the current time slot, denoted as ; the maximum calculation rate of the history is calculated according to the historical offloading decision, the energy allocation ratio and the offloading time allocation ratio, denoted as ; and the discrete loss function is calculated based on the maximum calculation rate of the previous time slot and the maximum calculation rate of the history as follows:

[0072]

[0073] In the formula, is the discrete loss function, is the i-th column of the probability matrix, is the absolute value of the maximum calculation rate difference, is the decision parameter, if is greater than , then ; otherwise .

[0074] Step 7, the maximum calculation rate of the current time slot is calculated according to the offloading decision, the energy allocation ratio and the offloading time allocation ratio of the current time slot, and the first continuous network branch and the second continuous network branch are formed to update the continuous loss function, and the time slot ends, and the training is repeated until all rounds are completed.

[0075] In the calculation of the continuous loss function, the reciprocal of the logarithm of the maximum calculation rate of the current time slot is taken as the continuous loss function.

[0076] In the network training process, the entire architecture contains three parallel neural networks (one discrete network and two continuous networks), and the neural network training batch size is 1, that is, the channel data of the current time slot is obtained from the environment each time to perform neural network model training update, and it is recorded as one round of iteration, and a total of 4W rounds of training are performed. The AdamW optimizer (Adam with Weight Decay) is used for the three neural networks, the learning rate of the discrete network is 3e-5, the learning rate of the two continuous networks is 3e-6, and the training is repeated until all rounds are completed. After the training is completed, the three parallel neural networks can output the offloading decision, the energy allocation ratio and the offloading time allocation ratio of each wireless network based on the channel data of the current time slot.

[0077] ​This invention proposes a baseline-based multi-scale pyramid network architecture, utilizing a multi-network parallel deep reinforcement learning method to solve the resource allocation problem in edge computing networks powered by renewable energy. It also proposes a baseline-based optimization strategy evaluation method to assess the merits of the current network output strategy and guide the update of neural network parameters. This invention constructs a multi-scale pyramid network architecture that can adapt to data inputs of different dimensions, transforming them into a unified feature vector. In edge computing networks, if the number of nodes changes, the network model can continue to be used directly without redesigning the network structure. This invention employs multi-network parallel computation, utilizing discrete networks to obtain data offloading decisions and multiple continuous networks to obtain energy and time allocation ratios, achieving simplicity and efficiency. This avoids the problems of high optimization algorithm difficulty and time-consuming iterations required in traditional optimization methods. The resources to be optimized in this invention include the data offloading ratio and the time allocated to each terminal device in time-division multiple access.

[0078] For example, in an edge computing network containing 5 wireless devices and 3 hybrid access points, the corresponding channel matrix h = [[1.8e-3, 5.2e-5, 2.1e-6], [5.1e-5, 2.7e-5, 9.6e-5], [1.3e-6, 1.9e-5, 2.9e-5], [2.1e-5, 9.2e-4, 6.1e-6], [2.1e-5, 3.4e-5, 6.1e-6]], bandwidth... Hertz, the number of cycles required for the CPU of each wireless device to process 1 bit. Dynamic power consumption coefficient Additional data transfer during data offloading Gaussian noise Each time slot Seconds, the energy each device receives in each time slot kilowatts, then the optimized equipment unloading decision is: Energy distribution ratio Time allocation ratio .

[0079] It should be understood that, although Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2At least one of the steps in the above-mentioned embodiments can include a plurality of sub-steps or a plurality of stages, which are not necessarily executed at the same time, but can be executed at different times, and the order of execution of the sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or at least a part of the sub-steps or stages of other steps.

[0080] The technical features of the above-mentioned embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features in the above-mentioned embodiments are described, but it should be understood that any combination of the technical features is within the scope of the present disclosure as long as there is no contradiction.

[0081] The above-mentioned embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the application. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A baseline-based automated edge computing network resource allocation method, wherein the edge computing network has A hybrid access point and A wireless device, characterized in that... Includes the following steps: After the start of each time slot, the channel gain between each wireless device and each hybrid access point is obtained to form the channel set of the current time slot; The channel set of the current time slot is input into a parallel discrete network branch to predict the offloading decision of each wireless device, thus obtaining the offloading decision set of the current time slot. Input the channel set and offload decision set of the current time slot into the first continuous network branch in parallel to predict the energy allocation ratio of each wireless device; The channel set and offload decision set of the current time slot are input into the parallel second continuous network branch to predict the offload time allocation ratio of each wireless device; Take the first in history After each iteration of training, the parallel discrete network branch, the first continuous network branch, and the second continuous network branch serve as the baseline. Based on the channel set of the current time slot, the historical offloading decision, energy allocation ratio, and offloading time allocation ratio are output. Based on the current time slot and historical unloading decisions, energy allocation ratios, and unloading time allocation ratios, the maximum computation rate is calculated to form a discrete loss function to update the discrete network branches. Based on the unloading decision, energy allocation ratio, and unloading time allocation ratio of the current time slot, calculate the maximum computation rate of the current time slot, form a continuous loss function to update the first and second continuous network branches, the time slot ends, and training is repeated until all rounds.

2. The baseline-based automated edge computing network resource allocation method according to claim 1, characterized in that, The discrete network branch includes a discrete network and a normalization layer. The discrete network includes an input layer, a multi-scale pyramid network, and an output layer. The channel set of the current time slot is input into the input layer of the discrete network. After passing through the multi-scale pyramid network, the output layer outputs the first matrix. The first matrix is ​​normalized by the SoftMax normalization operation of the normalization layer to obtain the probability matrix. Random offloading sampling is performed according to the probability distribution in the probability matrix, and the offloading decision of each wireless device is output.

3. The baseline-based automated edge computing network resource allocation method according to claim 1, characterized in that, The first continuous network branch includes a first continuous network and a normalization layer. The first continuous network includes an input layer, a multi-scale pyramid network, and an output layer. The current time slot's channel set and offload decision set are combined into a two-channel matrix, which is input into the input layer of the first continuous network. After passing through a multi-scale pyramid network, the output layer outputs a second matrix. After the second matrix undergoes SoftMax normalization operation in the normalization layer, the energy allocation ratio obtained by the local calculation of all wireless devices is obtained.

4. The baseline-based automated edge computing network resource allocation method according to claim 1, characterized in that, The second continuous network branch includes a second continuous network and a normalization layer. The second continuous network includes an input layer, a multi-scale pyramid network, and an output layer. The channel set and offloading decision set of the current time slot are combined into a two-channel matrix and input into the input layer of the second continuous network. After passing through the multi-scale pyramid network, the output layer outputs a third matrix. Based on the offloading decision set of the current time slot, the normalization layer performs SoftMax normalization on the element values ​​of the wireless devices offloading to the same hybrid access point in the third matrix to obtain the offloading time allocation ratio of each wireless device.

5. The baseline-based automated edge computing network resource allocation method according to any one of claims 2-4, characterized in that, The multi-scale pyramid network includes convolutional pyramid layers, pooling pyramid layers, and fully connected layers; The convolutional pyramid layer contains The pooling pyramid layer contains dilated convolutions with different kernel sizes. A pooled pyramid, and each pooled pyramid performs... Downsampling operations in different pooling regions, the input data is passed through a convolutional pyramid layer to obtain... Each feature map is processed through a pooling pyramid to obtain the final result. Each feature map will Each feature map is unfolded into a one-dimensional vector and concatenated. The concatenated features are then passed through a fully connected layer to obtain the output data of the multi-scale pyramid network.

6. The baseline-based automated edge computing network resource allocation method according to claim 1, characterized in that, The baseline is obtained as follows: After each round of training iterations, the discrete network branches, the first continuous network branch, and the second continuous network branch are stored in a cache queue. If the number of cached branches exceeds the size of the cache queue, the earliest cached element is removed from the cache queue according to the first-in-first-out principle, and the first element in the cache queue is taken as the baseline for training.

7. The baseline-based automated edge computing network resource allocation method according to claim 2, characterized in that, The process of forming a discrete loss function to update the discrete network branches includes: Calculate the maximum computing rate of the current time slot based on the offloading decision, energy allocation ratio, and offloading time allocation ratio of the current time slot; Calculate the historical maximum computing rate based on historical unloading decisions, energy allocation ratios, and unloading time allocation ratios; The discrete loss function is calculated based on the maximum computation rate of the previous time slot and the historical maximum computation rate as follows: ； In the formula, For discrete loss functions, The probability matrix is ​​the nth List, The absolute value of the difference in maximum computational speed. This represents the maximum computation rate for the current time slot. This represents the highest computing speed in history. For decision parameters, if > ,but ;otherwise .

8. The baseline-based automated edge computing network resource allocation method according to claim 1, characterized in that, The continuous loss function is the reciprocal of the logarithm of the maximum computation rate of the current time slot.

Images