Video stream adaptive code rate control method based on online convex optimization and loop learning

Abstract

The invention discloses a video stream adaptive code rate control method based on online convex optimization and loop learning. The method comprises the following steps: step 1, building a hierarchical objective QoE evaluation model; and step 2, an ABR controller based on online convex optimization. And step 3, integrating a DASH video playing system of the ABR controller. And the HELL is developed according to the requirements of the 5G scene on the network application. In the macroscopic strategy, the HELL pursues more efficient network bandwidth use, and the video block quality is limited by lower single application delay. In the calculation method of the code rate, the HELL uses the network QoS at a plurality of request moments to smooth the change of the network, and the controller is adjusted through the feedback of the last decision. In the aspect of a code rate switching strategy, the HELL uses a more careful switching strategy. The method aims at maximizing the utilization rate of the streaming media service to the network bandwidth, reducing the jamming risk caused by each application, and ensuring stable video quality; compared with an existing ABR controller, the HELL achieves the optimal QoE in a scene of 83.33%.

Description

Technical Field

[0001] The present invention relates to a method for adaptive bit rate control of video streams, and in particular to a method for adaptive bit rate control of 5G media video streams based on online convex optimization and cyclic learning, belonging to the technical field of video streaming media service networks. Background Art

[0002] With the rapid development of 5G technology, network technology is experiencing revolutionary changes. This technological advancement has not only driven the growth of the number of users, but also created new opportunities for network services such as streaming media. The proportion of video-related service traffic in total network service traffic has maintained an upward trend in recent years.

[0003] The expansion of the online video market has also promoted the application of network technology in streaming video. Among them, cloud computing and cloud storage technology have been proven to effectively improve the quality of video services. On the video service side, video service providers (VSPs) use cloud computing to perform video streaming operations such as video transcoding in the cloud, which greatly reduces the overall computational workload of video processing and ensures the timeliness of video services. Cloud storage technology allows VSPs to host processed video resources in edge cloud servers to reduce the delay in distributing video resources to users. However, as video quality improves, delays and packet loss in video transmission will be very common. The high transmission speed, low latency and low packet loss of 5G networks enable the combination of 5G and cloud computing technology to give VSPs the ability to process high-quality video streams in real time in the cloud.

[0004] On the client side, a series of video service quality optimization technologies rely on cloud computing to dynamically monitor the user's service status and reasonably distribute video resources configured in the cloud server to improve the user's QoE. Among them, the ABR controller has been widely used in video services due to its lightweight deployment and good optimization effect. Most existing network video applications (such as YouTube, Netflix) use the dynamic adaptive streaming HTTP (DASH) video system with ABR. The application logic of ABR in the DASH system can be simply summarized as follows: calculate the maximum video quality allowed under the current service status through the service parameters returned by the user end, and apply for the corresponding quality video block in the next video data request to the video server. The existing ABR strategy is mainly optimized through throughput prediction or buffer level control, but these two methods have been proven to have defects, specifically:

[0005] 1) For controllers based on throughput prediction, network fluctuations during the service process will affect the performance of the predictor, thereby increasing the possibility of jamming.

[0006] 2) In highly volatile and ultra-dense networks (such as cellular networks), due to the rapid and drastic changes in instantaneous signal states, both throughput prediction and buffer level control methods frequently adjust the bit rate, resulting in the problem of fluctuating video quality.

[0007] At the same time, another obstacle to improving the quality of video services is to accurately interpret the user experience based on the service parameters available to mobile network operators and service providers, that is, to perceive the QoE of video streaming services through the service parameters mastered by operators. A commonly used benchmark in this regard is the mean opinion score (MOS) defined by the International Telecommunication Union-Telecommunication Standardization Sector (ITU-T). This standard divides the user experience into five different categories, each of which represents a different level of user perception of the quality of streaming services. Summary of the invention

[0008] In view of the technical problems mentioned in the above background technology, the present invention first analyzes the impact of network QoS on the terminal video playback process in video streaming services. Using a self-built video service test platform, the changes in the terminal video playback process (for example, changes in the number of freezes) are observed by changing the network QoS, and then the impact of network QoS on service quality is analyzed. Based on the above analysis, HELL is constructed. This is an ABR controller based on feedback learning, maximizing bandwidth utilization, and minimizing service delay. HELL is developed according to the requirements of 5G scenarios for network applications. In terms of macro strategy, HELL pursues more efficient use of network bandwidth and limits the quality of video blocks by lower single application delay. In terms of bit rate calculation method, HELL uses network QoS at multiple request times to smooth network changes, and adjusts the controller through feedback from the last decision. In terms of bit rate switching strategy, HELL uses a more cautious switching strategy. This aims to maximize the utilization of network bandwidth by streaming services, reduce the risk of freezes caused by each application, and ensure stable video quality.

[0009] The technical solution adopted by the present invention is a video stream adaptive bit rate control method based on online convex optimization and cyclic learning. The implementation process of the method is as follows:

[0010] Step 1: Build a hierarchical objective QoE evaluation model. The present invention designs a multi-level, multi-dimensional QoE evaluation model. The five parameters of the first level are service parameters, among which bandwidth, packet loss rate, delay and jitter are network QoS parameters, and bit rate is a video parameter. The second-level parameters describe the service process of video streaming media, which are divided into three categories: network capacity, network quality and video quality. The third-level parameters describe the objective service quality, including the playback quality and picture quality during the video service process. Finally, the user QoE is obtained by synthesizing the playback quality and picture quality.

[0011] Step 2: ABR controller based on online convex optimization. The controller used in the present invention is based on the QoE evaluation model parameters, and the control effect of the controller is evaluated by the evaluation model. The controller converts the bitrate selection problem into a probability selection problem, discretizes and convexifies it, converts the ABR control problem into an online convex optimization problem, and uses a convex optimization solution method to select the optimal bitrate.

[0012] Step 3: DASH video playback system with integrated ABR controller. This system integrates the evaluation model and controller implemented in the first two stages. The video bitrate that the system sends to the server is the optimal bitrate given by the controller in step 2. The original video resource is encoded in blocks with multiple qualities and saved on the cloud server. At the beginning of the service, the media presentation description MPD file describing the video block is first parsed from the server, and then the ABR controller controls the quality of the video block requested by the client to the server. The user's local player monitors the current service and returns the service parameters to the ABR controller online. After receiving the parameters, the ABR controller uses the ABR algorithm to evaluate the current service, calculates the video quality when requesting a video block from the cloud server next time, and applies it in the next request.

[0013] The change process of the player buffer during DASH streaming service is as follows: k-1 ~t k When the player plays n k-1 , and download n k At t k ~t k+1 When the network quality is degraded, the player cannot complete n before the buffer is exhausted. k+1 Download, then in t k+1 ~t k+2 When the video segment download is completed, the buffer level increases by one video segment length, and the video segment download and playback continue.

[0014] The impact of network QoS on the terminal video service process is analyzed. By controlling the changes in QoS parameters, how it affects the terminal video playback is analyzed. Since this part of the test does not involve the ABR controller, that is, the client requests video blocks from the server at a constant bit rate, this requires that the network bandwidth is not less than the video bit rate, that is, the video can be played smoothly without delay and packet loss in an ideal environment.

[0015] In order to more accurately analyze the impact of network bandwidth on video playback, relative bandwidth is used as a variable for testing. The calculation formula is as follows

[0016]

[0017] Where B is the network bandwidth and R is the video bit rate. Relative bandwidth describes the network's carrying capacity for video transmission. R Indicates the relative bandwidth when the bandwidth is B and the video bit rate is R.

[0018] Relative bandwidth: The model prioritizes the highest possible relative bandwidth under the same network bandwidth to achieve the player's robustness to stuttering. The reduction in the number and duration of stuttering is positively correlated with the improvement of QoE, so allocating a video bitrate that is more in line with the current network conditions based on relative bandwidth will significantly optimize the video service quality.

[0019] Packet loss rate: Experimental results show that as the packet loss rate increases, the duration of a single freeze also increases, while the buffer level continues to decrease. The results also show that the impact of increased packet loss rate on high-bitrate videos is more significant than that on low-bitrate videos. Therefore, in order to improve the quality of video service, the model quickly reduces the video bitrate when the network packet loss rate is high. This bitrate switching strategy can achieve stable video streaming transmission and improve the quality of video service.

[0020] Latency: Increased network latency causes buffer levels to drop, increasing the frequency and duration of freezes. The model evaluates latency and buffer levels, and makes bitrate decisions based on the relative state of the buffer and latency. This strategy minimizes the impact of network latency on video block streaming, improving the QoE of video services by improving the stability of video playback.

[0021] HELL high bandwidth utilization low latency control logic and feedback learning control method. HELL perceives service scenarios through network QoS and uses parameters such as latency and throughput for bitrate control. The HELL bitrate controller is the core part. HELL monitors the network status and player buffer level to make bitrate decisions, generates decision bitrates and inputs bitrate switching logic. The bitrate switching logic is responsible for generating the requested bitrate based on the decision bitrate and video playback status, reading the video bitrate ladder from the video segment requester, generating the serial number of the requested video bitrate and downloading it to the video buffer.

[0022] HELL introduces smoothing processing for network bandwidth, which improves the ABR controller's resistance to instantaneous changes in services by smoothing the network status at the past few discrete application moments.

[0023]

[0024] Where I all represents the total data volume of the video block, ni represents the i-th requested video block, and t represents the transmission duration of the current video block.

[0025] In DASH streaming services, the client's request for video resources is discrete in the time domain. Therefore, in a high-rate network environment, bandwidth calculation needs to integrate the bandwidth of the link at multiple times to smooth the impact of network fluctuations on controller decisions. The core control logic of HELL is to improve network bandwidth utilization, and its calculation formula is as follows:

[0026]

[0027] a is the video play flag. When the ni-th video block is transmitted and there is a video playing in the streaming service, a is 1, otherwise it is 0.

[0028] In DASH service, the buffer is consumed when downloading video segments, and the buffer is replenished with space for the length of the video segment after the download is completed. Therefore, the core formula of the minimum latency strategy used by the mode is:

[0029]

[0030] Cni represents the bandwidth of network transmission, and BLni represents the buffer quality when transmitting the nith data block. ni is the playback speed gain. When the video is played at multiple speeds (1+p ni ) indicates the playback speed of the video. Vni indicates the nth i The playback time of the video block, Lnibi represents the network delay generated when transmitting the ni-th video block. i represents the cumulative delay caused by downloading video segments in streaming services. i <0 holds true, then BL ni =0, playback will stop due to buffer underflow to download video resources, that is, a buffering event occurs.

[0031] In summary, HELL's macro control logic controls the bit rate increase in order to improve the utilization of the total network bandwidth by streaming media services, and at the same time uses a low-latency strategy to limit the bit rate to ensure that the service is within the carrying capacity of the network and the user end, thereby achieving high-quality streaming media services and sacrificing controllable buffer levels in exchange for high-quality video.

[0032] The ABR control process based on feedback learning is as follows:

[0033] Simplify (1)-(4) and organize the control logic of HELL as follows:

[0034]

[0035]

[0036] Where BL min and BL maxRepresents the minimum and maximum values ​​of the buffer level. min and R max Represents the maximum and minimum selectable video bitrates.

[0037] The control process of HELL is as follows. Formulas (9) and (10) give the function for controlling the video bit rate:

[0038]

[0039]

[0040] Formula (10) gives the value of η when R is too large. ni >1. Formula (11) gives the penalty for R ni For online ABR control, the uncertainty of network QoS between adjacent applications makes it possible to maintain BL in formula (6) ni ∈[0,BL max ], It becomes a very complex control problem. Taking advantage of the fact that the total service time in streaming media service is much longer than the single decision time of the controller, we believe that {R n1 ,R n2 ,…,R nN} is an infinite length sequence. Based on this, consider BL ni is bounded on average. That is, BL is allowed ni The upper and lower bounds are violated instantaneously, and Rn is calculated using formulas (8) and (9). i Penalize, so that BL satisfies (7) in a statistical sense.

[0041] The full rate ladder is used in the decision-making process of the ABR controller to calculate the requested bitrate. Therefore, let {R1, R2, … R K} is the code rate ladder, considering the probability simplex ω={ω1,...,ω k ,...,ω K},satisfy

[0042] Ω={ω∈R N :ω≥0∧||ω||1=1} (10)

[0043] where ω k Indicates decision Rn i =R k The probability of decision code rate Rn i,req is the inner product of the code rate ladder and ω. This operation convexifies formulas (9) and (10) and solves them using convex optimization theory. Adjust formulas (8) and (9) to formulas (11) and (12), and rearrange formulas (6) and (7) to formula (13);

[0044]

[0045]

[0046]

[0047] At this point, the rate control problem has been transformed into a minimization problem. Since (13) requires However, it is unknown in the decision-making process. Consider using the objective function and constraint function about The first-order Taylor expansion estimates the two applications n i The predicted value at time , and the logarithmic penalty function is introduced The control bit rate is applied to satisfy (7). Consider:

[0048]

[0049] Where A is the rate decision control coefficient, and κ is the penalty intensity control coefficient.

[0050] For the minimization problem, the constraints are satisfied. So it can be converted to Minimization problem. Use gradient descent method to solve, that is, calculate

[0051]

[0052] Solving (14) yields ω t The update formula is:

[0053]

[0054] Proj[·] is a Euclidean transformation, so that satisfies the probability simplex.

[0055] Based on HELL's macro control logic and learning method, In the calculation of and is the impact of the last adjustment on bandwidth utilization, and The introduction of feedback enables HELL to gradually achieve continuous control of the service. That is, after the service starts, HELL learns the feedback generated by past control in multiple video applications, gradually improves the accuracy of perception of service status, and makes adjustments according to the control logic to achieve optimal control of the bit rate.

[0056] Compared with the existing technology, this method first analyzes the impact of network QoS on terminal video playback and finds that compared with the changes in packet loss rate and relative bandwidth, the impact of delay on the video streaming service process is the most obvious. After that, HELL, an ABR controller based on feedback learning, is proposed. HELL perceives the quality of streaming service by monitoring network QoS. In the ABR control logic, HELL achieves bit rate control by giving priority to improving bandwidth utilization through controllable compression buffer levels. In the ABR control method, HELL calculates the optimal bit rate for each video application based on maximizing bandwidth utilization and minimizing service delay each time a video segment is requested, and the impact of perceived control on streaming playback is achieved based on feedback learning, thereby achieving learning optimization. In addition, HELL is developed entirely based on an open source platform and tested using real 5G streaming service network data, making it more widely used in the growing 5G video streaming business. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Diagram of the objective QoE evaluation model in video streaming services.

[0058] Figure 2 This is a diagram of a DASH video playback system equipped with an ABR controller.

[0059] Figure 3 A graph showing the changes in buffer levels in a video streaming service.

[0060] Figure 4 Diagram of the QoS-terminal video service test system.

[0061] Figure 5 This is the structural diagram of HELL in the DASH system.

[0062] Figure 6 Figure 2. Network traces used in the ABR controller test.

[0063] Figure 7 The test results of the ABR controller are as follows. (a) Average bit rate; (b) Average bandwidth utilization; (c) Buffering duration; (d) Average buffer level; (e) QoE;

[0064] Figure 8 This is the average buffer level distribution map.

[0065] Fig. 9 This is the code rate distribution diagram.

[0066] Fig.10 This is the QoE distribution diagram. DETAILED DESCRIPTION

[0067] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0068] In order to be closer to the real service scenario, a DASH video service system is designed for HELL, and simulation tests are carried out with various ABR controllers in a real 5G network data set. The experimental results show that HELL can effectively improve the video quality and service QoE by controllably reducing the buffer level in the experimental network scenario compared with the existing ABR controller. The QoE of video streaming services is improved by adaptive bit rate control, so the QoE evaluation problem is first discussed. The main methods for QoE evaluation are subjective and objective evaluation. Subjective methods, such as those described in the work of the International Telecommunication Union (ITU), involve measuring QoE in a multimedia environment in combination with user perception. This type of method focuses on the user's subjective perception of service quality, that is, evaluating the service quality based on the user's subjective feelings about the service. The results of this type of evaluation are affected by many factors such as personal preferences, emotions, and environmental factors, so they have certain uncertainty and variability. Objective evaluation evaluates service quality by quantifying the service process, that is, quantifying service quality through service parameters such as network performance (e.g., network bandwidth), picture presentation quality (e.g., video bit rate), and video playback process (e.g., freeze duration) during the service process. This type of evaluation result is more stable and repeatable, but lacks consideration of users' subjective feelings. Figure 1 A hierarchical objective QoE evaluation model is given. The five parameters of the first level are service parameters, among which bandwidth, packet loss rate, delay and jitter are network QoS parameters, and bit rate is a video parameter. The second level parameters describe the service process of video streaming, which is divided into three categories: network capacity, network quality and video quality. The third level parameters describe the objective service quality, including the playback quality during the video service process (for example, if freezes occur frequently, the playback quality will decrease) and picture quality (for example, if a flower screen occurs, the picture quality will decrease). Finally, the user QoE is obtained by synthesizing the playback quality and picture quality.

[0069] The issue of QoE evaluation in video streaming services has been extensively studied, and the service parameters involved are summarized as follows:

[0070] Network QoS: This type of parameters mainly includes bandwidth, packet loss rate, jitter, latency, and end-to-end round-trip delay. Network QoS mainly describes the network conditions in video streaming services. Network QoS can be used to analyze network quality and then perceive the possible impact of changes in network quality on terminal video playback.

[0071] Picture presentation quality: This type of parameters mainly includes video bit rate and video quality assessment (VQA) algorithm. Compared with video bit rate, VQA has a better ability to perceive user feelings. The more commonly used ones are SSIM proposed by Wang et al. and VMAF proposed by Li et al. This type of parameter is intended to evaluate the user's subjective feelings about the video picture. Since the picture delay quality depends on the transcoding size and the type of video streaming service (such as on-demand, live broadcast, real-time video streaming) and the impact of the network on video transmission, the single use of picture presentation quality is often unable to accurately and comprehensively evaluate the service QoE.

[0072] Video playback quality: This type of parameter mainly focuses on describing the jamming events during the service process, because jamming events are one of the important factors for accurately predicting the overall QoE. In addition, the buffer level during the service process is added, because the root cause of jamming is buffer underflow. The buffer level during the service process can be used to analyze the ABR controller's ability to balance video quality and buffer level and the potential jamming risk during the service process. Table 1 lists the parameters used in these studies to evaluate QoE and marks the parameters that belong to network QoS.

[0073] Table 1. Comparison of QoE evaluation model parameters

[0074]

[0075] Among all the parameters that affect QoE, network QoS parameters have been proven to have a greater impact on QoE. The relationship between network-based QoS parameters and user-perceived QoE can provide an important reference for QoE management of video streaming services. In particular, for network providers who control network resource planning and configuration mechanisms, the mapping relationship between QoS and QoE can achieve controllable service quality improvement. Researchers have conducted extensive research on the mapping relationship between QoS and QoE. Vega et al. focused on the potential relationship between QoS and the picture quality of different types of videos. By analyzing the picture quality of 2D and 3D videos in different network environments (including different packet loss rates, delays, and bandwidths), it was found that there were differences in resilience between different types of 2D video streaming services compared with 3D videos. Miranda et al. used a network probe based on the Internet Control Message Protocol (ICMP) to detect network quality for video on demand (VoD) services, and based on this, perceived the QoE of VoD services. The advantage of this evaluation method is that it does not require the detection of client devices, modification of existing network elements, deployment of monitoring tools at multiple network points, or in-depth inspection of video traffic, which greatly reduces the perceived cost of service QoE. In response to the perception of QoE in the 5G environment, Nightingale et al. used network bandwidth and video bit rate to measure the current network level's ability to carry video streams, and fitted a QoE calculation model through a data set.

[0076] In summary, perceiving QoE through network QoS is an important method to quantify the quality of video streaming service systems. Therefore, streaming algorithms in real video streaming services widely use network QoS parameters to guide the dynamic control of video service parameters (such as video bit rate) to achieve improved video service quality.

[0077] ABR control in video streaming services guides video service control through QoS to improve service QoE. The following will introduce the specific implementation of this control method. Figure 2 A DASH video playback system equipped with an ABR controller is demonstrated. The original video resource is encoded into multiple qualities and stored on the cloud server. At the beginning of the service, the user first parses the media presentation description (MPD) file describing the video chunks from the server, and then the ABR controller controls the quality of the video chunks requested by the client to the server. The user's local player monitors the current service and returns the service parameters to the ABR controller. After receiving the parameters, the ABR controller uses the ABR algorithm to evaluate the current service, calculates the video quality when requesting the video chunk from the cloud server next time, and applies it in the next request.

[0078] Figure 3This shows the changes in the player buffer during DASH streaming service. k-1 ~t k When the player plays n k-1 , and download n k At t k ~t k+1 When the network quality is degraded, the player cannot complete n before the buffer is exhausted. k+1 Download, then in t k+1 ~t k+2 When the video segment download is completed, the buffer level increases by one video segment length, and the video segment download and playback continue.

[0079] Existing ABR controllers can be mainly divided into heuristic and learning-based according to their bitrate control methods. Unlike heuristic ABR controllers, learning-based ABR controllers continuously update control parameters by learning service status during the service process, that is, the controller self-adjusts in real time according to the service to track changes in the service environment. Compared with learning-based controllers, the behavior of heuristic controllers is usually more predictable and stable because they do not rely on external data. However, their control performance in unknown and changing network environments is weaker than that of learning-based controllers. This is because learning-based controllers do not rely on fixed heuristic methods, but try to learn to modify control parameters from the system environment to adapt to system dynamics.

[0080] The impact of network QoS on the terminal video service process is analyzed. By controlling the changes in QoS parameters, how it affects the playback of terminal videos is analyzed. Since this part of the test does not involve the ABR controller, that is, the client requests video blocks from the server at a constant bit rate, this requires that the network bandwidth is not less than the video bit rate, that is, the video can be played smoothly under ideal conditions (no delay and no packet loss). Table 2 gives a list of symbols used.

[0081] Table 2. List of symbols

[0082] N Total video chunks for video streaming services n Video block i, i∈[1,N] <![CDATA[R ni ]]> <![CDATA[The bit rate R of the nth i video block mi ∈ [R min , R max > <![CDATA[B ni ]]> <![CDATA[Network bandwidth when the nth i video block]]> <![CDATA[C ni ]]> <![CDATA[Real-time network throughput at the nth i video block]]> <![CDATA[BI ni ]]> <![CDATA[Buffer level at the nth i video block]]> <![CDATA[T ni ]]> <![CDATA[Cumulative freeze duration at the nth i video block]]> <![CDATA[I ni ]]> <![CDATA[Network latency when the nth i video block]]> <![CDATA[t ni ]]> <![CDATA[Download duration of the nth i video block]]> <![CDATA[BW R ]]> Relative bandwidth <![CDATA[Vn i ]]> <![CDATA[n i Length (seconds)]]> a Video playback flag α Bit rate compensation factor β Stuttering compensation factor γ Delay compensation coefficient σ Service process control coefficient μ Caton risk control coefficient λ Bit rate change control coefficient <![CDATA[I ni ]]> Total user link traffic <![CDATA[η t ]]> Streaming service bandwidth utilization. <![CDATA[p t ]]> Playback speed change factor A Bit rate decision control coefficient κ Penalty intensity control coefficient. K Bandwidth smoothing factor

[0083] 1. Quantitative analysis system

[0084] Figure 4 The architecture of the system for testing QoS-endpoint video services. DASH videos are pre-encoded and hosted on an Apache web server. Mahimahi is used to simulate network conditions. The client video player is a Chromium browser that supports H.264 and HEVC playback. The client and server run on the same computer. During video playback, the client monitors the service process and generates a log file after the service ends.

[0085] 2. Parameter settings

[0086] In the selection of test parameters, in order to more accurately analyze the impact of network bandwidth on video playback, relative bandwidth is used as a variable for testing. The calculation formula is as follows:

[0087]

[0088] Relative bandwidth describes the network's ability to carry video transmission. R The larger the value, the stronger the current network's carrying capacity for the bitrate video, and vice versa. Set the relative bandwidth change to ±10%. In the delay analysis, according to the delay range of the 5G streaming service (30-75ms), set the delay change range to 30-70ms. For packet loss rate, set the PLR ​​range to 1-10%. When performing delay and packet loss rate tests, select BW for network bandwidth. R =1. The video material uses Big-Buck-Bunny and Spring4K used in dash.js player, and is tested using three bitrate configurations: high, medium, and low. The bitrate differences between adjacent configurations are 5018Kbps and 4751Kbps. Table 3 gives the data of the video samples. The maximum buffer level is set to 30s.

[0089] Table 3. Test video samples

[0090] Resolution Bitrate(Kbps) Duration 1080pCartoon 1920x1080 9914 0：10：34 4Kcartoon 3840x2160 14932 0：10：34 4KDocumentary 3840X2160 19683 0：02：45

[0091] 3. QoS Analysis

[0092] Tables 4 to 6 show the impact of QoS changes on the video service process. The total buffering time (s), single buffering time (s) and average buffer level (s) of the video service are selected for analysis.

[0093] Relative bandwidth: As the relative bandwidth increases, the total service freeze duration shows a downward trend, while the buffer level increases significantly. This shows that the increase in bandwidth reduces the possibility of buffer underflow, resulting in a decrease in the duration and possibility of freezes. It is worth noting that when the relative bandwidth is not less than 1, that is, the network carrying capacity is not less than the maximum bandwidth requirement of the video, freezes still occur, and the average buffer level also reaches the maximum value. This shows that bandwidth configuration cannot be converted into corresponding throughput, that is, the conversion relationship between bandwidth and effective throughput is also affected by conditions such as video bit rate and video slice size. Comparing the bit rate, as the relative bandwidth increases from 90% to 110%, the total freeze duration of medium bit rate and high bit rate videos decreases by 6.29s and 7.11s respectively, while the total freeze duration of low bit rate videos does not change significantly. This is because the basic bandwidth requirements of medium bit rate and high bit rate videos are high, so the increase in relative bandwidth brings a significant increase in bandwidth, so it has a significant effect on reducing the total freeze duration. However, the basic bandwidth of low bit rate videos is low, and the change in relative bandwidth cannot significantly optimize the bandwidth, so the optimization effect on the total freeze duration is poor.

[0094] Packet loss rate: When the packet loss rate increases from 0% to 4%, the total jamming time is relative to BW R =1 has increased significantly, which is consistent with the result that the duration of the freeze is highly correlated with the change in packet loss rate when PLR changes initially. For low-bitrate and medium-bitrate videos, when the packet loss rate continues to deteriorate to 10%, the resulting freeze duration is not significantly worse than when it was 4%, while the high-bitrate video is significantly worse. This is because high bitrates have higher data density, so the impact of packet loss rate deterioration on high-bitrate videos is more serious, which makes retransmission and buffering take longer. As the packet loss rate increases, the duration of a single freeze also increases, while the buffer level continues to decrease. This is consistent with the impact of network environment deterioration on video services.

[0095] Latency: When latency increases, three parameters of the service deteriorate significantly, namely, total latency and single latency increase, and buffer level decreases. Compared with the relative bandwidth and packet loss rate deterioration, the deterioration of latency has the most obvious impact on the service process. This is because video playback usually relies on "streaming", where data is sent continuously and played gradually. If the network latency increases, the speed at which data packets reach the player will slow down, so the terminal loads slower, causing the buffer level to decrease, which in turn leads to an increase in the duration of the freeze.

[0096] Table 4 Total delay of video playback

[0097]

[0098] Table 5 Average single delay

[0099]

[0100] Table 6 Average buffer levels

[0101]

[0102] In summary, the impact of latency on the video service process is significantly higher than the other two QoS. For PLR, transmission packet loss is usually an occasional event and can be repaired by the retransmission mechanism. High latency is a persistent problem that affects the transmission time of each data packet, thereby continuously affecting the entire video segment transmission process. As for relative bandwidth, in some video applications, such as video conferencing, online games, or real-time data transmission, the low or high latency directly affects the user experience. Even if the bandwidth is sufficient and the packet loss rate is low, high latency can still cause significant performance issues, such as asynchronous conversations and delayed game operations. Therefore, in video applications that require fast response or real-time interaction, the impact of latency on the service process is often more significant. This is because latency directly affects the transmission speed and synchronization of data, which has a more direct impact on the user experience.

[0103] HELL high bandwidth utilization low latency control logic and feedback learning control method. HELL perceives service scenarios through network QoS and uses parameters such as latency and throughput for bit rate control. Figure 5 The basic structure of HELL in DASH service is given. The HELL bitrate controller is the core part. HELL monitors the network status and the player buffer level to make bitrate decisions, generates the decision bitrate and inputs the bitrate switching logic. The bitrate switching logic is responsible for generating the requested bitrate based on the decision bitrate and video playback status, reading the video bitrate ladder from the video segment requester, generating the sequence number of the requested video bitrate and downloading it to the video buffer.

[0104] 5G network scenarios require network applications to perform well in high-throughput and mobile network environments and support use in ultra-dense networks (UDN). Since 5G is deployed ultra-densely to provide higher data transmission speeds, lower latency, and better connection stability, mobile devices may frequently switch connections to different base stations in ultra-dense networks, which will cause instantaneous and rapid changes in network status. The existing ABR controller lacks smoothing processing for network status and has the defect of frequent bit rate adjustments in high-change rate network environments. Therefore, HELL introduces smoothing processing for network bandwidth, which improves the ABR controller's resistance to instantaneous changes in services by smoothing the network status of the past few discrete application moments. Consider

[0105]

[0106] (3) The network bandwidth after smoothing calculation is given. According to the previous analysis, the impact of latency on QoE is more obvious, so HELL introduces the consideration of network latency to fully perceive the network service status. Compared with using the throughput at each moment for calculation, this has higher resistance to network fluctuations.

[0107] In DASH streaming services, the transmittable size of the video stream is limited by the network bandwidth utilization. At the same time, 5G network applications need to have high working efficiency to be deployed on the mobile edge computing (MEC) platform. This requires it to have high efficiency. Therefore, controlling the video bit rate according to the network bandwidth during the network service to improve the bandwidth utilization of the streaming service is a feasible strategy for ABR control. By improving the bandwidth utilization, the utilization of network resources can be maximized, thereby achieving better control effects in resource-constrained environments. Al-Jawad et.al proposed a streaming service controller based on different link bandwidth utilization. The core logic of the controller is:

[0108]

[0109] where d l,f , is the selection flag of media stream f for link l, a f , is the download video bit rate, C l is the link bandwidth. Formula (2) maximizes the occupancy rate of different link bandwidths of the streaming media service at the request time, thereby improving the total bandwidth utilization of the streaming media service and thus improving the service quality.

[0110] In DASH streaming services, the client's request for video resources is discrete in the time domain. Therefore, in a high-rate network environment, bandwidth calculation needs to integrate the bandwidth of the link at multiple times to smooth the impact of network fluctuations on controller decisions.

[0111] The core control formula of HELL is as follows:

[0112]

[0113] When there is a video playing in the streaming service at time t, a is 1, otherwise it is 0.

[0114] When the video bit rate increases, the bandwidth utilization of streaming media services increases. However, the increase in video bit rate also increases the download time, which may cause buffer underflow. Therefore, the quality of video service is controlled by service delay. Specifically, when η ni >1, this situation is permissible in a short period of time. However, due to network fluctuations, η ni >1 means that the current streaming service traffic exceeds the network's carrying capacity, so the video quality is unreasonable. ni<1 means that the current network is capable of supporting video services at this bit rate, the video quality is reasonable, and the video bit rate can be increased to increase bandwidth utilization. When increasing the requested bit rate, the bit rate increase is limited according to the minimum delay strategy so that the bit rate increase is not too large and exceeds the link carrying capacity.

[0115] The common control strategy of existing learning-based ABR algorithms is the minimum delay strategy, which ensures that the buffer will not be exhausted by minimizing the delay of video services, thereby reducing the occurrence of buffering events. The results of QoS tests show that the service process of video streaming services at the terminal is most significantly affected by delays, so the minimum delay strategy is introduced to reduce the possibility of buffer underflow and thus cause freezes. The core formula of this strategy is

[0116]

[0117]

[0118] (1) represents the change in the buffer level before and after a video chunk request. In DASH services, the buffer is consumed when a video segment is downloaded, and the buffer is replenished with space equal to the length of the video segment after the download is completed. i represents the cumulative delay caused by downloading video segments in streaming services. i <0 holds true, then BL ni =0, playback will stop due to buffer underflow to download video resources, that is, a buffering event occurs.

[0119] In summary, HELL's macro control logic is to improve the utilization of the total network bandwidth by streaming media services to control the bit rate increase, and at the same time use low-latency strategies to limit the bit rate to ensure that the service is within the carrying capacity of the network and the user end, and to achieve high-quality streaming media services. This logic of maximizing bandwidth utilization will lead to compression of the buffer level, but the low-latency strategy can limit the compression and make it controllable. This leads to the core logic of HELL: high-quality video in exchange for a controllable sacrifice of the buffer level.

[0120] ABR control based on feedback learning

[0121] Simplify (1)-(4) and organize the control logic of HELL as follows:

[0122]

[0123]

[0124] The control process of HELL is as follows. Formulas (10) and (11) give the control of R t Functions for controlling

[0125]

[0126]

[0127] Formula (10) gives the value when R t Too large leads to η ni >1. Formula (11) gives the penalty for R ni For online ABR control, the uncertainty of network QoS (such as throughput, bandwidth) between adjacent applications makes it necessary to maintain the BL in formula (6) ni ∈[0,BL max ], It becomes a very complex control problem. Taking advantage of the fact that the total service time in streaming media service is much longer than the single decision time of the controller, we believe that {R n1 ,R n2 ,…,R nN} is an infinite length sequence. Based on this, consider BL ni is bounded on average. That is, BL is allowed ni The upper and lower bounds are violated instantaneously, and Rn is calculated using formulas (7) and (8). i Penalize, so that BL satisfies (6) in a statistical sense.

[0128] The full rate ladder is used in the decision-making process of the ABR controller to calculate the requested bitrate. Therefore, let {R1, R2, … R K} is the code rate ladder, considering the probability simplex ω={ω1,...,ω k ,...,ω K}, which satisfies

[0129] Ω={ω∈R N :ω≥0∧||ω||1=1} (29)

[0130] where ω k Indicates decision Rn i =R k The probability of decision code rate Rn i,req is the inner product of the code rate ladder and ω. This operation convexifies formulas (9) and (10) and solves them using convex optimization theory. Adjust formulas (9) and (10) to formulas (12) and (13), and rearrange formulas (5) and (6) to formula (14);

[0131]

[0132]

[0133]

[0134] Since (14) requires However, it is unknown in the decision-making process. Consider using the objective function and constraint function about The first-order Taylor expansion estimates the two applications n i The predicted value at time , and the penalty function is introduced The control bit rate is applied to satisfy (9). Consider

[0135]

[0136]

[0137] Where A is the rate decision control coefficient, and κ is the penalty intensity control coefficient.

[0138] Since (12) is a minimization problem and the constraints satisfy So it can be converted to The minimization problem of . Use the gradient descent method to solve (13), that is, calculate

[0139]

[0140] Solving (15) yields ω t The update formula

[0141]

[0142] Proj[·] is a Euclidean transformation, so that satisfies the probability simplex.

[0143] Based on HELL's macro control logic and learning method, In the calculation of and is the impact of the last adjustment on bandwidth utilization, and The introduction of feedback enables HELL to gradually achieve continuous control of the service. That is, after the service starts, HELL learns the feedback generated by past control in multiple video applications, gradually improves the accuracy of perception of service status, and makes adjustments according to the control logic to achieve optimal control of the bit rate.

[0144] The application process of HELL in streaming media service is as follows:

[0145] 1. HELL continuously monitors the client service status and network QoS to perceive changes in video streaming services.

[0146] 2. The root loop learns the feedback generated by the last control on the service, and calculates the code rate step probability of the current application according to formula (19).

[0147] 3. Calculate the current optimal application bitrate based on the bitrate step probability and the server bitrate step, and submit an application to the server.

[0148] Next, performance tests are conducted on HELL and other ABR controllers to compare the performance of HELL.

[0149] Example

[0150] Experimental setup

[0151] Network data set: Network data is randomly cut and selected from Channel 5G to construct a network data set for testing different ABR controllers. The length of each data is 40s. All these data are collected from real data, covering network data of different streaming service operators (Netflix and Amazon Prime), different scenarios (mobile services and fixed services) and different video service types (animation, movies and dramas). The statistical characteristics of the data set are shown in Table 7. Considering that the low-bandwidth network environment is more challenging for the ABR controller, because the ABR controller is usually required to make a trade-off between the buffer level and the video quality in a low-bandwidth environment, Trace3 and Trace6 are used as low-bandwidth test environments. Figure 5 The changes of all network data over time are given.

[0152] Test platform: The test platform built mainly includes (i) a network environment simulator used to simulate network scenarios based on network data sets. The simulator is implemented based on Mahimahi. (ii) a DASH video playback system that supports custom ABR controllers. The system is developed based on dash.js and deployed in the Chrome browser. The test video material is "bbb_30fps" used in the dash.js reference client. The video block is 5s long and has 10 bitrate levels (254Kbps-14931Kbps). All test resources are hosted in the Apache web server to cooperate with the network environment simulator to perform tests in simulated network scenarios.

[0153] ABR controller: Throughput-based TH, buffer-based BOLA, latency-minimized L2A, and LoL+ in dash.js were selected for performance testing together with HELL.

[0154] Quantitative QoE evaluation model: Evaluating the performance of ABR controllers by quantifying QoE is a common method in ABR control research. In the measurement of model performance, the QoE evaluation model is used. Since the QoE in DASH streaming services is mainly affected by video quality, latency, and freezes, the QoE calculation formula used is the following formula to quantify the average QoE of video streaming services:

[0155]

[0156] The parameters are listed in Table 1.

[0157] Different from other models, the quantitative evaluation of buffer level is introduced in QoE calculation. Since buffer underflow is the direct cause of jamming, by introducing the quantification of buffer level, the ABR controller's ability to balance service quality and buffer level and the possibility of jamming in the service process can be evaluated.

[0158] Summary: The test environment is Ubuntu 22.04.3LTS, the test processor is Intel 11800H, and the Chrome browser version is 119.0.6046.159. The video client and server run on the same device, and the end-to-end delay is set to 30ms. The bit rate decision control coefficient A is set to 4 0.99 , the penalty intensity control coefficient is set to 1. A single test lasts two minutes. After the test, the test system summarizes the service status data and generates a test log.

[0159] Table 7. Characteristics of network traces

[0160]

[0161] Test results and analysis

[0162] 1. Overall analysis

[0163] Figure 5 The test results of the ABR controller are given, including average bit rate, average buffer level, total buffer time, average bandwidth utilization, QoE and QoE improvement level. Figure 5 The results in (a) show that, in all six network scenarios, HELL achieved the best QoE in five of them and achieved the best control effect. Figure 5 .(b) and Figure 5 The results in (c) prove that HELL can effectively improve the video bit rate and thus improve the service quality by controllably reducing the buffer level. Figure 5 (d) shows the total buffering time of each service scenario. The buffering time of HELL ranges from 2s to 12.37s, accounting for 1.67% to 10.31% of the total service time. In the scenario with the longest buffering time, HELL achieves a QoE improvement of 51.83% to 78.18% through controllable compression buffer level, realizing the original design intention of HELL.

[0164] 2. Comparative analysis of ABR controllers

[0165] Combined with the logic of the ABR controller selected in the experimental design, HELL is compared with the other five controllers to show HELL's ability to control the service process and improve QoE. The results of this part are shown in Figure 5-Figure 8 and Table 5.

[0166] Figure 5 (a) shows the average video bitrate of the ABR controller in each scenario, and HELL achieves the highest average bitrate in five of them. Table 5 shows that HELL achieves the highest average bitrate of 6.08Mbps in all scenarios, proving the superiority of HELL in improving bitrate. Figure 7 The results show that HELL's bitrate is concentrated at high bitrates, which is significantly different from the other controllers, that is, HELL increases the bitrate by compressing the buffer level.

[0167] In terms of buffer control, TH achieves the highest buffer level among all ABR controllers, with an average buffer level of 17.81s, significantly higher than HELL's 4.98s. HELL has the lowest average buffer level in all scenarios, and this feature is also Figure 6 This is reflected in the service buffer level distribution diagram. 90% of HELL's buffer levels are below 9.02s, while the values ​​of other controllers are around 30s. As for the duration of freezes, LoL+ has the lowest average freeze time of 0.99s. The average freeze time of HELL is 5.05s, accounting for 4.21% of the total service time. The sacrifice of buffer level has been exchanged for a 30% to 40% improvement in QoE. Fig.10 It proves that HELL's QoE is concentrated in high-level areas compared to the other four ABR algorithms, so this sacrifice is reasonable.

[0168] Combined with the average bandwidth utilization of all controllers in each scenario, this is the core focus of HELL. The results show that HELL has achieved the best bandwidth utilization in all five scenarios. The average bandwidth utilization given in Table 8 shows that HEL has improved by more than 20% in bandwidth utilization compared with other controllers, which shows that HELL achieves the expected design purpose.

[0169] Table 8. Average test results

[0170] LoL+ Bola HELL L2A TH Average bitrate(Mbps) 3.65 3.78 6.08 3.22 3.60 Average bandwidth efficiency(%) 3220 3550 55.38 3132 3188 Average buffer level(s) 1429 1712 498 14.73 17.81 Average rebuffering duration(s) 0.99 443 505 235 221 QoE 2663 2813 60.65 23.22 2629 QoE improvement (%) 3674 3805 0 44.25 3334

[0171] In summary, HELL has effectively improved the video quality by improving bandwidth utilization to compress the buffer level and then improve the control logic of bit rate. At the same time, HELL limits the compression of the buffer based on the low-latency restriction logic, so that the video service can still be relatively smooth. Based on the above logic and feedback learning control, HELL has achieved the best QoE and the highest network bandwidth utilization in 83.33% of the scenarios, proving that HELL has achieved the design purpose of achieving high-quality streaming media services by improving network bandwidth utilization.

Claims

1. A video stream adaptive bitrate control method based on online convex optimization and cyclic learning, characterized in that: The implementation process of this method is as follows: Step 1: Build a hierarchical objective user experience evaluation model, namely the QoE evaluation model; design a multi-level, multi-dimensional QoE evaluation model; the five parameters of the first level are service parameters, among which bandwidth, packet loss rate, delay and jitter are network QoS parameters, and bit rate is a video parameter; the second-level parameters describe the service process of video streaming, which is divided into three categories: network capacity, network quality and video quality; the third-level parameters describe the objective service quality, including the playback quality and picture quality during the video service process; finally, the current user QoE is obtained by synthesizing the playback quality and picture quality; Step 2: ABR controller based on online convex optimization; the adaptive video bitrate controller used is based on the QoE evaluation model parameters built in step 1, and the control effect of the controller is also evaluated by the hierarchical objective user experience evaluation model built in step 1; the bitrate controller converts the bitrate selection problem into a probabilistic selection problem, discretizes and convexifies it, realizes the conversion of the ABR control problem into an online convex optimization problem, and uses the convex optimization solution method to output the selected optimal video bitrate; Step 3: DASH video playback system with integrated ABR controller; the video playback system integrates the evaluation model implemented in the first two stages with the ABR controller, and the video bitrate requested from the video server is the optimal bitrate given by the ABR controller based on online convex optimization in step 2; the original video resource is encoded into multiple qualities and saved on the cloud server; at the beginning of the service, the media presentation description MPD file describing the video block is first parsed from the video server, and the ABR controller controls the quality of the video block requested by the client to the server; the user's local player monitors the current service and returns the video service parameters to the ABR controller online, including network bandwidth, delay and current video bitrate; after receiving the above parameters, the ABR controller uses the ABR algorithm to evaluate the current service, calculates the video bitrate when requesting the video block from the video server next time, and applies it in the next request.

2. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 1 is characterized in that: The change process of the player buffer during DASH streaming service is as follows: k-1 ~t k When the player plays n k-1 , and download n k ; at t k ~t k+1 When the network quality is degraded, the player cannot complete n before the buffer is exhausted. k+1 Download, then in t k+1 ~t k+2 A freeze occurs; when the video segment download is completed, the buffer level increases by the length of the video segment, and the video segment download and playback continue.

3. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 1 is characterized in that: Using relative bandwidth as a variable for testing, the calculation formula is as follows: Where B is the network bandwidth and R is the video bit rate. The relative bandwidth describes the network's carrying capacity for video transmission. BW R Indicates the relative bandwidth when the bandwidth is B and the video bit rate is R; Relative bandwidth: The model prioritizes high relative bandwidth under the same network bandwidth to achieve the player's robustness to stuttering. The reduction in the number and duration of stuttering is positively correlated with the improvement of QoE. Therefore, allocating a video bitrate that is more in line with the current network conditions based on relative bandwidth will significantly optimize the video service quality. Packet loss rate: As the packet loss rate increases, the duration of a single freeze also increases, while the buffer level continues to decrease; The results show that the impact of the increase in packet loss rate on high-bitrate videos is more significant than that on low-bitrate videos. To improve the quality of video services, the model quickly reduces the video bitrate when the network packet loss rate is high. The bitrate switching strategy achieves stable video stream transmission and improves the quality of video services. Latency: Increased network latency causes buffer levels to drop, increasing the frequency and duration of freezes. The model evaluates latency and buffer levels, and makes bitrate decisions based on the relative status of the buffer and latency. The strategy reduces the impact of network latency on video block streaming, improving the QoE of video services by increasing the stability of video playback.

4. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 1 is characterized in that: HELL high bandwidth utilization and low latency control logic and feedback learning control method; HELL bitrate controller perceives service scenarios through network QoS, and uses delay and throughput parameters for bitrate control; HELL bitrate controller is the core part, which uses HELL control algorithm for bitrate control. HELL bitrate controller monitors network status and player buffer level to make bitrate decisions, generates decision bitrate and inputs bitrate switching logic; bitrate switching logic is responsible for generating application bitrate according to decision bitrate and video playback status, reading video bitrate ladder from video segment applicant, generating serial number of application video bitrate and downloading it to video buffer.

5. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 4 is characterized in that: HELL introduces smoothing processing for network bandwidth, which improves the ABR controller's resistance to instantaneous service changes by smoothing the network status at the past few discrete application moments; Where I all Indicates the total data size of the video block, n i represents the i-th requested video block, and t represents the transmission duration of the current video block; i represents the sequential index of the currently playing video block, j represents the index of all data sent from the player to the server and the player between the i-1th video block and the ith video block, B ni Indicates the network bandwidth when applying for the ni-th video block.

6. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 5 is characterized in that: In DASH streaming services, the client's application for video resources is discrete in the time domain. In a high-change rate network environment, bandwidth calculation needs to integrate the bandwidth of the link at multiple times to smooth the impact of network fluctuations on controller decisions. The core control logic of HELL is to improve network bandwidth utilization, and its calculation formula is as follows: a ni is the playback flag of the ni-th video block; when the ni-th video block is transmitted and there is a video playing in the streaming service, a ni is 1, otherwise it is 0; Ini represents the total amount of data transmitted between the user and the video server when transmitting the ni-th video block, and Bni represents the network bandwidth between the user and the video server when transmitting the ni-th video block.

7. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 6 is characterized in that: In DASH service, the buffer is consumed when downloading video segments, and the buffer is replenished with space for the length of the video segment after the download is completed; the core formula of the minimum latency strategy used is: C ni Indicates the bandwidth of network transmission, BL ni Indicates the transmission of the nth i The buffer quality when the data blocks are ni is the playback speed gain. When the video is played at multiple speeds (1+p ni ) indicates the playback speed of the video; V ni Indicates the nth i The playback time of the video block, Lnibi indicates the transmission time of the nth video block. i The network delay caused by the video block; Δ i represents the cumulative delay caused by downloading video segments in streaming services; if Δ i <0 holds true, then BL ni =0, playback will stop due to buffer underflow to download video resources, that is, a buffering event occurs.

8. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 4 is characterized in that: The ABR control process based on feedback learning is as follows: Simplifying (1)-(4), the control logic of HELL is as follows: Where BL min and BL max Represents the minimum and maximum values ​​of the buffer level; R min and R max Represents the maximum and minimum video bit rates that the HELL bit rate controller can apply to the video server; The control process of the HELL bitrate control algorithm is as follows. Formulas (9) and (10) give the function for controlling the video bitrate: Formula (10) gives the value of η when R is too large. ni >1 penalty; Formula (11) gives the penalty for R ni For online ABR control, the uncertainty of network QoS between adjacent applications makes it possible to maintain BL in formula (6) ni ∈[0,BL max ], It becomes a very complex control problem. Taking advantage of the fact that the total service time in streaming media service is much longer than the single decision time of the controller, we believe that {R n1 ,R n2 ,…,R nN } is an infinite length sequence; consider BL ni is bounded on average; that is, BL is allowed ni The upper and lower bounds are violated instantaneously, and Rn is calculated using formulas (8) and (9). i Penalize, so that BL satisfies formula (7) in a statistical sense; In the decision-making process of the ABR controller, the full rate ladder is used to calculate the requested rate. Let {R1, R2, … R K } is the code rate ladder, considering the probability simplex ω={ω1,...,ω k ,...,ω K },satisfy: Ω={ω∈R N :ω≥0<||ω||1=1} (10) where ω k Indicates decision Rn i =R k The probability of decision code rate Rn i,req is the inner product of the code rate ladder and ω; Formulas (9) and (10) are convexified and solved using convex optimization theory; Formulas (8) and (9) are adjusted to Formulas (11) and (12), and Formulas (6) and (7) are rearranged to Formula (13); The rate control problem is transformed into a minimization problem. Since formula (13) requires Participate in the calculation, but it is unknown in the decision-making process; consider using the objective function and constraint function about The first-order Taylor expansion estimates the two applications n i The predicted value at time , and the logarithmic penalty function is introduced Apply the control bit rate to satisfy (7); consider: Where A is the rate decision control coefficient, κ is the penalty intensity control coefficient; For the minimization problem, the constraints are satisfied. So it converts to Minimization problem; use gradient descent method to solve, that is, calculate Solving (14) yields ω t The update formula is: Proj[·] is a Euclidean transformation, so that satisfies the probability simplex.

9. The video stream adaptive bit rate control method based on online convex optimization and cyclic learning according to claim 8, characterized in that: Based on HELL's macro control logic and learning method, In the calculation of and is the impact of the last adjustment on bandwidth utilization, and It is the impact of the last adjustment on the service delay; the introduction of feedback enables HELL to gradually achieve continuous control of the service, that is, after the service starts, HELL learns the feedback generated by past control in multiple video applications, gradually improves the accuracy of perception of the service status, and makes adjustments according to the control logic to achieve optimal control of the bit rate.

Images