Cloud game acceleration method based on QoE prediction

Abstract

The present application relates to the cross field of cloud game and artificial intelligence, and particularly to a cloud game acceleration method based on QoE prediction, which compares real experience data such as client frame rate, lag and subjective score with the predicted value of the model, and converts the prediction error into a reward signal of reinforcement learning: if the prediction is accurate, a positive reward is given to strengthen the current decision logic, and if the prediction is biased, a negative reward is given to drive parameter adjustment. Through this continuous online learning closed loop, the expert large model continuously deepens the understanding of individual players and general laws, making the deduction of the digital twin model increasingly accurate and promoting the overall continuous evolution of the system. As a result, the system realizes the transition from the traditional mode of "passive response, one-sided optimization and one-size-fits-all" to the personalized service of "active prediction, global collaboration and one strategy for one person", which can dynamically adapt to the equipment, network and habits of each player and truly provide customized optimal experience guarantee.

Description

Technical Field

[0001] This invention relates to the intersection of cloud gaming and artificial intelligence, specifically a cloud gaming acceleration method based on QoE prediction. Background Technology

[0002] Cloud gaming shifts the computation and rendering tasks of a game from the local terminal to a cloud server and transmits real-time video streams to the user via the network. This core architecture places extremely stringent requirements on network connection latency, stability, and bandwidth. Any fluctuation in network quality can directly lead to lag, degraded image quality, or malfunctions on the user's end, severely impacting the experience. To address the aforementioned challenges, the industry currently employs network acceleration technologies, such as CDN distribution, intelligent routing, or protocol optimization, to improve data transmission paths and efficiency. However, these mainstream technologies have fundamental limitations, and their optimization models are typically reactive and unidimensional.

[0003] Existing technologies are mostly "passive response" models. They usually trigger acceleration actions based on preset static rules or thresholds; this mechanism is essentially a remedy after the user experience quality has deteriorated, and there is an unavoidable decision lag, making it difficult to prevent the occurrence of poor experiences such as lag and screen tearing.

[0004] Secondly, the optimization goals are often one-sided. Traditional methods mostly focus on network service quality indicators in isolation, such as latency and packet loss rate. However, the quality of the end user experience is a complex and comprehensive indicator determined by network conditions, server computing power, client device performance, game content complexity, and even the player's personal operating habits. Simply optimizing a single QoS indicator may not bring about a substantial improvement in QoE.

[0005] Finally, there is a lack of personalized optimization capabilities. Existing acceleration strategies typically adopt a "one-size-fits-all" approach, treating all users as homogeneous network entities. This ignores the significant differences among individual users in terms of device performance, network environment tolerance, and operating habits, failing to provide truly optimal experience guarantees for any specific user. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies and solve at least one of the technical problems mentioned in the background, this invention proposes a cloud gaming acceleration method based on QoE prediction.

[0007] The technical solution adopted by this invention to solve its technical problem is: a cloud gaming acceleration method based on QoE prediction, comprising the following steps; S1. Data Acquisition: Real-time aggregation of user-end data, network link data, and cloud server data through probes deployed on the client side, network side, and cloud. S2. Real-time construction and synchronization of player digital twin models: Real-time collection and aggregation of user-end data, network link data and cloud server data to form a structured and dynamically updated player digital twin model; S3. QoE extrapolation and prediction based on digital twin model: The player's digital twin model is input into a specialized expert model, which simulates the evolution of the QoE state of the digital twin model in the future time window after applying different candidate acceleration strategies, and outputs the corresponding predicted QoE curve. S4. Optimal path selection and instruction issuance: Based on the preset optimization objective, evaluate multiple predicted QoE curves, select the candidate acceleration strategy corresponding to the optimal QoE evolution path, and parse it into a multi-objective acceleration instruction set and issue it to the corresponding execution unit. S5. Model Calibration and Evolution: Collect real QoE data after strategy execution, compare it with the predicted QoE curve, and use the deviation as a feedback signal to perform reinforcement learning optimization on the expert large model, so as to realize the calibration of the digital twin model and the self-evolution of the system.

[0008] Preferably, the user-side data in S2 includes device model, CPU / GPU load, local rendering frame rate, player operation command frequency and complexity; the network link data includes round-trip latency, packet loss rate, network jitter and available bandwidth; and the cloud server data includes node geographical location, computing power load, game type and scene complexity.

[0009] Preferably, the candidate acceleration strategies in S2 include, but are not limited to, one or more combinations of switching cloud service nodes, adjusting video stream bitrate, changing network transmission protocol, and adjusting cloud rendering priority.

[0010] Preferably, the multi-target acceleration instruction set in S4 includes resource layer instructions, network layer instructions, and encoding layer instructions, which are used to achieve cross-layer global collaborative optimization; wherein, the resource layer instructions are used to dynamically adjust the allocation of cloud GPU rendering resources; the network layer instructions are used to switch network transmission paths or protocols; and the encoding layer instructions are used to adaptively adjust video encoding parameters according to the game screen content.

[0011] Preferably, the reinforcement learning optimization process in S5 specifically involves using the deviation between the predicted result and the actual result as a reward / penalty signal, and updating the parameters of the expert large model online using a policy gradient method or a value function method.

[0012] Preferably, the digital twin construction module in S2 is used to collect multi-source heterogeneous data in real time, construct and synchronously update the player's digital twin model.

[0013] Preferably, the S3 integrates the expert big model, which is used to receive the digital twin model, perform scenario deduction, and output the predicted QoE curves under each candidate strategy.

[0014] Preferably, S4 includes a decision and instruction generation module for evaluating the predicted QoE curve, selecting the optimal strategy, and generating a multi-objective acceleration instruction set across resource, network, and coding layers.

[0015] Preferably, S5 includes an execution and feedback module, which is used to send the instruction set to the corresponding execution unit of the cloud gaming platform and collect real QoE data to feed back to the inference and prediction module for the reinforcement learning calibration of the model.

[0016] The advantages of this invention are: This invention comprehensively calculates the frame rate, lag duration, and subjective ratings reported by client-side data points and compares them with previous predictions. The prediction error is converted into a reward signal in reinforcement learning. For example, if the prediction is accurate, a positive reward is given to reinforce the model's internal decision-making logic that led to the accurate prediction; if the prediction deviation is large, a negative reward is given to prompt the model to adjust its internal inference parameters. Through this continuous online learning, the expert model's understanding of specific players and general patterns deepens, making the digital twin model's deductions increasingly accurate. The entire system evolves continuously, achieving personalized service from "one-size-fits-all" to "one-person-one-policy" based on a precise digital twin model. The system can dynamically understand and adapt to each player's device, network, and habits, providing a truly tailored optimal experience. This solves three limitations of existing cloud gaming acceleration technologies: first, passive response, relying on fixed rules to remedy experience degradation, resulting in delayed decision-making; second, one-sided optimization, focusing only on network indicators, while the actual experience is determined by multiple factors; and third, lack of personalization, adopting a "one-size-fits-all" strategy that cannot adapt to user differences and cannot provide optimal protection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is the QoE curve of the present invention. Detailed Implementation

[0019] 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.

[0020] The following is in conjunction with the appendix Figure 1 and Figure 2 This application will be described in further detail. This application discloses a cloud gaming acceleration method based on QoE prediction, including the following steps; S1. Data Acquisition: Real-time aggregation of user-end data, network link data, and cloud server data through probes deployed on the client side, network side, and cloud. S2. Real-time construction and synchronization of player digital twin models: Real-time collection and aggregation of user-end data, network link data and cloud server data to form a structured and dynamically updated player digital twin model; S3. QoE extrapolation and prediction based on digital twin model: The player's digital twin model is input into a specialized expert model, which simulates the evolution of the QoE state of the digital twin model in the future time window after applying different candidate acceleration strategies, and outputs the corresponding predicted QoE curve. S4. Optimal path selection and instruction issuance: Based on the preset optimization objective, evaluate multiple predicted QoE curves, select the candidate acceleration strategy corresponding to the optimal QoE evolution path, and parse it into a multi-objective acceleration instruction set and issue it to the corresponding execution unit. S5. Model Calibration and Evolution: Collect real QoE data after strategy execution, compare it with the predicted QoE curve, and use the deviation as a feedback signal to perform reinforcement learning optimization on the expert large model, so as to realize the calibration of the digital twin model and the self-evolution of the system.

[0021] Example 1 S1. Data Acquisition: Real-time data acquisition is performed synchronously at three levels using a lightweight software probe (SDK) or hardware probe.

[0022] Client-side probe: Integrated into the cloud gaming client, it collects user-side data. This includes: terminal device model and operating system version; real-time CPU usage, GPU usage, memory usage, and temperature; local video decoding frame rate, display frame rate, and frame rendering time; player input commands; such as the timing, frequency, and spatial complexity of key presses and mouse movements, for example, the operation density per unit time.

[0023] Network-side probes: Deployed at user terminal access points, edge networks, and core network nodes to collect network link data. Specifically, this includes: end-to-end round-trip latency from the client to the cloud game server, and RTT to each candidate edge node; packet loss rate on the network path; network jitter; and currently available bandwidth estimated through active probing or passive analysis.

[0024] Cloud server probe: Deployed in the cloud gaming rendering server and management cluster to collect data from the cloud server. Specifically, this includes: the physical geographical location of the server or virtual machine currently providing services to players; the server's real-time GPU / CPU computing load and video memory usage; the identifier and version of the currently running game application; and rendering scene complexity information fed back in real time by the game engine, such as the number of polygons, lighting and shadow effect levels, and the number of players on screen simultaneously.

[0025] S2. Real-time Construction and Synchronization of Player Digital Twin Models: Three types of heterogeneous data streams from S1 are transmitted in real time to a centralized twin-building engine. This system aggregates these three types of heterogeneous data in real time through probes deployed on clients, network nodes, and servers: client-side data, such as device model, CPU / GPU load, local rendering frame rate, and the frequency and complexity of player input commands; network link data, such as round-trip latency, packet loss rate, network jitter, and available bandwidth; and cloud server data, such as node geographic location, computing load, currently running game type, and scene complexity. After cleaning and normalization, this data is injected into a structured dynamic data body, forming the player's current "digital twin" model, which is then updated synchronously in real time at a high frequency.

[0026] S3. QoE Inference and Prediction Based on Digital Twin Model: The player's digital twin model generated in step S2 is used as input and fed into a large expert model trained specifically for cloud gaming QoE prediction, such as a deep neural network based on the Transformer architecture. This model has a built-in understanding of the cloud gaming business logic; its core operation is to perform parallel "What-if" scenario inference: the model internally simulates how the digital twin state will evolve if the system adopts a series of predefined candidate acceleration strategies within a short period in the future, such as the next 5-10 seconds. The candidate acceleration strategies include, but are not limited to: switching to another geographically or load-optimized edge computing node; dynamically increasing or decreasing the encoding bitrate of the video stream; switching from TCP protocol to a private low-latency transmission protocol based on UDP; increasing or decreasing the rendering priority of the player session on the cloud GPU.

[0027] For each simulation strategy, the expert model comprehensively extrapolates the impact of multiple intermediate variables, such as changes in network latency, potential packet loss, cloud rendering queuing time, and client decoding pressure, ultimately outputting a quantified predicted QoE curve. The QoE value can be a comprehensive score, such as 1-5, which quantifies the user experience level at each moment within the prediction period.

[0028] like Figure 2 As shown: The baseline (current state, blue curve) represents the unchanged acceleration strategy. QoE has steadily decreased from 65 to 40, indicating a gradual deterioration in user experience quality.

[0029] Strategy A (Switch to node B and increase bitrate, orange curve): Switching to node B and increasing bitrate resulted in a steady increase in QoE from 68 to 88, achieving the best prediction performance.

[0030] Strategy B (Keep the current node but adopt a more aggressive transport protocol, green curve): Keep the current node but adopt a more aggressive transport protocol. QoE improved from 66 to 85 and then stabilized, showing good prediction performance.

[0031] Through the above "What-if" deduction, the expert big model can generate a complete QoE evolution curve for each candidate acceleration strategy. The system can select the optimal strategy based on these prediction curves, realizing a paradigm shift from "passive response" to "active prediction".

[0032] S4. Optimal Path Selection and Command Issuance: The decision engine receives multiple predicted QoE curves output by the expert large model. Based on a preset global optimization objective, such as "maximizing the average QoE value within the prediction window" or "minimizing the variance of the QoE value to ensure stationarity," the engine evaluates and compares these curves to select the optimal QoE evolution path and its corresponding candidate acceleration strategy.

[0033] Once a strategy is selected, the instruction generator parses it into a set of specific, executable multi-target acceleration instructions, and simultaneously distributes them to the corresponding execution units on the cloud gaming platform to achieve collaborative optimization. Resource layer command: Issued to the cloud resource scheduler. The command may be: to increase the scheduling priority of the player's rendering instance on the GPU[GPU_ID] by one level, or to reserve additional video memory buffers for it.

[0034] Network layer command: Issued to the SD-WAN controller or intelligent routing module. The command might be: Switch the game stream path destined for the IP from the current node [Node_A] to the edge node [Node_B], enable forward error correction (FEC) coding, and set the redundancy ratio to 20%. Encoding layer instruction: Sent to the cloud video encoder. The instruction may be: For the game scene [Scene_ID], shorten the GOP length from 250 frames to 60 frames and enable perceptual encoding to focus on ensuring the image quality of the central area of ​​the screen.

[0035] S5. Model Calibration and Evolution: During an observation period following the implementation of the acceleration strategy, the system collects real QoE data through client probes and the business quality monitoring system. This data can be obtained through methods such as questionnaires and comprehensive calculation of edge performance indicators. The real QoE change curve is then compared with the prediction curve of the corresponding strategy in step S3 to calculate the prediction bias.

[0036] This bias is translated into a reward or penalty signal within the reinforcement learning framework. For example, a positive reward is given if the prediction is accurate, and a negative reward is given if the prediction is overly optimistic or pessimistic. This signal is fed back to the expert large model in step S3. The model uses this signal to fine-tune its internal network parameters online or near-line through policy gradients, such as the PPO algorithm, or value function updates, such as the DQN algorithm. Through a massive number of "prediction-execution-feedback" loops, the expert large model's ability to infer complex causal relationships continuously improves, making the predictions of the player's digital twin model increasingly accurate. The entire system possesses the ability to self-evolve and automatically calibrate, adapting to constantly changing network environments, game content, and user behavior patterns.

[0037] Example 2 In some embodiments, the method may be implemented using a dedicated system. This system may include: The digital twin building module is configured to execute steps S1 and S2, which are responsible for data collection, aggregation, and model building.

[0038] The inference and prediction module, which integrates the expert big model, is configured to execute step S3, which is responsible for receiving the digital twin model and performing scenario inference.

[0039] The decision and instruction generation module is configured to execute step S4, which is responsible for evaluating curves, selecting strategies, and generating cross-layer instruction sets.

[0040] The execution and feedback module is configured to distribute instruction sets to each execution unit and collect real QoE data to feed back to the inference and prediction module in order to execute the reinforcement learning optimization process in step S5.

[0041] These modules can be deployed on the cloud control plane in the form of software, hardware, or a combination of both, and work together to achieve the methods described in this invention.

[0042] Working principle: After the data collected by S1 is cleaned, time-aligned and normalized, it is combined to form a digital twin that can dynamically and comprehensively reflect the current joint state of the player-network-cloud environment.

[0043] Subsequently, the digital twin model is fed into the inference and prediction module. The core of this module is a large expert model pre-trained and continuously fine-tuned using massive amounts of cloud gaming scenario data. Upon receiving the current state S_t of the digital twin and a set of candidate acceleration strategies {A, B, C...}, the model performs forward inference in parallel. For example, it simulates how the state of the digital twin will evolve over time to S_{t+Δt} if strategy A (switching to edge node B and enabling FEC) is adopted, and calculates the predicted QoE curve QoE_A for this path. Similarly, QoE_B, QoE_C, etc., are obtained. The inference considers the interactions of various dimensions; for example, network switching may slightly increase latency but improve jitter, thus affecting coding efficiency and the final subjective experience.

[0044] The decision-making and instruction generation module receives all predicted curves and performs a comprehensive evaluation based on preset optimization objectives (such as maximizing the expected QoE value within the next 5 seconds while minimizing its variance) to select the optimal path. Assuming strategy B is selected, the module will decompose it into specific execution instructions. For example... Figure 2 As shown, this set of instructions is coordinated: To the resource layer: Send instructions to the cloud rendering cluster to appropriately increase the GPU rendering priority of the current game instance.

[0045] At the network layer: Send instructions to the SD-WAN controller to switch user flow paths to the better-predicted node and adjust transport protocol parameters.

[0046] To the encoding layer: Send instructions to the video encoder to adaptively adjust the GOP structure and enable region of interest encoding for upcoming high-dynamic game scenes.

[0047] After the command is issued, the system enters the feedback and calibration phase. The execution and feedback module collects the actual QoE data generated, which can be comprehensively calculated using frame rate, stutter duration, and subjective scores reported by client-side tracking points, and compared with previous predictions. Prediction errors are converted into reward signals in reinforcement learning. For example, if the prediction is accurate, a positive reward is given to reinforce the model's internal decision-making logic that led to the accurate prediction; if the prediction deviation is large, a negative reward is given to prompt the model to adjust its internal inference parameters. Through this continuous online learning, the expert model's understanding of specific players and general patterns deepens, making the digital twin model's inferences increasingly accurate, and allowing the entire system to continuously evolve.

[0048] This invention represents a fundamental shift in cloud gaming acceleration technology, moving from a "passive response" to a "proactive prediction" model. By constructing and extrapolating a digital twin model of a "virtual player," the decision-making basis transforms from lagging, single-point QoS data to proactively predicting future QoE states, thus preventing experience degradation earlier. Building upon this, it elevates the dimension from "network optimization" to "global collaboration." By issuing collaborative instruction sets across resource, network, and encoding layers, a global optimization system is constructed, unlocking potential untapped by traditional methods. Finally, relying on a precise digital twin model, it achieves personalized service, moving from a "one-size-fits-all" approach to a "one-person-one-policy" system. The system dynamically understands and adapts to each player's device, network, and habits, providing a truly tailored optimal experience. These three transformations work together to deliver an unprecedentedly smooth, stable, and highly personalized cloud gaming experience for users.

[0049] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A cloud gaming acceleration method based on QoE prediction, characterized in that: Includes the following steps; S1. Data Acquisition: Real-time aggregation of user-end data, network link data, and cloud server data through probes deployed on the client side, network side, and cloud. S2. Real-time construction and synchronization of player digital twin models: Real-time collection and aggregation of user-end data, network link data and cloud server data to form a structured and dynamically updated player digital twin model; S3. QoE extrapolation and prediction based on digital twin model: The player's digital twin model is input into a specialized expert model, which simulates the evolution of the QoE state of the digital twin model in the future time window after applying different candidate acceleration strategies, and outputs the corresponding predicted QoE curve. S4. Optimal path selection and instruction issuance: Based on the preset optimization objective, evaluate multiple predicted QoE curves, select the candidate acceleration strategy corresponding to the optimal QoE evolution path, and parse it into a multi-objective acceleration instruction set and issue it to the corresponding execution unit. S5. Model Calibration and Evolution: Collect real QoE data after strategy execution, compare it with the predicted QoE curve, and use the deviation as a feedback signal to perform reinforcement learning optimization on the expert large model, so as to realize the calibration of the digital twin model and the self-evolution of the system.

2. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The user-side data in S2 includes device model, CPU / GPU load, local rendering frame rate, and frequency and complexity of player operation commands; the network link data includes round-trip latency, packet loss rate, network jitter, and available bandwidth; and the cloud server data includes node geographical location, computing load, game type, and scene complexity.

3. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The candidate acceleration strategies in S2 include, but are not limited to, switching cloud service nodes, adjusting video stream bitrate, changing network transmission protocol, and adjusting cloud rendering priority, or one or more combinations thereof.

4. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The multi-target acceleration instruction set in S4 includes resource layer instructions, network layer instructions, and encoding layer instructions, which are used to achieve cross-layer global collaborative optimization. Among them, the resource layer instructions are used to dynamically adjust the allocation of cloud GPU rendering resources; the network layer instructions are used to switch network transmission paths or protocols; and the encoding layer instructions are used to adaptively adjust video encoding parameters according to the game screen content.

5. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The reinforcement learning optimization process in S5 specifically involves using the deviation between the predicted result and the actual result as a reward / penalty signal, and updating the parameters of the expert large model online through the policy gradient method or the value function method.

6. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The digital twin construction module in S2 is used to collect multi-source heterogeneous data in real time, construct and synchronously update the player's digital twin model.

7. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The S3 integrates the expert big model, which is used to receive the digital twin model, perform scenario deduction, and output the predicted QoE curves under each candidate strategy.

8. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The S4 includes a decision and instruction generation module, which is used to evaluate the predicted QoE curve, select the optimal strategy, and generate a multi-objective acceleration instruction set across resource, network, and coding layers.

9. The cloud gaming acceleration method based on QoE prediction according to claim 1, characterized in that: The S5 includes an execution and feedback module, which is used to send the instruction set to the corresponding execution unit of the cloud gaming platform and collect real QoE data to feed back to the inference and prediction module for the reinforcement learning calibration of the model.

Images