Real-time scheduling method of wireless network based on large language model

Abstract

The invention provides a real-time scheduling method for a wireless network based on a large language model, the wireless network comprises a scheduling node and a plurality of user nodes, and the scheduling node is provided with a scheduling agent, a historical memory pool, an reflection agent and a suggestion buffer area. The method comprises the following steps executed in each time slot: step S1, the scheduling node acquires global information of a wireless network under the current time slot and calculates states of all data streams under the current time slot according to a preset calculation rule; and S2, the scheduling agent generates a scheduling decision of the current time slot according to a preset scheduling prompt word by using a large language model configured by the scheduling agent based on the states of all data streams under the current time slot and decision suggestions stored in the suggestion buffer area and by taking a maximized reward as a target. According to the technical scheme of the invention, by introducing the large language model, the problem of insufficient adaptability in a dynamic network environment in the prior art is solved.

Description

Technical Field

[0001] This invention relates to information scheduling technology in wireless networks, specifically to data flow scheduling technology in wireless networks, and more specifically, to a real-time scheduling method for wireless networks based on a large language model. Background Technology

[0002] With the widespread application of real-time wireless networks, various scenarios with extremely high latency requirements are gradually becoming more common, such as virtual reality and intelligent unmanned cluster communication. In these scenarios, it is essential to ensure that data packets are delivered within the specified deadline; otherwise, expired data packets will lose their value and be discarded.

[0003] In recent years, how to schedule real-time data streams to better meet the demand for timely throughput has attracted widespread attention. However, most work has focused on the frame-synchronized traffic model. This model assumes that all data streams generate new data packets simultaneously at the start of each fixed period (a frame consisting of T time slots), and that all data packets have the same, relatively long deadline (usually related to the frame period). However, this highly simplified model fails to reflect the complexity of many real-world application scenarios. In reality, the end-to-end latency requirements of many real-time services are much shorter than the data packet generation period (i.e., frame length T), causing traditional scheduling algorithms based on the frame-synchronized model (such as priority scheduling and round-robin) to be unable to effectively meet these more pressing deadline constraints, resulting in poor performance in actual deployments.

[0004] To overcome the limitations of frame synchronization models, some research has turned to more general traffic models, which allow packets to arrive in any time slot and have their own independent, potentially short, deadlines. However, scheduling algorithms for general models face significant challenges: their computational complexity is often extremely high (e.g., involving complex combinatorial optimization), or they struggle to provide stable and reliable performance guarantees in dynamically changing wireless network environments. This often results in unsatisfactory performance of these algorithms in practical applications.

[0005] Traditional deterministic scheduling methods, such as greedy algorithms or dynamic programming, have fundamental shortcomings when applied to real-time wireless network scheduling. These methods typically fail to adequately account for the inherent unreliability of wireless channels (such as time-varying channel quality, interference, and packet loss). They assume that transmission will always succeed or are based on idealized models, making it difficult to achieve ideal real-time throughput performance in real-world wireless environments where channel fading exists.

[0006] In recent years, intelligent scheduling methods based on deep reinforcement learning (DRL) have shown potential in handling complex wireless environments and optimizing real-time throughput, typically outperforming traditional methods. However, these methods also have significant bottlenecks: insufficient generalization ability, making it difficult to effectively handle new scenarios with increased data flow in dynamic network environments (i.e., model performance may significantly degrade at flow scales not seen during training); and the training process is too time-consuming, requiring substantial data and computational resources for offline training, limiting their applicability in real-time systems that require rapid responses to network changes.

[0007] In summary, existing technologies for real-time wireless network scheduling face the following main problems: (1) Mainstream research relies on unrealistic frame synchronization traffic models, which cannot meet the urgent delay requirements that are common in practical applications and are much shorter than the generation period; (2) For more general traffic models, existing algorithms suffer from excessive computational complexity or performance difficulties; (3) Traditional scheduling methods (such as greedy algorithms and dynamic programming) ignore the unreliability of wireless channels, resulting in poor actual performance; (4) Although deep reinforcement learning-based scheduling methods perform well in terms of throughput, they have poor generalization ability (cannot adapt to the dynamic increase of the number of streams) and excessively high training costs. These shortcomings together limit the effectiveness and practicality of existing technologies in meeting the increasingly demanding real-time wireless application scenarios.

[0008] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solutions of the present invention, and does not imply that the relevant information is necessarily prior art. In the absence of evidence indicating that the relevant information was disclosed before the filing date of this invention, the relevant information should not be considered prior art. Summary of the Invention

[0009] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a real-time scheduling method for wireless networks based on a large language model and a wireless network.

[0010] The objective of this invention is achieved through the following technical solutions.

[0011] According to a first aspect of the present invention, a real-time scheduling method for a wireless network based on a large language model is provided, for determining the data streams scheduled by the wireless network in each time slot. The wireless network includes a scheduling node and multiple user nodes. The scheduling node is configured with a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer. The scheduling agent is configured with a large language model to generate scheduling decisions for each time slot. The historical memory pool is used to store the scheduling decisions for each time slot. The reflective agent is configured with a large language model to generate decision suggestions. The suggestion buffer is used to store decision suggestions. The method includes performing the following steps in each time slot: Step S1: The scheduling node obtains the global information of the wireless network in the current time slot and calculates the state of all data streams in the current time slot according to a preset calculation rule; wherein the global information of the wireless network includes a quintuple corresponding to each data stream. And the data packets contained in each data stream; the five-tuple corresponding to each data stream includes the data stream transmission start point, data stream period, data packet delay, data packet arrival rate, and data packet transmission probability; Step S2, the scheduling agent uses its own configured large language model to generate the scheduling decision for the current time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer, with the goal of maximizing the reward, according to the preset scheduling prompt words, to determine the data stream to be scheduled in the current time slot, and uses the data packet transmission probability corresponding to the data stream as the decision reward for the current time slot; and stores the state, scheduling decision, and decision reward of all data streams corresponding to the current time slot as the decision information for the current time slot in the historical memory pool, so that the reflective agent uses its own configured large language model to generate decision suggestions based on the decision information of each time slot stored in the historical memory pool.

[0012] In some embodiments of the present invention, the preset calculation rule is as follows:

[0013]

[0014] in,

[0015]

[0016]

[0017]

[0018]

[0019] in, express Time slot data stream state, express Time slot data stream The degree of urgency, express Time slot data stream The arrival factor of the data packets. express Time slot data stream The number of data packets that have not expired. Represents data stream The Middle One incoming data packet, express Time slot data stream The Middle The remaining time before the arrival of each data packet expires. express Time slot data stream The Middle The expiration time of each arriving data packet. express Time slot data stream The Middle The arrival time of each arriving data packet. Represents data stream The corresponding data packet latency, Represents data stream The starting point of data stream transmission Represents data stream The corresponding data stream cycle.

[0020] In some embodiments of the present invention, the method further includes: the reflective agent selecting decision information from the historical memory pool of multiple time slots with the most recent consecutive preset values ​​at preset time intervals, so that the large language model configured by itself generates new decision suggestions based on the decision information of multiple time slots according to preset reflective prompt words, and replaces the decision suggestions previously stored in the suggestion buffer with the newly generated decision suggestions.

[0021] In some embodiments of the present invention, the preset time interval is 10 time slots.

[0022] In some embodiments of the present invention, the preset value is 10.

[0023] In some embodiments of the present invention, the large language model configured for the scheduling agent and the reflective agent is GPT-4o.

[0024] According to a second aspect of the present invention, a wireless network is provided, the wireless network including a scheduling node and a plurality of user nodes, the scheduling node being configured with a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer, the scheduling agent being configured with a large language model to generate scheduling decisions for each time slot, the historical memory pool being used to store the scheduling decisions for each time slot, the reflective agent being configured with a large language model to generate decision suggestions, and the suggestion buffer being used to store decision suggestions, wherein: the wireless network is configured to implement real-time scheduling using the method described in the first aspect of the present invention to determine the data flow scheduled by the wireless network in each time slot.

[0025] Compared with the prior art, the advantages of the present invention are: the introduction of a large language model to generate scheduling decisions for each time slot based on the state of all data streams and decision suggestions in each time slot, so as to realize data stream scheduling for each time slot; and the introduction of a reflection mechanism to enable another large language model to generate new decision suggestions based on decision information from multiple time slots to guide the generation of decision suggestions, thus solving the problem of insufficient adaptability in the face of dynamic network environments. Attached Figure Description

[0026] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0027] Figure 1 This is a flowchart illustrating a real-time scheduling method for a wireless network based on a large language model according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the framework of a real-time scheduling method for wireless networks based on a large language model according to an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram illustrating an example of a wireless network according to an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram illustrating experimental results in a dense traffic network environment, using the number of data streams as a variable, according to an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram illustrating experimental results in a sparse traffic network environment, using the number of data streams as a variable, according to an embodiment of the present invention.

[0032] Figure 6 This is a schematic diagram illustrating experimental results in a dense traffic network environment, using the delay period as a variable, according to an embodiment of the present invention.

[0033] Figure 7 This is a schematic diagram illustrating experimental results in a sparse traffic network environment with time limit as the variable, according to an embodiment of the present invention.

[0034] Figure 8This is a schematic diagram illustrating experimental results in a dynamic network environment according to an embodiment of the present invention;

[0035] Figure 9 A schematic diagram illustrating the results of an experiment verifying the reflection mechanism according to an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0037] Before describing the present invention, a brief introduction to the scheduling background of wireless networks will be provided to facilitate a better understanding of the present invention.

[0038] Wireless networks typically employ Time Division Multiple Access (TDMA) mechanisms. This mechanism divides time into fixed-length time slots and allocates these slots to different data streams. Based on the principle of time division multiplexing, each time slot can only be allocated to one data stream, effectively avoiding interference problems caused by multiple transmissions occurring simultaneously. If multiple data streams are incorrectly allocated to the same time slot, all data streams will fail to transmit due to collisions. Therefore, the core issue of wireless network scheduling is how to determine the data stream scheduled for each time slot.

[0039] As mentioned in the background section, existing technologies for real-time wireless network scheduling face the following main problems: (1) Mainstream research relies on unrealistic frame synchronization traffic models, which cannot meet the urgent delay requirements that are much shorter than the generation period in practical applications; (2) For more general traffic models, existing algorithms suffer from excessive computational complexity or performance that is difficult to guarantee; (3) Traditional scheduling methods (such as greedy algorithms and dynamic programming) ignore the unreliability of wireless channels, resulting in poor actual performance; (4) Although deep reinforcement learning-based scheduling methods perform well in terms of throughput, they have poor generalization ability (cannot adapt to the dynamic increase of the number of streams) and excessively high training costs. These defects together limit the effectiveness and practicality of existing technologies in meeting the increasingly demanding real-time wireless application scenarios.

[0040] To address the aforementioned issues, the inventors propose introducing a large language model into wireless networks to generate scheduling decisions in real time, thereby meeting the effectiveness and practicality requirements of real-time wireless applications. Based on this, the present invention proposes a real-time scheduling method for wireless networks based on a large language model. In this method, a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer are first configured for the scheduling nodes in the wireless network. The scheduling agent is configured with a large language model to generate scheduling decisions; the historical memory pool stores these decisions; the reflective agent is configured with a large language model to generate decision suggestions to guide the scheduling agent in generating scheduling decisions; and the suggestion buffer stores these decision suggestions. Then, in each time slot, data flow scheduling is implemented as follows: the scheduling node obtains global information about the wireless network in the current time slot to calculate the state of all data flows in the current time slot; the scheduling agent uses its configured large language model to generate scheduling decisions based on the states of all data flows in the current time slot and the decision suggestions, thereby determining the data flow to be scheduled in the current time slot.

[0041] In summary, such as Figure 1 As shown, this invention provides a real-time scheduling method for wireless networks based on a large language model, used to determine the data streams scheduled by the wireless network in each time slot. The wireless network includes a scheduling node and multiple user nodes. The scheduling node is configured with a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer. The scheduling agent is configured with a large language model to generate scheduling decisions for each time slot. The historical memory pool stores the scheduling decisions for each time slot. The reflective agent is configured with a large language model to generate decision suggestions. The suggestion buffer stores the decision suggestions. The method includes performing the following steps in each time slot: Step S1: The scheduling node obtains the global information of the wireless network in the current time slot and calculates the state of all data streams in the current time slot according to a preset calculation rule. The global information of the wireless network includes the quintuple corresponding to each data stream and the state of each... Each data stream contains data packets; the five-tuple corresponding to each data stream includes the data stream transmission start point, data stream period, data packet delay, data packet arrival rate, and data packet transmission probability; Step S2: The scheduling agent uses its own configured large language model to generate the scheduling decision for the current time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer, with the goal of maximizing the reward, according to the preset scheduling prompt words, to determine the data stream to be scheduled in the current time slot, and uses the data packet transmission probability corresponding to the data stream as the decision reward for the current time slot; and stores the state, scheduling decision, and decision reward of all data streams corresponding to the current time slot as the decision information for the current time slot in the historical memory pool, so that the reflective agent uses its own configured large language model to generate decision suggestions based on the decision information of each time slot stored in the historical memory pool.

[0042] To better understand the present invention, each step will be described in detail below with reference to specific embodiments.

[0043] I. Step S1

[0044] In step S1, the scheduling node obtains the global information of the wireless network in the current time slot and calculates the status of all data streams in the current time slot according to the preset calculation rules; wherein, the global information of the wireless network includes the five-tuple corresponding to each data stream and the data packets contained in each data stream; the five-tuple corresponding to each data stream includes the data stream transmission start point, data stream period, data packet delay, data packet arrival rate and data packet transmission probability.

[0045] The quintuple corresponding to each data stream can be represented as: , This is the starting point of the data stream transmission, representing the offset of the arrival time of the first data packet in the data stream relative to the start time of the wireless network, i.e., the moment when the first data packet in the data stream arrives; The data stream period represents the time interval between the arrival of two adjacent data packets in the data stream, i.e., every [number] data packets in the data stream. The time may be until a data packet arrives (for periodic traffic, the data stream period is fixed; for bursty or random traffic, the data stream period is not fixed). Data packet delay refers to the time interval of data packets in a data stream, that is, the maximum allowed time from the arrival of a data packet to the completion of its transmission (e.g., ...). The delay period for data packets arriving at a specific time is... It is necessary to Transmission must be completed before the specified time; otherwise, it will expire and be discarded. The packet arrival rate represents the rate at which packets arrive in the data stream at intervals of 10 ... The actual probability of a data packet arriving at a given time. The data packet transmission probability represents the probability that a data packet will be successfully transmitted on the link corresponding to the data stream (data stream). The probability of a data packet being successfully transmitted is... The probability of transmission failure is , This parameter directly reflects the unreliability of the channel in a wireless network; a lower value indicates a lower reliability. A value of 0 indicates that more retransmission attempts may be needed to successfully transmit data packets, suggesting poor channel quality of the link through which the data stream travels.

[0046] Suppose there are two data streams with quintuples as follows: and Among them, data stream The data stream transmission starts at 0, indicating a data stream. Data packets may arrive in time slot 0; data stream The data stream cycle is 3, indicating that the data stream... A data packet may arrive every 3 time slots; data stream The data packet latency is 4, indicating that the data stream... Data packets need to be transmitted within 3 time slots after arrival; data stream A data packet arrival rate of 1 indicates a data stream. The probability of a data packet arriving every 3 time slots is 1; data stream The data packet transmission probability is 0.9, indicating that the data stream... The probability of successfully transmitting data packets on the corresponding link is 0.9. Data Stream The data stream transmission starts at 0, indicating a data stream. Data packets may arrive in time slot 0; data stream The data stream cycle is 2, indicating that the data stream... A data packet may arrive every two time slots; data stream The data packet latency is 3, indicating that the data stream... Data packets need to be transmitted within two time slots after arrival; data stream The data packet arrival rate is 0.9, indicating a data flow... The probability of a data packet arriving every two time slots is 0.9; data stream The data packet transmission probability is 0.7, indicating that the data stream... The probability of successfully transmitting data packets on the corresponding link is 0.7.

[0047] II. Step S2

[0048] In step S2, the scheduling agent uses its configured large language model to generate a scheduling decision for the current time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer, aiming to maximize the reward. This decision determines the data stream scheduled for the current time slot and uses the transmission probability of the corresponding data packet as the decision reward for the current time slot. Furthermore, the state, scheduling decisions, and decision rewards of all data streams corresponding to the current time slot are stored as decision information in the historical memory pool. This allows the reflective agent to generate decision suggestions based on the decision information of each time slot stored in the historical memory pool using its configured large language model. It should be noted that after determining the data stream scheduled for the current time slot, the data packet that is about to expire in the queue of that data stream is selected for transmission.

[0049] According to one embodiment of the present invention, the preset calculation rule is as follows:

[0050]

[0051] in,

[0052]

[0053]

[0054]

[0055]

[0056] in, express Time slot data stream state, express Time slot data stream The degree of urgency, express Time slot data stream The arrival factor of the data packets. express Time slot data stream The number of data packets that have not expired. Represents data stream The Middle One incoming data packet, express Time slot data stream The Middle The remaining time before the arrival of each data packet expires. express Time slot data stream The Middle The expiration time of each arriving data packet. express Time slot data stream The Middle The arrival time of each arriving data packet. Represents data stream The corresponding data packet latency, Represents data stream The starting point of data stream transmission Represents data stream The corresponding data stream cycle. Among them, The larger the value, the more likely it is to represent Time slot data stream The more urgent, Time indicates Time slot data stream A new data packet will be received. Time indicates Time slot data stream No new data packets will be received.

[0057] As can be seen from the foregoing embodiments, the state of each data stream in each time slot includes the urgency level and packet arrival factor of that data stream. Therefore, the states of all data streams in each time slot can be represented as a set. ,in, express The state of all data streams under a time slot, express Time slot data stream The degree of urgency, express Time slot data stream The arrival factor of the data packets. express The number of data streams in a time slot.

[0058] As can be seen from the foregoing embodiments, the state, scheduling decisions, and decision rewards of all data streams corresponding to each time slot are stored in the historical memory pool as the decision information for that time slot. The decision information for each time slot can be represented as a triple, for example, The decision information of a time slot can be represented as , express The state of all data streams under a time slot, express Time slot scheduling decisions, ( Indicates in Time slot scheduling data stream ), express Time slot decision rewards .

[0059] As described in the foregoing embodiments, the scheduling agent generates scheduling decisions in each time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer. Considering the complexity and variability of the network environment, the decision suggestions stored in the suggestion buffer should also be adjusted as the network environment changes to better guide the scheduling agent in generating scheduling decisions. Therefore, the inventors propose a reflection mechanism to transform the decision information of historical time slots into practical decision suggestions to cope with the complex and variable network environment.

[0060] According to one embodiment of the present invention, the method further includes: the reflexive agent selecting decision information from multiple time slots of the most recent consecutive preset values ​​from the historical memory pool at preset time intervals, so that its configured large language model generates new decision suggestions based on the decision information of the multiple time slots according to preset reflexive prompts, and replaces the decision suggestions previously stored in the suggestion buffer with the newly generated decision suggestions. This reflexive mechanism can transform the decision information of historical time slots into decision suggestions, enabling the scheduling agent to generate scheduling decisions adapted to different network environments under the guidance of these decision suggestions, achieving high performance under various traffic patterns, channel conditions, and traffic environments.

[0061] According to one embodiment of the present invention, the preset time interval is 10 time slots. It should be noted that the preset time interval can be flexibly determined according to actual needs, such as being set to 5, 7 or 9 time slots, etc., and the specific value is not limited.

[0062] According to one embodiment of the present invention, the preset value is 10. It should be noted that the preset value can be determined according to actual needs, such as 5, 10, 15, etc., and the specific value is not limited.

[0063] According to one embodiment of the present invention, the large language model configured for the scheduling agent and the reflective agent is GPT-4o. The large language model is an artificial intelligence system pre-trained on massive text corpora, possessing powerful language understanding, reasoning analysis, and knowledge integration capabilities. This type of model adopts a Transformer architecture, capable of processing complex contextual information and generating coherent natural language output. Therefore, using a large language model can dynamically handle varying data stream volumes without pre-defining fixed input structure dimensions. This overcomes the inherent limitations of existing technologies in handling high-traffic scenarios and can seamlessly adapt to the expansion of wireless network scale. Furthermore, the large language model can be directly deployed and adapted to various wireless network scenarios without requiring additional iterative training for specific wireless network scenarios.

[0064] To better understand this invention, the following will be used... Figure 2 The process framework diagram shown is used for illustration.

[0065] Depend on Figure 2It is known that in each time slot, the scheduling agent set in the scheduling node uses its own configured large language model to generate a scheduling decision for the current time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer, with the goal of maximizing the reward, according to the preset scheduling prompt words. This determines the data stream to be scheduled in the current time slot, and stores the state, scheduling decision, and decision reward of all data streams corresponding to the current time slot as the decision information of the current time slot in the historical memory pool. Unlike the scheduling agent, the reflective agent set in the scheduling node selects the decision information of the 10 most recent consecutive time slots from the historical memory pool every 10 time slots. This allows its own configured large language model to generate new decision suggestions based on the decision information of the 10 time slots according to the preset reflective prompt words, and replaces the decision suggestions previously stored in the suggestion buffer with the newly generated decision suggestions.

[0066] The scheduling prompts for the scheduling agent follow a structured design, comprising five parts: Role, Workflow, Decision Process, Rules, and OutputFormat. The Role definition clarifies the large language model's role as a "network scheduling policy generator." The Workflow section describes the time slot characteristics and scheduling requirements of the wireless network, explaining that each time slot needs to select a data stream for transmission, and clarifies the selection criteria and objectives to determine the time constraints and basic process for the large language model to understand scheduling decisions. The Decision Process section guides the large language model in generating scheduling decisions through four steps: first, analyzing the urgency of each data stream; second, evaluating the packet arrival factor of each data stream to predict future network conditions; third, comparing the urgency and packet arrival factors of different data streams; and finally, selecting the data stream that maximizes rewards for scheduling. The Rules section defines that the scheduling decisions generated by the large language model must use a standardized JSON format and explicitly requires that the selected actions (scheduling decisions) be represented in integer form. The OutputFormat defines the structure of the standardized JSON format.

[0067] The reflection prompts for the reflective agent employ a three-stage reflection process to guide the large language model in generating decision suggestions. The specific three-stage reflection process is as follows: The first stage extracts temporal patterns, analyzing the decision information from the ten most recent consecutive time slots in the historical memory pool to identify recurring patterns in network traffic and scheduling decisions. The second stage evaluates decisions, identifying critical scheduling moments, paying particular attention to situations where data packets are approaching their latency deadlines, and analyzing the potential performance improvements from different scheduling choices. The third stage outputs decision suggestions in natural language based on the analysis results from the first and second stages. For example, when resource contention is detected, it suggests adjusting priorities based on channel quality and deadlines, or reserving resources in advance before periodic traffic peaks.

[0068] As can be seen from the foregoing embodiments, the real-time scheduling method for wireless networks based on a large language model proposed in this invention continuously analyzes the state of all data streams in the wireless network at each time slot, enabling the scheduling agent to generate scheduling decisions based on the state of all data streams and the decision suggestions stored in the suggestion buffer. Specifically, to adapt to various dynamic changes in the wireless network environment, the decision suggestions stored in the suggestion buffer are updated periodically. The update process is as follows: the reflective agent periodically selects decision information from multiple consecutive time slots from the historical memory pool, enabling its configured large language model to generate new decision suggestions based on the decision information from multiple time slots according to preset reflective prompts, and replaces the previously stored decision suggestions in the suggestion buffer with the newly generated decision suggestions.

[0069] Unlike existing technologies, the real-time scheduling method for wireless networks based on a large language model proposed in this invention continuously analyzes historical decision information and generates decision suggestions through a reflective agent set on the scheduling node. This enables the scheduling agent to generate scheduling decisions for each time slot under the guidance of continuously updated decision suggestions, combined with the state of all data streams analyzed in each time slot. As a result, the generated scheduling decisions can adapt to changes in the wireless network environment (such as changes in traffic patterns, channel condition fluctuations, and changes in the number of streams), solving the problem of insufficient adaptability of existing technologies when facing dynamic wireless network environments.

[0070] Based on the foregoing embodiments, the present invention also proposes a wireless network, which includes a scheduling node and multiple user nodes. The scheduling node is configured with a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer. The scheduling agent is configured with a large language model to generate scheduling decisions for each time slot. The historical memory pool is used to store the scheduling decisions for each time slot. The reflective agent is configured with a large language model to generate decision suggestions. The suggestion buffer is used to store decision suggestions. The wireless network is configured to implement real-time scheduling using the method described in the foregoing embodiments to determine the data flow scheduled by the wireless network in each time slot.

[0071] To better understand this invention, the following will be used... Figure 3 The wireless network shown is used as an example to illustrate how to implement data flow scheduling in a wireless network, in conjunction with the aforementioned embodiments.

[0072] Depend on Figure 3It is known that this wireless network adopts a centralized star network topology centered on the scheduling node (AP), connecting multiple client nodes / user nodes (such as client 1, client 2, and client 3). It supports multiple independent latency-sensitive data streams through downlinks (links between the scheduling node and clients), uplinks (links between clients and the scheduling node), and direct links (links between clients), with each data stream assigned to a specific link for transmission. For example, data stream 3 on client 3 is assigned to link 3, data stream 1 is assigned to link 2, and data stream 2 is assigned to link 1.

[0073] At each time slot, Figure 3 In the wireless network shown, the scheduling node first collects global information about the wireless network in the current time slot via a dedicated link and calculates the state of all data streams in the current time slot according to preset calculation rules. Then, the scheduling agent in the scheduling node uses its configured large language model to generate a scheduling decision for the current time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer, aiming to maximize the reward, according to preset scheduling prompts, to determine the data stream to be scheduled in the current time slot. For example, if it is determined that data stream 3 will be scheduled in time slot t, then the fastest expiring data packet in data stream 3 will be transmitted to the scheduling node (AP) via link 3 in time slot t.

[0074] To verify that the real-time scheduling method proposed in this invention has better performance than existing technologies, the inventors conducted various experiments for multi-dimensional comparisons, and obtained the following results: Figures 4-8 The experimental results are presented. In the experiments, existing scheduling strategies RAC, RL-Centralized, LDF, EDF, and RAC-Approx were used to compare the real-time scheduling method proposed in this invention with existing technologies. Timely throughput was used as an indicator to evaluate the performance of the existing technologies and the real-time scheduling method proposed in this invention. Timely throughput represents the ratio of the number of packets successfully reaching the destination node to the total number of packets generated in the network; a higher timely throughput indicates better scheduling performance.

[0075] in, Figure 4 Demonstrates the effectiveness of high-traffic network environments ( This paper compares the real-time throughput performance of the proposed real-time scheduling method with other scheduling strategies as the number of data streams increases (from 2 to 14). The proposed real-time scheduling method significantly outperforms scheduling strategies such as EDF, LDF, and RAC-Approx in dense traffic network environments, and its real-time throughput is close to the theoretical optimal value. This indicates that the proposed real-time scheduling method can deeply understand the network scenario and efficiently utilize resources for scheduling. In contrast, other scheduling strategies show a continuous decline in real-time throughput as the number of data streams increases (competition between data streams becomes more intense), especially when the number of data streams is ≥8, the RAC scheduling strategy completely fails due to excessive computational complexity.

[0076] Figure 5 Demonstrates the application of sparse traffic network environments ( This paper compares the real-time throughput performance of the proposed real-time scheduling method with other scheduling strategies as the number of data streams increases (from 2 to 14). The proposed real-time scheduling method significantly outperforms EDF, LDF, and RAC-Approx scheduling strategies in sparse traffic network environments. Furthermore, the real-time throughput (vertical axis) of the proposed real-time scheduling method consistently approaches the theoretical optimum, indicating that it can still deeply understand the network scenario and efficiently schedule resources in sparse environments. In contrast, other scheduling strategies show a continuous decline in real-time throughput as the number of data streams increases (competition between data streams becomes more intense). Especially when the number of data streams is ≥12, the RAC scheduling strategy completely fails due to excessive computational complexity.

[0077] Figure 6 Demonstrates the characteristics of dense traffic network environments ( The real-time scheduling method proposed in this invention is compared with other scheduling strategies in terms of timely throughput performance as the data flow latency increases (from 1 to 5). Figure 7 This demonstrates the sparse traffic network environment ( This invention compares the real-time scheduling method proposed in this paper with other scheduling strategies in terms of timely throughput performance as the data flow latency increases (from 1 to 5). Figure 6 and Figure 7 It can be seen that the real-time throughput (vertical axis) of the real-time scheduling method proposed in this invention is always close to the theoretical optimal and better than some scheduling strategies. Figure 6 and Figure 7This indicates that as the data stream delay period increases, the data packet survival time also increases accordingly, and the competition between data streams weakens. The real-time scheduling method proposed in this invention and other scheduling strategies have improved the timely throughput performance. Only when the data stream delay period is 1, the data packet survival time is short and the competition between data streams is more intense. At this time, the timely throughput performance of the LDF scheduling strategy differs greatly from other scheduling strategies.

[0078] Figure 8 This paper demonstrates the real-time throughput performance of the proposed real-time scheduling method compared to other scheduling strategies in a dynamic network environment. The evaluation includes four stages. Stage A comprises 10 data streams, each corresponding to a different quintuple representation, and each data stream corresponds to a quintuple. The values ​​of each parameter are independently and randomly drawn from a uniform distribution, where , , , as well as Phase B still contains 10 data streams, each corresponding to a different quintuple representation, and each data stream corresponds to a quintuple. The values ​​of each parameter in stage A conform to the random distribution constraints of stage A; the quintuples corresponding to each data stream in stage C are consistent with those in stage A, but an additional data stream is added, and the parameters in the quintuples corresponding to the added data stream conform to the random distribution constraints of stage A; the number of data streams in stage D satisfies k ∈ [2, 10], and each data stream corresponds to a quintuple. The values ​​of each parameter conform to the random distribution constraints of stage A.

[0079] Depend on Figure 8 As can be seen, the real-time scheduling method proposed in this invention performs excellently in each stage, stably achieving near-theoretical optimal timely throughput, while other scheduling strategies such as RAC-Approx, LDF, and EDF fail to converge to the optimal solution over long periods. Although the scheduling strategies RAC and RL-Centralized can achieve optimal timely throughput, they have certain limitations. For example, the RAC scheduling strategy has high computational complexity and requires recalculation to obtain the optimal scheduling strategy when the environment changes, which limits its practical application. For instance, in stage C, when a new data flow is added to the network, the RAC scheduling strategy requires nearly three hours of computation time to obtain the optimal scheduling scheme, which is unacceptable in practical wireless network applications. In contrast, although the RL-Centralized scheduling strategy can converge to the optimal solution in a short time, its training cost is high.

[0080] Furthermore, to verify that adding a reflection mechanism to the real-time scheduling method proposed in this invention can achieve better timely throughput performance, the inventors proposed performing data flow scheduling using both the real-time scheduling method proposed in this invention and the real-time scheduling method without the reflection mechanism in the same wireless network environment, and obtained the following results. Figure 9 The experimental results are shown.

[0081] Depend on Figure 9 It can be seen that the real-time scheduling method proposed in this invention achieves near-theoretical optimal throughput under different data stream quantities, and outperforms the real-time throughput performance of the real-time scheduling method without the reflection mechanism under different data stream quantities. This indicates that the reflection mechanism can transform historical time slot decision information into practical decision guidance to cope with complex and ever-changing network environments, achieving near-theoretical optimal scheduling performance.

[0082] The beneficial effects of this invention are as follows: a large language model is introduced to generate scheduling decisions for each time slot based on the state of all data streams and decision suggestions in each time slot, so as to realize data stream scheduling in each time slot; and a reflection mechanism is introduced to enable another large language model to generate new decision suggestions based on decision information from multiple time slots to guide the generation of decision suggestions, thus solving the problem of insufficient adaptability in the face of dynamic network environments.

[0083] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0084] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0085] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0086] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A real-time scheduling method for a wireless network based on a large language model, used to determine the data flow scheduled in each time slot of the wireless network, wherein the wireless network includes a scheduling node and multiple user nodes, characterized in that, The scheduling node is configured with a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer. The scheduling agent is equipped with a large language model to generate scheduling decisions for each time slot. The historical memory pool stores the scheduling decisions for each time slot. The reflective agent is equipped with a large language model to generate decision suggestions. The suggestion buffer stores the decision suggestions. The method includes performing the following steps in each time slot: Step S1: The scheduling node obtains the global information of the wireless network in the current time slot and calculates the state of all data streams in the current time slot according to the preset calculation rules; wherein, the global information of the wireless network includes the five-tuple corresponding to each data stream and the data packets contained in each data stream; the five-tuple corresponding to each data stream includes the data stream transmission start point, data stream period, data packet delay, data packet arrival rate and data packet transmission probability; Step S2: The scheduling agent uses its own configured large language model to generate a scheduling decision for the current time slot based on the state of all data streams in the current time slot and the decision suggestions stored in the suggestion buffer, with the goal of maximizing the reward, according to the preset scheduling prompt words, to determine the data stream to be scheduled in the current time slot, and uses the data packet transmission probability corresponding to the data stream as the decision reward for the current time slot; and stores the state, scheduling decision, and decision reward of all data streams corresponding to the current time slot as the decision information of the current time slot in the historical memory pool, so that the reflective agent uses its own configured large language model to generate decision suggestions based on the decision information of each time slot stored in the historical memory pool.

2. The method according to claim 1, characterized in that, The preset calculation rule is as follows: in, in, express Time slot data stream state, express Time slot data stream The degree of urgency, express Time slot data stream The arrival factor of the data packets. express Time slot data stream The number of data packets that have not expired. Represents data stream The Middle One incoming data packet, express Time slot data stream The Middle The remaining time before the arrival of each data packet expires. express Time slot data stream The Middle The expiration time of each arriving data packet. express Time slot data stream The Middle The arrival time of each arriving data packet. Represents data stream The corresponding data packet latency, Represents data stream The starting point of data stream transmission Represents data stream The corresponding data stream cycle.

3. The method according to claim 2, characterized in that, The method further includes: The reflective agent selects decision information from multiple time slots of the most recent consecutive preset values ​​from the historical memory pool at preset time intervals, so that its configured large language model generates new decision suggestions based on the decision information of multiple time slots according to preset reflective prompts, and replaces the decision suggestions previously stored in the suggestion buffer with the newly generated decision suggestions.

4. The method according to claim 3, characterized in that, The preset time interval is 10 time slots.

5. The method according to claim 4, characterized in that, The preset value is 10.

6. The method according to claim 5, characterized in that, The large language model configured for the scheduling agent and the reflective agent is GPT-4o.

7. A wireless network, comprising a scheduling node and multiple user nodes, characterized in that, The scheduling node is configured with a scheduling agent, a historical memory pool, a reflective agent, and a suggestion buffer. The scheduling agent is equipped with a large language model to generate scheduling decisions for each time slot. The historical memory pool is used to store the scheduling decisions for each time slot. The reflective agent is equipped with a large language model to generate decision suggestions. The suggestion buffer is used to store decision suggestions. The wireless network is configured to implement real-time scheduling using the method described in any one of claims 1-6 to determine the data stream scheduled by the wireless network in each time slot.

8. A computer-readable storage medium, characterized in that, It contains a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1-6.

9. An electronic device, characterized in that, include: One or more processors, and memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method of any one of claims 1-6 by executing the executable instructions.

Images