Distributed sender rate adaptation method and system using multiple agents

Abstract

This invention discloses a distributed transmitter rate adaptive adjustment method and system utilizing multiple agents, comprising: First, each transmitter node collects available bandwidth, buffer queue occupancy, and round-trip latency to construct a node state representation vector, and a local rate adjustment agent generates candidate transmission rates. Nodes receive neighbor rate adjustment experience messages, and through a multi-agent rate decision module, fuse neighbor action trajectories, state transitions, and throughput change gradients to obtain a cooperative rate adjustment offset. This offset dynamically corrects the candidate rates, generates a target transmission rate, and configures it to the transmission protocol stack for transmission.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for adaptive rate adjustment of a distributed transmitter using multiple agents. Background Technology

[0002] Existing methods for adjusting transmitter rates often rely on isolated decisions based on single-node network state awareness, lacking effective utilization of the collaborative effects among multiple transmitter nodes. As network applications scale up, multiple transmitters share bottleneck links, each independently adjusting its transmission rate based on locally available bandwidth, buffer occupancy, or latency information, easily leading to uncoordinated rate adjustment behavior. This lack of coordination can easily trigger vicious bandwidth contention between nodes, synchronous oscillations in transmission rates, and significant fluctuations in bottleneck link buffer queues when network load changes dynamically, resulting in decreased overall throughput, increased end-to-end latency, and low network resource utilization.

[0003] While rate coordination methods based on centralized control or global information interaction can improve the coordination effect among multiple nodes, they often require a central coordination node, which poses a single point of failure risk. Furthermore, the collection and distribution of global states result in high communication overhead and decision-making delays, making it difficult to meet the real-time requirements of large-scale distributed scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for adaptive rate adjustment of a distributed transmitter using multiple agents.

[0005] In a first aspect, embodiments of the present invention provide a distributed transmitter rate adaptive adjustment method utilizing multiple agents, the method comprising:

[0006] A set of network status awareness information from multiple sending nodes is obtained, wherein the set of network status awareness information includes the channel available bandwidth sequence, buffer queue occupancy sequence, and message round-trip delay sequence collected by each sending node;

[0007] By deploying local rate adjustment agents at each transmitting node, the available channel bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence are constructed into a node state representation vector. The rate adjustment strategy function inside the local rate adjustment agent is updated based on the node state representation vector to generate candidate transmission rate parameters for the current adjustment period.

[0008] Receive a neighbor rate adjustment experience message broadcast by a neighboring sending node. The neighbor rate adjustment experience message includes the rate adjustment action trajectory of the neighboring sending node in adjacent historical adjustment periods, the network state transition sequence corresponding to the rate adjustment action trajectory, and the throughput change gradient sequence triggered by the network state transition sequence.

[0009] The neighbor rate adjustment experience message and the candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period are input into the multi-agent rate decision module. The multi-agent rate decision module performs cooperative experience fusion processing on the rate adjustment action trajectory and the network state transition sequence to obtain the cooperative rate adjustment offset.

[0010] The candidate transmission rate parameters are dynamically corrected based on the cooperative rate adjustment offset to generate target transmission rate parameters, and the target transmission rate parameters are configured into the transmission protocol stack of the sending node to perform data transmission operations.

[0011] Secondly, embodiments of the present invention provide a distributed transmitter rate adaptive adjustment system utilizing multiple agents, comprising at least one service unit;

[0012] The service unit includes a storage unit and a computing unit; the storage unit is used to store program code; the computing unit is used to run the program code to execute the distributed transmitter rate adaptive adjustment method using multiple agents as described in the first aspect.

[0013] Compared with existing technologies, the beneficial effects provided by this invention include: The method and system for distributed transmitter rate adaptive adjustment using multiple agents, as disclosed in the embodiments of this invention, acquires a set of network state awareness information from multiple transmitter nodes, including channel available bandwidth sequences, buffer queue occupancy sequences, and message round-trip delay sequences collected by each transmitter node; through local rate adjustment agents deployed on each transmitter node, the above sequences are constructed into node state representation vectors and the rate adjustment strategy function is updated to generate candidate transmission rate parameters for the current adjustment period; neighbor rate adjustment experience messages broadcast by neighbor transmitter nodes are received, which contain the rate adjustment action trajectories of neighbor nodes in adjacent historical adjustment periods, the corresponding network state transition sequences, and throughput change gradient sequences; the neighbor rate adjustment experience messages and local candidate transmission rate parameters are input into a multi-agent rate decision module, and the rate adjustment action trajectories and network state transition sequences are subjected to collaborative experience fusion processing to obtain a collaborative rate adjustment offset; the candidate transmission rate parameters are dynamically corrected based on this offset to generate target transmission rate parameters and configure them to the transmission protocol stack of the transmitter node to execute data transmission. This invention utilizes a multi-agent distributed collaboration and experience fusion mechanism, enabling each sender to adaptively adjust its transmission rate based on local network conditions and neighbor adjustment experience, effectively improving network throughput and link utilization while reducing the probability of congestion and end-to-end transmission latency. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the steps of a distributed transmitter rate adaptive adjustment method utilizing multiple agents provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the process for obtaining network state awareness information set according to an embodiment of the present invention;

[0017] Figure 3 This is a schematic block diagram of a computer device structure provided for an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the distributed transmitter rate adaptive adjustment method using multiple agents provided in this embodiment of the disclosure. The following is a detailed description of this distributed transmitter rate adaptive adjustment method using multiple agents.

[0021] Step S201: Obtain a network status awareness information set of multiple sending nodes. The network status awareness information set includes the channel available bandwidth sequence, buffer queue occupancy sequence and message round-trip delay sequence collected by each sending node.

[0022] Step S202: By deploying local rate adjustment agents on each transmitting node, the available channel bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence are constructed into a node state representation vector. The rate adjustment strategy function inside the local rate adjustment agent is updated according to the node state representation vector to generate candidate transmission rate parameters in the current adjustment period.

[0023] Step S203: Receive a neighbor rate adjustment experience message broadcast by a neighboring sending node. The neighbor rate adjustment experience message includes the rate adjustment action trajectory of the neighboring sending node in adjacent historical adjustment periods, the network state transition sequence corresponding to the rate adjustment action trajectory, and the throughput change gradient sequence triggered by the network state transition sequence.

[0024] Step S204: Input the neighbor rate adjustment experience message and the candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period into the multi-agent rate decision module. The multi-agent rate decision module performs cooperative experience fusion processing on the rate adjustment action trajectory and the network state transition sequence to obtain the cooperative rate adjustment offset.

[0025] Step S205: Dynamically correct the candidate transmission rate parameters based on the cooperative rate adjustment offset to generate target transmission rate parameters, and configure the target transmission rate parameters to the transmission protocol stack of the sending node to perform data transmission operations.

[0026] In this embodiment of the invention, for example, the server is a sending server in a video live streaming, cloud gaming, or industrial data backhaul service, and multiple sending nodes are sending nodes deployed in different transmission processes within the same server cluster, edge server group, or the same physical server. Each sending node sends data packets to different receiving users or different transmission paths, and shares the same operator access network, wireless backhaul link, or data center egress link. The server executes a distributed sending rate adaptive adjustment process within each rate adjustment cycle, so that each sending node no longer relies solely on its own instantaneous measurement results to increase or decrease its rate, but instead combines the learning results of the local agent and the historical adjustment experience of neighboring sending nodes to generate a more stable target transmission rate that adapts to network changes.

[0027] The server first acquires a set of network status awareness information from multiple sending nodes. In a specific scenario, the server operates three sending nodes: Node A, Node B, and Node C. Node A is sending high-definition video segments to a mobile terminal, Node B is sending real-time control data to another receiving end, and Node C is transmitting cloud desktop image frames. At the beginning of each adjustment period, the server sends a sequence of probe messages to the corresponding receiving end through the channel probing module within each sending node, and reads the sending and receiving timestamps from the acknowledgment messages after they are returned. The server calculates a time interval parameter based on the arrival interval of consecutive acknowledgment messages, calculates an acknowledgment message arrival rate parameter based on the number of acknowledgment messages arriving per unit time, and then inputs these parameters into a bandwidth estimation filter to filter out instantaneous jitter, thus obtaining the available channel bandwidth sequence for each sending node. For node A, the bandwidth values ​​obtained by the server are 48Mbps, 46Mbps, 43Mbps and 41Mbps, indicating that the bandwidth of the path where node A is located is decreasing; for node B, the bandwidth sequence remains around 12Mbps, indicating that its path is relatively stable; for node C, the bandwidth sequence fluctuates between 30Mbps and 55Mbps, indicating that there is sudden contention traffic on its path.

[0028] The server simultaneously reads the current number of bytes occupied in the sending buffer queue of each sending node through the buffer queue monitoring module, forming a buffer queue occupancy sequence. Node A's sending buffer queue gradually increases from 2MB to 6MB, indicating that the current sending rate is higher than the actual carrying capacity of the path; Node B's buffer queue remains in a low range for a long time, indicating that its sending scheduling is relatively sufficient; Node C's buffer queue shows periodic increases and decreases, indicating a phased mismatch between its application layer data generation rate and the available bandwidth of the link. The server also calculates the packet round-trip time (RTT) measurement value through the latency measurement module, forming a packet RTT sequence. Node A's RTT increases from 38ms to 92ms, Node B remains around 20ms, and Node C varies between 35ms and 70ms. The server performs anomaly jump point processing, trend term extraction processing, and outlier filtering processing on the aforementioned channel available bandwidth sequence, buffer queue occupancy sequence, and message round-trip delay sequence, respectively. Abnormal values ​​caused by single acknowledgment message loss, short-term processing delays at the receiver, or system scheduling interruptions are removed. The processed three types of sequences are then combined into a normalized three-channel time series data structure, serving as a network state awareness information set. This set reflects both link carrying capacity and local transmission pressure and transmission delay changes, providing a unified input for subsequent agent decision-making.

[0029] Subsequently, the server, through local rate adjustment agents deployed at each sending node, constructs a node state representation vector from the available channel bandwidth sequence, buffer queue occupancy sequence, and message round-trip delay sequence. Based on this node state representation vector, the server updates the rate adjustment strategy function within the local rate adjustment agent, generating candidate transmission rate parameters for the current adjustment period. Specifically, the server extracts the three-channel sequence of node A within the current adjustment period from the network state awareness information set. It performs time-domain feature encoding on the available channel bandwidth sequence to obtain a channel state encoding vector representing the average bandwidth, descent slope, fluctuation amplitude, and short-term available capacity. It also performs time-domain feature encoding on the buffer queue occupancy sequence to obtain a buffer state encoding vector representing the queue occupancy ratio, queue growth rate, and remaining queue space. Finally, it performs time-domain feature encoding on the message round-trip delay sequence to obtain a delay state encoding vector representing the average delay, delay gradient, jitter amplitude, and a sustained increase indicator. The server concatenates these three types of encoded vectors into a node state representation vector for node A, enabling this vector to express the pre-congestion state of "bandwidth decrease, queue increase, and delay increase."

[0030] The server inputs the node's state representation vector into the rate adjustment strategy function of node A. The rate adjustment strategy function contains a set of action selection parameters and a set of value estimation parameters. The action selection parameter set outputs candidate transmission rate parameters based on the current state, and the value estimation parameter set outputs the estimated state value corresponding to continuing to use the current strategy in this state. In the node A scenario, the local rate adjustment agent adjusts the transmission rate from 50Mbps in the previous cycle to a candidate transmission rate of 42Mbps based on the trend of decreasing bandwidth and increasing latency. In the node B scenario, the agent recognizes that bandwidth, buffer, and latency are relatively stable, and maintains the transmission rate of 10Mbps in the previous cycle at 10.5Mbps. In the node C scenario, the agent cautiously increases the candidate transmission rate from 28Mbps to 32Mbps based on the state of bandwidth recovery but still fluctuating latency. After the actual transmission is completed, the server calculates the actual throughput of each node and updates the rate adjustment strategy function based on the state value estimate, allowing the local agent to gradually learn the differences in benefits from increasing, decreasing, or maintaining the transmission rate under different network conditions.

[0031] After the local candidate transmission rate is generated, the server receives neighbor rate adjustment experience messages broadcast by neighboring sending nodes. In a specific scenario, nodes A, B, and C each maintain their local neighbor node lists, and neighbor relationships are determined by network topology similarity, shared bottleneck links, the same edge data center exit, or wireless coverage area. Node B experienced a process of "rapid increase in buffer queue, synchronous increase in round-trip latency, and stable throughput after transmission rate reduction" within adjacent historical adjustment periods. The server sends a local rate adjustment experience sharing message to neighboring node A through node B's agent message broadcasting module. This message contains node B's rate adjustment action trajectory within consecutive adjustment periods, such as decreasing from 14Mbps to 12Mbps and then to 10.5Mbps; it also contains the network state transition sequence corresponding to the action trajectory, such as shifting from "slight decrease in bandwidth, moderate queue occupancy, and slight increase in latency" to "decrease in queue and stable latency"; and it also contains the throughput change gradient sequence triggered by network state transitions, such as a small decrease in throughput but reduced packet loss after the first rate reduction, and effective throughput recovering and remaining stable after the second rate reduction.

[0032] The server listens for broadcast messages through the agent message receiving module of node A, parses the sender node identifier, and matches and verifies this identifier against node A's local neighbor node list. When the sender node B exists in node A's neighbor node list, the server directly extracts node B's neighbor rate adjustment action trajectory, neighbor network state transition sequence, and neighbor throughput change gradient sequence. When a neighboring node is not recorded in the neighbor list but its broadcast signal strength exceeds a preset reception strength threshold, the server also includes it as a temporary reference neighbor in the reception range. The server performs time alignment, invalid field filtering, and normalization encapsulation processing on the extracted neighbor experience, combining them to form a neighbor rate adjustment experience message. For real-time business scenarios, the server also calculates the experience reference value score based on the experience freshness decay weight and the neighbor throughput change gradient sequence, removing outdated neighbor experiences and those with insignificant throughput improvement from the input of subsequent multi-agent rate decision modules to avoid outdated experiences interfering with current decisions.

[0033] Next, the server inputs the neighbor rate adjustment experience messages and the candidate transmission rate parameters generated by the local rate adjustment agent within the current adjustment period into the multi-agent rate decision module. This module performs collaborative experience fusion processing on the rate adjustment action trajectory and network state transition sequence to obtain the collaborative rate adjustment offset. Specifically, the server first inputs the neighbor rate adjustment action trajectory into the experience analysis submodule, extracting rate adjustment direction preference features and rate adjustment magnitude preference features from node B's continuous rate reduction actions. This identifies that node B tends to adopt moderate rate reduction rather than drastic rate reduction under similar congestion precursors. The server then inputs the neighbor network state transition sequence into the experience analysis submodule to mine state transition triggering condition patterns, identifying that "continuous increase in buffer queue occupancy and a positive packet round-trip delay gradient" is a crucial condition for node B to trigger rate reduction actions. Finally, the server inputs the neighbor throughput change gradient sequence into the experience analysis submodule to analyze the relationship between rate adjustment actions and throughput changes, obtaining the positive and negative effect mapping relationships for node B. The positive effect mapping relationship indicates that when the queue is close to overflowing and the latency continues to rise, moderately reducing the transmission rate can improve the effective throughput stability; the negative effect mapping relationship indicates that when the bandwidth is still sufficient and the queue occupancy is low, excessive speed reduction will cause insufficient link utilization.

[0034] The server inputs Node A's candidate transmission rate of 42Mbps into the local action coding submodule to generate a local rate adjustment action intent vector. This vector represents Node A's current intention to reduce the transmission rate and alleviate queue pressure. The collaborative attention calculation submodule calculates the action similarity attention weight between this action intent vector and Node B's rate adjustment direction preference feature, and also calculates the amplitude matching attention weight between this action intent vector and Node B's rate adjustment amplitude preference feature. Since Node A also currently experiences bandwidth reduction, queue increase, and latency increase, the server determines that Node A's state is highly similar to Node B's historical state. Therefore, it performs attention weighting to enhance the positive effect mapping relationship in Node B's experience and attention weighting to suppress the negative effect mapping relationship that causes excessive rate reduction and insufficient link utilization. Subsequently, the server performs state pattern matching between Node B's state transition triggering condition pattern and Node A's current node state representation vector, calculates the similarity distance between Node A's state and multiple neighbor state condition templates, and selects the best-matching neighbor state transition triggering condition as the local transferable experience fragment. The server inputs the local portable experience fragment into the cooperative offset generation submodule. Based on the positive effect mapping relationship, it generates a negative rate offset suggestion amount, such as suggesting a further reduction of 3Mbps from the local candidate rate of 42Mbps. Based on the negative effect mapping relationship, it generates a negative rate offset suppression amount, such as limiting the reduction to no more than 5Mbps. Finally, the two are vectorized to obtain a cooperative rate adjustment offset amount that includes the offset direction and offset magnitude, i.e., the direction is to reduce the rate and the magnitude is 3Mbps.

[0035] Finally, the server dynamically corrects the candidate transmission rate parameters based on the cooperative rate adjustment offset, generates the target transmission rate parameters, and configures the target transmission rate parameters into the transmission protocol stack of the sending node to execute the data transmission operation. In a specific scenario, the server obtains the cooperative rate adjustment offset of node A from the multi-agent rate decision module, parses out the offset direction identifier as negative and the offset magnitude value as 3Mbps, and then algebraically superimposes the candidate transmission rate of node A (42Mbps) with this offset magnitude in the negative direction to obtain an initial corrected transmission rate of 39Mbps. The server continues to read the latest observation value of the buffer queue occupancy sequence within the current adjustment period of node A, calculates the buffer queue free capacity and the degree of over-limit risk. The queue occupancy of node A is already close to the preset high water level. Based on this, the server generates a buffer protection adjustment factor. This factor has an inverse adjustment relationship with the degree of over-limit risk. When the queue is close to overflow, it outputs an additional negative correction amount, further reducing the initial corrected transmission rate to 37Mbps.

[0036] The server then reads the latest observations of the round-trip delay sequence of node A's packets and calculates the direction of delay gradient change. When multiple consecutive adjustment cycles show that the round-trip delay is still increasing, the server generates a delay gradient suppression coefficient and multiplies the buffer-aware corrected transmission rate with this suppression coefficient to obtain the delay-aware corrected transmission rate. For example, multiplying 37Mbps by 0.92 yields approximately 34Mbps. The server also obtains the maximum transmission unit limit parameter and the physical layer link rate limit parameter on the network path associated with node A, limiting the delay-aware corrected transmission rate between zero and the physical layer link rate limit to avoid negative rates or configuration values ​​exceeding link capacity. Simultaneously, based on the maximum transmission unit limit parameter, the server performs length constraints or fragmentation processing on the data packets to be transmitted, ensuring that the length of a single data packet does not exceed the allowed packet length range. After the above limiting and packet length processing, the server writes the target transmission rate parameter of node A into the transmission rate register in the transmission protocol stack congestion control module. The transmission protocol stack's transmission scheduling module reads the target transmission rate parameter from this register and controls the transmission interval between adjacent data packets accordingly, so that node A transmits video segment data at a rate not exceeding the target transmission rate.

[0037] Within the same adjustment period, the server performs the same process for nodes B and C. Node B, due to its stable local state and the absence of significant congestion risk from neighbor experience, has a target transmission rate close to its candidate transmission rate. Node C, due to bandwidth fluctuations, is handled with a slight increase strategy by the server, considering neighbor experience, buffer risk, and latency gradient, to avoid excessively increasing the transmission rate due to a sudden bandwidth rebound. The server also calculates a fluctuation variability parameter based on the rate correction magnitude between the candidate and target transmission rates, and adjusts the neighbor message broadcast radius and neighbor experience reception threshold accordingly. When the target transmission rates of multiple nodes remain stable over a long period and the fluctuation range is below the preset fluctuation convergence threshold, the server reduces the call frequency of the multi-agent rate decision module and updates the experience replay optimization parameter set using historical experience fragments in the local experience replay buffer during non-call periods. For multi-path transmission scenarios, the server instantiates a local rate adjustment agent and a multi-agent rate decision instance for each transmission path, generates a path-level target transmission rate for each path, and then generates a path quality score based on the latest round-trip delay for each path and allocates data packets to be sent. In the event of an unrecoverable failure of a path, the corresponding instance is frozen and the unsuccessfully sent data packets are reallocated. Through this process, the server can integrate local network status, neighbor agent experience, buffer protection, latency suppression, and link limiting constraints in a distributed transmission environment to form a stable, collaborative, and sustainably adaptive sender rate adjustment result.

[0038] Please refer to the following: Figure 2 , Figure 2This is a schematic diagram of the process for obtaining a network state awareness information set according to an embodiment of the present invention. In this embodiment of the present invention, the process of obtaining a network state awareness information set of multiple sending nodes can be implemented through the following example.

[0039] By deploying channel detection modules at each sending node, a sequence of detection messages is sent to the receiving node. Based on the receiving timestamp and sending timestamp carried in the acknowledgment message sequence returned by the receiving node, the time interval parameter and the arrival rate parameter of the acknowledgment message are calculated between adjacent acknowledgment messages.

[0040] The time interval parameter and the acknowledgment message arrival rate parameter are input into the bandwidth estimation filter for recursive filtering to generate a channel available bandwidth sequence within a continuous time window. The bandwidth estimation filter uses a weighted moving average method based on a time decay factor to iteratively update the bandwidth estimation value.

[0041] By deploying a buffer queue monitoring module on each sending node, the current number of bytes occupied in the data buffer queue in the transmission protocol stack is read according to a preset sampling period. The current number of bytes occupied collected in multiple consecutive sampling periods is arranged in time sequence to form a buffer queue occupancy sequence. The buffer queue occupancy sequence reflects the dynamic cumulative change trend of the amount of data to be sent by the sending node.

[0042] By deploying a latency measurement module at each sending node, a sending timestamp is marked for each sent data packet, and the receiving timestamp when the corresponding acknowledgment packet is received is recorded. The difference between the receiving timestamp and the sending timestamp is calculated as the round-trip latency measurement value of the data packet. The round-trip latency measurement values ​​corresponding to multiple consecutive data packets are arranged in chronological order to form a packet round-trip latency sequence.

[0043] The available channel bandwidth sequence is subjected to abnormal jump point identification processing. Abnormal observation points in the available channel bandwidth sequence that deviate from adjacent observations by more than a preset jump threshold are removed. Local linear interpolation is then performed using the normal observations adjacent to the abnormal observation points to fill the gaps, resulting in a smoothed available channel bandwidth sequence.

[0044] The buffer queue occupancy sequence is subjected to trend term extraction processing, separating the long-term trend component and the short-term fluctuation component in the buffer queue occupancy sequence. The long-term trend component is retained as a valid component of the buffer queue occupancy sequence, while the short-term fluctuation component is discarded.

[0045] Outlier filtering is performed on the message round-trip delay sequence. The local median absolute deviation of the message round-trip delay sequence is calculated. Message round-trip delay measurements that exceed a preset multiple of the local median absolute deviation are marked as outliers and removed from the message round-trip delay sequence to obtain the message round-trip delay sequence after outlier filtering.

[0046] The smoothed channel available bandwidth sequence, the buffer queue occupancy sequence after trend term extraction, and the message round-trip delay sequence after outlier filtering are combined into a time-domain aligned three-channel time series data structure.

[0047] The values ​​of the three channels corresponding to each sampling time in the three-channel time series data structure are normalized and mapped to a unified numerical range to obtain a normalized three-channel time series data structure.

[0048] The identifiers of each transmitting node in the normalized three-channel time series data structure are associated and stored with the corresponding normalized three-channel time series data to form the network state awareness information set.

[0049] The server, acting as the unified execution entity for multiple sending nodes, synchronously collects the link status of nodes A, B, and C. At the beginning of each adjustment period, the server controls the channel sounding module within each sending node to continuously send a sequence of sounding packets to the corresponding receiving node. Each sounding packet carries a sending timestamp, and the acknowledgment packet sequence returned by the receiving node carries both the receiving timestamp and the original sending timestamp. The server reads the return times of adjacent acknowledgment packets, calculates the time interval between adjacent acknowledgment packets, and counts the number of acknowledgment packets arriving per unit time to obtain the acknowledgment packet arrival rate parameter. For example, if the interval between consecutive acknowledgment packets received by node A during high-definition video transmission increases from 8ms to 15ms, the server identifies this as a slowdown in the acknowledgment return speed and a decreasing trend in the available bandwidth of the link.

[0050] The server inputs the time interval parameter and acknowledgment message arrival rate parameter into the bandwidth estimation filter, which then performs recursive filtering using a weighted moving average method based on a time decay factor. Newer acknowledgment message observations receive higher weights, while earlier observations have their impact reduced by the time decay factor, thus preventing single bursts of jitter from directly altering the bandwidth assessment results. The server then generates a sequence of available channel bandwidth within a continuous time window; for example, node A generates a sequence of 48Mbps, 46Mbps, 43Mbps, and 41Mbps, while node B generates a sequence with minor fluctuations around 12Mbps.

[0051] Meanwhile, the server, through the buffer queue monitoring module of each sending node, reads the current number of bytes occupied in the data buffer queue of the transmission protocol stack according to a preset sampling period, and arranges the occupancy values ​​of multiple consecutive sampling periods in chronological order to form a buffer queue occupancy sequence. The buffer occupancy of node A increases from 2MB to 6MB, and the server records that the video segments to be sent are dynamically accumulating, indicating that the current transmission pressure is higher than the link's processing capacity. The server also uses a latency measurement module to mark the sending timestamp of each sent data packet and records the receiving timestamp when the corresponding acknowledgment packet is received. The difference between the two is used as the round-trip latency measurement value, and the measurement values ​​of multiple consecutive packets are arranged in chronological order to form a packet round-trip latency sequence.

[0052] The server then cleans and reshapes the three types of sequences. For the available bandwidth sequence, the server identifies anomalous jump points where the deviation from adjacent observations exceeds a preset jump threshold, such as a momentary 1Mbps anomaly caused by a single acknowledgment message delay. It then uses normal observations before and after the anomaly point for local linear interpolation to fill the gap, resulting in a smoothed bandwidth sequence. For the buffer queue occupancy sequence, the server separates the long-term trend component from the short-term fluctuation component, retaining the long-term trend that reflects the continuous accumulation or release of the queue, and discarding short-term spikes caused by instantaneous writes from the application layer. For the message round-trip delay sequence, the server calculates the local median absolute deviation, marking delay measurements exceeding a preset multiple as outliers and removing them, resulting in a more stable delay sequence.

[0053] After preprocessing, the server aligns the smoothed channel available bandwidth sequence, the buffer queue occupancy sequence after trend term extraction, and the message round-trip delay sequence after outlier filtering in the time domain according to the same sampling time, combining them into a three-channel time series data structure. The server further normalizes the bandwidth, queue occupancy, and delay values ​​at each sampling time, mapping different units such as Mbps, bytes, and milliseconds to the same numerical range. Finally, the server associates the identifiers of nodes A, B, and C with their corresponding normalized three-channel time series data, forming a network state awareness information set. This set is used by the subsequent local rate adjustment agent to construct node state representation vectors and generate candidate transmission rate parameters.

[0054] In this embodiment of the invention, the local rate adjustment agent deployed at each transmitting node constructs the available channel bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence into a node state representation vector, and updates the rate adjustment strategy function inside the local rate adjustment agent according to the node state representation vector to generate candidate transmission rate parameters for the current adjustment period. This can be implemented through the following example.

[0055] The available channel bandwidth sequence of the corresponding transmitting node in the current adjustment period is extracted from the network state awareness information set. The available channel bandwidth sequence is subjected to time-domain feature encoding processing to extract the mean trend feature, variance fluctuation feature and instantaneous rate of change feature of the channel bandwidth. The mean trend feature, variance fluctuation feature and instantaneous rate of change feature are concatenated into a channel state encoding vector.

[0056] Extract the buffer queue occupancy sequence of the corresponding sending node in the current adjustment period from the network state awareness information set, perform time-domain feature encoding processing on the buffer queue occupancy sequence, extract the cumulative fill degree feature, consumption rate feature and queue length autocorrelation feature of the buffer queue, and concatenate the cumulative fill degree feature, the consumption rate feature and the queue length autocorrelation feature into a buffer state encoding vector.

[0057] The round-trip delay sequence of the corresponding sending node within the current adjustment period is extracted from the network state awareness information set. The round-trip delay sequence is processed by time-domain feature encoding to extract the delay level feature, delay jitter amplitude feature and delay growth persistence feature of the round-trip delay. The delay level feature, the delay jitter amplitude feature and the delay growth persistence feature are concatenated into a delay state encoding vector.

[0058] The channel state coding vector, the buffer state coding vector, and the delay state coding vector are concatenated along the feature dimension to generate the node state representation vector of the transmitting node in the current adjustment period.

[0059] Obtain the rate adjustment strategy function stored by the local rate adjustment agent in the previous adjustment cycle. The rate adjustment strategy function is a parameterized mapping relationship with node state representation vector as input and rate adjustment action as output. The rate adjustment strategy function includes a value estimation parameter set and an action selection parameter set.

[0060] The node state representation vector within the current adjustment period is input into the rate adjustment strategy function. The action dominance score corresponding to each candidate rate adjustment action is calculated through the action selection parameter set. Based on the action dominance score, candidate transmission rate parameters within the current adjustment period are generated.

[0061] The node state representation vector within the current adjustment period is input into the rate adjustment strategy function, and the state value estimate corresponding to the node state representation vector is calculated through the value estimation parameter set. The state value estimate is used to represent the expected cumulative throughput gain under the current network state.

[0062] After the current adjustment period ends, the actual throughput value within the current adjustment period is obtained by the throughput monitoring module deployed on the sending node. The value estimation loss gradient is calculated based on the time difference error between the actual throughput value and the state value estimate.

[0063] The value estimation parameter set is updated by backpropagation based on the value estimation loss gradient to obtain the updated value estimation parameter set. The action selection parameter set is then updated by policy gradient ascent based on the actual throughput value to obtain the updated action selection parameter set.

[0064] The updated set of value estimation parameters and the updated set of action selection parameters are combined into an updated rate adjustment policy function, and the updated rate adjustment policy function is stored in the local rate adjustment agent for use in the next adjustment cycle.

[0065] In this embodiment of the invention, for example, after the server completes the construction of the network state awareness information set, it calls the local rate adjustment agent for each sending node. Taking node A as an example, the server extracts the available channel bandwidth sequence of node A within the current adjustment period from the network state awareness information set, for example, the bandwidth value decreases from 48Mbps to 41Mbps within a continuous time window. The server performs time-domain feature encoding on this sequence, calculates the mean bandwidth and the decreasing trend to obtain the mean trend feature; calculates the degree of deviation of the bandwidth value from the mean to obtain the variance fluctuation feature; and calculates the slope of change between adjacent sampling points to obtain the instantaneous rate of change feature. The server concatenates the three types of features into a channel state encoding vector, so that the vector can express that the available bandwidth of node A's current link is shrinking and has a continuous decreasing trend.

[0066] The server continues to extract the buffer queue occupancy sequence of node A from the network state awareness information set, for example, the buffer occupancy gradually increases from 2MB to 6MB. The server performs temporal feature encoding on this sequence, calculates the overall occupancy ratio of the queue within the current window to obtain the cumulative fill rate feature; calculates the consumption rate feature based on the decreasing segment of the queue occupancy value and the amount of data sent out of the queue; and calculates the queue length autocorrelation feature based on the correlation between queue lengths within consecutive sampling periods. The server concatenates the above features into a buffer state encoding vector, which is used to represent that the data to be sent by node A is continuously accumulating, the queue release rate is insufficient, and the queue growth is continuous.

[0067] The server then extracts the round-trip time (RTT) sequence of messages from node A, for example, if the RTT increases from 38ms to 92ms. The server encodes this sequence, extracting the average delay and high-order delay to form a delay level feature, extracting the delay difference between adjacent messages to form a delay jitter amplitude feature, and extracting whether multiple consecutive sampling points maintain a positive increase to form a delay growth persistence feature. The server concatenates these three types of delay features into a delay state encoding vector. Subsequently, the server concatenates the channel state encoding vector, buffer state encoding vector, and delay state encoding vector along the feature dimensions to generate a node state representation vector for node A within the current adjustment period. This vector fully describes the pre-congestion state of "bandwidth decrease, queue accumulation, and continuously increasing delay."

[0068] In one optional implementation, the node state representation vector can be constructed using a fixed field order. Specifically, the node state representation vector includes at least a bandwidth state field, a buffer state field, and a latency state field. The bandwidth state field includes the average available bandwidth, bandwidth change trend, and bandwidth fluctuation level within the current adjustment period; the buffer state field includes the buffer queue occupancy ratio, buffer queue growth trend, and remaining buffer queue capacity; and the latency state field includes the average round-trip latency, round-trip latency change trend, and round-trip latency jitter level. When generating the node state representation vector, the server first calculates the above fields according to a preset sampling window, and then arranges the fields in a preset order to form a vector structure of the same length. For different sending nodes, the same field order and the same normalization range are used, enabling the state representations of different nodes to be uniformly identified and compared by the local rate adjustment agent and the multi-agent rate decision module. For example, the node state representation vector can sequentially include: average bandwidth, bandwidth decrease flag, bandwidth fluctuation value, queue occupancy ratio, queue growth flag, remaining queue ratio, average latency, latency increase flag, and latency jitter value. The flag values ​​in the above fields can be represented by 0 or 1, or by normalized continuous numerical values.

[0069] The server retrieves the rate adjustment policy function stored by the local rate adjustment agent of node A in the previous adjustment cycle. This policy function takes the node state representation vector as input and the rate adjustment action as output, and internally includes a set of action selection parameters and a set of value estimation parameters. The server inputs the current node state representation vector into this policy function and calculates the action advantage scores of candidate actions such as "increase rate", "maintain rate", "small decrease rate", and "large decrease rate" using the action selection parameter set. Since node A is currently experiencing queue accumulation and latency growth, the small decrease rate action receives the highest score. Based on this, the server adjusts the transmission rate from 50 Mbps in the previous cycle to the candidate transmission rate parameter of 42 Mbps in the current adjustment cycle.

[0070] In one optional implementation, the local rate adjustment agent generates candidate transmission rate parameters using a preset action set. The preset action set may include significant rate reduction, slight rate reduction, rate maintenance, slight rate increase, and significant rate increase. Each action corresponds to a rate adjustment ratio or rate adjustment step size. For example, significant rate reduction corresponds to reducing the transmission rate of the previous adjustment cycle by a preset large ratio, slight rate reduction corresponds to reducing the transmission rate of the previous adjustment cycle by a preset small ratio, rate maintenance corresponds to maintaining the transmission rate of the previous adjustment cycle, and slight and significant rate increases correspond to different magnitudes of rate increases. The local rate adjustment agent calculates the action dominance score for each rate adjustment action based on the node state representation vector and selects the rate adjustment action with the highest score as the candidate action for the current adjustment cycle. The server then obtains the candidate transmission rate parameters based on the transmission rate used in the previous adjustment cycle and the adjustment ratio or adjustment step size corresponding to the candidate action. For example, when the transmission rate in the previous adjustment period was 50Mbps and the candidate action with the highest score was a slight decrease in speed, the server can adjust the transmission rate to 42Mbps according to a preset slight decrease in speed ratio; when the candidate action with the highest score was to maintain the rate, the server sets the candidate transmission rate parameter to a value close to the transmission rate in the previous adjustment period; when the candidate action with the highest score was to slightly increase the speed, the server adds a preset step size to the transmission rate in the previous adjustment period to obtain the candidate transmission rate parameter.

[0071] Simultaneously, the server calculates the state value estimate corresponding to the current node's state representation vector using the value estimation parameter set. This value represents the expected cumulative throughput gain that can be obtained by continuing to adjust according to the existing strategy under the current network state. After the current adjustment cycle ends, the server uses the throughput monitoring module of node A to count the actual throughput value and compares it with the state value estimate to calculate the time-series difference error. The server uses this error to form a value estimation loss gradient and performs gradient backpropagation update on the value estimation parameter set to make the value estimation result closer to the actual throughput performance. At the same time, the server performs policy gradient ascent update on the action selection parameter set based on the actual throughput value to increase the probability of selecting actions that effectively improve throughput. Finally, the server recombines the updated value estimation parameter set and action selection parameter set into an updated rate adjustment policy function and stores it in the local rate adjustment agent of node A for continued use in the next adjustment cycle.

[0072] In this embodiment of the invention, receiving the neighbor rate adjustment experience message broadcast by the neighbor sending node can be implemented through the following example.

[0073] Each sending node maintains a local neighbor node list, which includes node identifiers and communication address information of other sending nodes that share a portion of the communication link with the sending node or are located in the same network topology area.

[0074] The agent message broadcasting module deployed on each sending node sends a local rate adjustment experience sharing message to all neighbor sending nodes in the local neighbor node list at a preset broadcast period. The local rate adjustment experience sharing message includes the rate adjustment action trajectory of the sending node in the most recent consecutive adjustment periods, the network state transition sequence corresponding to the rate adjustment action trajectory, and the throughput change gradient sequence triggered by the network state transition sequence.

[0075] The intelligent agent message receiving module deployed on each sending node listens for broadcast messages from other sending nodes. When a broadcast message is received from a neighboring sending node, the sending source node identifier of the broadcast message is parsed, and the sending source node identifier is matched and verified with the local neighbor node list.

[0076] If the source node identifier exists in the local neighbor node list, extract the neighbor rate adjustment action trajectory, neighbor network state transition sequence, and neighbor throughput change gradient sequence carried in the broadcast message;

[0077] If the source node identifier does not exist in the local neighbor node list, determine whether the signal strength of the broadcast message exceeds a preset reception strength threshold. If it exceeds the preset reception strength threshold, add the source node identifier to the local neighbor node list and extract the neighbor rate adjustment action trajectory, neighbor network state transition sequence and neighbor throughput change gradient sequence carried in the broadcast message.

[0078] Based on the trajectory length parameter of the neighbor rate adjustment action trajectory, the neighbor rate adjustment action trajectory is recursively aligned in time to align the neighbor rate adjustment action trajectories with different start timestamps broadcast by different neighbor sending nodes to a unified time reference benchmark.

[0079] The neighbor rate adjustment action trajectory after time recursive alignment is processed by action space mapping, which maps the different discretization granularities or different numerical ranges of rate adjustment action trajectories that may be used by each neighbor sending node to the unified action space used by the local rate adjustment agent of the sending node.

[0080] Based on the state difference parameter between adjacent network states in the neighbor network state transition sequence, redundant state transition pairs with a state difference lower than a preset difference threshold are removed from the neighbor network state transition sequence to obtain a simplified neighbor network state transition sequence.

[0081] The throughput change gradient values ​​in the neighbor throughput change gradient sequence are matched with the time index of the simplified neighbor network state transition sequence, and time period data without corresponding throughput change gradient values ​​are removed.

[0082] The neighbor rate adjustment action trajectory processed by action space mapping, the simplified neighbor network state transition sequence, and the neighbor throughput change gradient sequence processed by time index matching are combined and encapsulated into the neighbor rate adjustment experience message.

[0083] In this embodiment of the invention, for example, the server maintains a local neighbor node list for each sending node. Taking node A as an example, the server records the node identifiers, communication address information, and shared link markers of nodes B and C in node A's neighbor node list. Node B and node A share the same data center egress link, and node C and node A are located in the same edge network topology area. The server uses this neighbor node list to limit the scope of experience interaction, allowing node A to preferentially receive rate adjustment experience that is highly relevant to its own network environment.

[0084] The server deploys an agent message broadcasting module on each sending node. Node B sends a local rate adjustment experience sharing message to Node A and Node C in each preset broadcast period. This message records the rate adjustment action trajectory of Node B in the most recent consecutive adjustment periods, such as "14Mbps to 12Mbps, 12Mbps to 10.5Mbps, 10.5Mbps unchanged"; at the same time, it records the network state transition sequence corresponding to the action trajectory, such as "buffer usage increases and latency increases" changing to "buffer usage decreases and latency stabilizes"; it also records the throughput change gradient sequence triggered by the above state transitions, such as the effective throughput first decreasing slightly and then stabilizing and resuming growth.

[0085] The server continuously listens for broadcast messages from other sending nodes through the agent message receiving module of node A. When node A receives a broadcast message from node B, the server parses the source node identifier in the broadcast message and matches and verifies node B's identifier against node A's local neighbor node list. Since node B already exists in node A's neighbor node list, the server directly extracts the neighbor rate adjustment action trajectory, neighbor network state transition sequence, and neighbor throughput change gradient sequence carried in the broadcast message. For another newly emerging node D, after node A receives node D's broadcast message, the server identifies that node D does not yet exist in its local neighbor node list and then reads the signal strength of the broadcast message. When the signal strength of node D's broadcast message exceeds a preset received strength threshold, the server adds node D's node identifier and communication address information to node A's local neighbor node list and extracts the experience data carried by node D, enabling node A to incorporate new neighbor sender experience.

[0086] After extracting neighbor experience, the server performs unified processing on the action trajectories of different neighbor sending nodes. Since the start timestamps of the broadcast messages from nodes B and D are different, and their trajectory lengths may also differ, the server performs time recursive alignment processing based on the trajectory length parameter of the action trajectory according to the neighbor rate, uniformly mapping the historical adjustment cycles of different neighbors to the same time reference base before node A's current adjustment cycle. Subsequently, the server performs action space mapping processing on the aligned action trajectories. For example, if node B uses a continuous adjustment granularity of "reducing by 1.5 Mbps each time," and node D uses a discrete adjustment granularity of "reducing by 10%, maintaining, and increasing by 10%," the server uniformly maps both to the standard action space used by node A's local agent, such as "increasing rate, maintaining rate, slightly decreasing rate, and significantly decreasing rate."

[0087] The server further processes the neighbor network state transition sequence. For the state sequence broadcast by node B, the server calculates the state difference parameter between adjacent network states and removes redundant state transition pairs with a difference lower than a preset threshold. For example, if bandwidth, queue size, and latency remain almost unchanged over two consecutive periods, the server deletes the duplicate state segment, retaining only state transitions that reflect key changes such as "queue size continuously increasing" and "latency starting to decrease." Then, the server matches the neighbor throughput change gradient sequence with the time index of the simplified neighbor network state transition sequence, removing throughput gradient data without corresponding state transitions to ensure a one-to-one correspondence between actions, states, and throughput changes.

[0088] Finally, the server combines and encapsulates the neighbor rate adjustment action trajectory (processed by action space mapping), the simplified neighbor network state transition sequence, and the neighbor throughput change gradient sequence (processed by time index matching) into a neighbor rate adjustment experience message. This message is then input into the multi-agent rate decision module of node A to determine whether the adjustment actions of neighboring nodes in similar network states can be transferred to the candidate transmission rate correction process of node A's current adjustment period.

[0089] In this embodiment of the invention, the step of inputting the neighbor rate adjustment experience message and the candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period into the multi-agent rate decision module, and then performing collaborative experience fusion processing on the rate adjustment action trajectory and the network state transition sequence through the multi-agent rate decision module to obtain the collaborative rate adjustment offset can be implemented through the following example.

[0090] The neighbor rate adjustment action trajectory contained in the neighbor rate adjustment experience message is input into the experience parsing submodule of the multi-agent rate decision module. The trajectory feature extraction processing is performed on the neighbor rate adjustment action trajectory to extract the rate adjustment direction preference feature and rate adjustment amplitude preference feature of the neighbor sending node in the continuous adjustment period.

[0091] The neighbor network state transition sequence contained in the neighbor rate adjustment experience message is input into the experience parsing submodule. The state transition pattern mining process is performed on the neighbor network state transition sequence to extract the state transition triggering condition pattern of the neighbor sending node triggering rate adjustment operation under different network state conditions.

[0092] The neighbor throughput change gradient sequence contained in the neighbor rate adjustment experience message is input into the experience parsing submodule. The gradient effect analysis processing of the neighbor throughput change gradient sequence is performed to extract the positive effect mapping relationship and negative effect mapping relationship between the rate adjustment action trajectory of the neighbor sending node and the throughput change.

[0093] The candidate transmission rate parameters generated by the local rate adjustment agent within the current adjustment period are obtained, and the candidate transmission rate parameters are input into the local action encoding submodule of the multi-agent rate decision module. The candidate transmission rate parameters are then processed by action intent encoding to generate a local rate adjustment action intent vector.

[0094] The collaborative attention calculation submodule of the multi-agent rate decision module calculates the action similarity attention weight between the local rate adjustment action intention vector and the rate adjustment direction preference feature, and calculates the amplitude matching degree attention weight between the local rate adjustment action intention vector and the rate adjustment amplitude preference feature.

[0095] The collaborative attention calculation submodule combines the action similarity attention weight and the amplitude matching attention weight to apply to the positive effect mapping relationship and the negative effect mapping relationship. It performs attention weighting enhancement processing on the positive effect mapping relationship and attention weighting suppression processing on the negative effect mapping relationship to obtain the effect mapping set after attention regulation.

[0096] The state transition triggering condition pattern is matched with the node state representation vector generated by the local rate regulation agent in the current regulation period, and the similarity distance between the node state representation vector and each state condition template in the state transition triggering condition pattern is calculated.

[0097] Based on the similarity distance, determine the neighbor state transition triggering condition that best matches the node state representation vector, and extract the attention-modulated effect mapping relationship corresponding to the best-matching neighbor state transition triggering condition as a local transferable experience fragment.

[0098] The local transferable experience fragment is input into the collaborative offset generation submodule of the multi-agent rate decision module. A positive rate offset proposal is generated based on the positive effect mapping relationship in the local transferable experience fragment, and a negative rate offset suppression is generated based on the negative effect mapping relationship in the local transferable experience fragment.

[0099] The positive rate offset proposal and the negative rate offset suppression are vector-synthesized to obtain a coordinated rate adjustment offset that includes the offset direction and offset magnitude.

[0100] In this embodiment of the invention, for example, after node A has generated a candidate transmission rate parameter of 42 Mbps, the server inputs the neighbor rate adjustment experience messages broadcast by nodes B and D into the multi-agent rate decision module corresponding to node A. The server first reads the neighbor rate adjustment action trajectories through the experience parsing submodule. For example, node B has continuously executed the action of "small speed reduction, small speed reduction, and maintaining the rate" in the last three adjustment cycles, while node D has executed the action of "maintaining the rate and small speed reduction". The server extracts features from these trajectories to obtain the directional preference features of neighboring sending nodes in the pre-congestion state that favor reducing the transmission rate, and further calculates the proportion of each speed reduction to the original transmission rate to form the rate adjustment amplitude preference features.

[0101] The server continues to input the neighbor network state transition sequence into the experience analysis submodule to mine the state transition patterns that trigger adjustment operations by neighbor nodes. Node B's experience shows that when the available channel bandwidth continuously decreases, buffer queue occupancy continuously increases, and the packet round-trip delay gradient is positive, its local agent triggers a rate-down operation; when queue occupancy begins to decrease and delay stabilizes, its local agent stops further rate-down. The server thus extracts the state transition triggering condition pattern of "bandwidth decrease, queue occupancy increase, delay increase" corresponding to a small rate-down. Subsequently, the server inputs the neighbor throughput change gradient sequence into the experience analysis submodule to analyze the relationship between adjustment actions and throughput changes. After a small rate-down, Node B's effective throughput fluctuations decrease, packet loss decreases, and it gradually returns to stability; the server establishes this part as a positive effect mapping relationship. Node D, with excessive rate-down when there is still spare capacity on the link, leads to a decrease in effective throughput; the server establishes this part as a negative effect mapping relationship.

[0102] In one optional implementation, the positive effect mapping relationship is used to record the beneficial results brought about by a neighboring sending node performing a certain rate adjustment action under specific network conditions. The beneficial results include at least one of increased effective throughput, reduced throughput fluctuation, decreased buffer queue occupancy, decreased round-trip latency, or reduced packet loss. The negative effect mapping relationship is used to record the adverse results brought about by a neighboring sending node performing a certain rate adjustment action under specific network conditions. The adverse results include at least one of decreased effective throughput, decreased link utilization, continued increase in buffer queue occupancy, continued increase in round-trip latency, or increased packet loss. When constructing the effect mapping relationship, the server associates and stores the state conditions in the neighboring network state transition sequence, the action type in the neighboring rate adjustment action trajectory, and the result changes in the neighboring throughput change gradient sequence. For example, when a neighboring node performs a small rate reduction action in a state of "increased queue occupancy and continuously increasing latency," and subsequently the effective throughput stabilizes and the latency decreases, the server treats this record as a positive effect mapping relationship; when a neighboring node performs a large rate reduction action in a state of "sufficient bandwidth and low queue occupancy," and subsequently the effective throughput decreases, the server treats this record as a negative effect mapping relationship.

[0103] The server obtains the candidate transmission rate parameter of 42Mbps for node A in the current adjustment period and inputs it into the local action encoding submodule. This submodule, combined with node A's transmission rate of 50Mbps in the previous period, encodes 42Mbps as a local rate adjustment action intent vector of "significantly reduced speed but not achieving strong suppression." The collaborative attention calculation submodule then calculates the action similarity attention weight between this action intent vector and the neighbor's direction preference features, and calculates the amplitude matching attention weight between this action intent vector and the neighbor's amplitude preference features. Since node A's current action direction is consistent with node B's historical successful action direction, and the speed reduction magnitude is similar, the server assigns node B a higher empirical attention weight; since node D has a record of excessive speed reduction leading to a decrease in throughput, the server retains a negative empirical suppression effect on it.

[0104] In one optional implementation, the action similarity attention weight is determined based on whether the local rate adjustment action intention is consistent with the neighbor's rate adjustment direction. When the local rate adjustment action intention is to decrease speed, and the neighbor's historical action trajectory also shows a decrease in speed under similar conditions, the action similarity attention weight is set to a higher value; when the local rate adjustment action intention is to increase speed, and the neighbor's historical action trajectory shows a decrease or maintenance of speed, the action similarity attention weight is set to a lower value. The amplitude matching attention weight is determined based on the closeness between the local candidate rate adjustment amplitude and the neighbor's historical rate adjustment amplitude. When the local candidate rate adjustment amplitude and the neighbor's historical effective adjustment amplitude are within the same preset amplitude range, the amplitude matching attention weight is set to a higher value; when the difference between the two exceeds a preset amplitude difference threshold, the amplitude matching attention weight is set to a lower value. The collaborative attention calculation submodule merges the action similarity attention weight and the amplitude matching attention weight into a neighbor experience adoption weight. The higher the neighbor experience adoption weight, the more suitable the neighbor experience is for correcting the local candidate transmission rate parameter; the lower the neighbor experience adoption weight, the lower the reference value of the neighbor experience for the current adjustment cycle.

[0105] The server combines action similarity attention weights and amplitude matching attention weights for both positive and negative effect mappings. It weights and amplifies the positive experience of a small decrease in speed at node B, and weights and suppresses the negative experience of an excessive decrease in speed at node D, resulting in a set of attention-modulated effect mappings. Next, the server performs state pattern matching between the neighbor state transition triggering condition patterns and the current node state representation vector of node A, calculating the similarity distance between node A's state and each neighbor's state template. Since node A currently exhibits decreased bandwidth, increased queue size, and continuously increasing latency, the server determines node B's congestion precursor template as the best-matching neighbor state transition triggering condition and extracts its corresponding attention-modulated effect mapping as a locally transferable experience fragment.

[0106] The server inputs the locally transferable experience fragment into the collaborative offset generation submodule. Based on the positive effect mapping relationship, it generates a positive rate offset suggestion to further reduce the transmission rate, for example, suggesting a further reduction of 3Mbps from 42Mbps. Simultaneously, based on the negative effect mapping relationship, it generates a negative rate offset suppression amount, for example, limiting the total reduction to no more than a certain percentage of the candidate rate in this cycle, to prevent node A from wasting available bandwidth due to excessive conservatism. Finally, the server performs vector synthesis of the positive rate offset suggestion and the negative rate offset suppression amount to obtain a collaborative rate adjustment offset that includes the offset direction and offset magnitude, i.e., the offset direction is to reduce the transmission rate and the offset magnitude is 3Mbps, which is used for subsequent dynamic correction of the candidate transmission rate of node A.

[0107] In one optional implementation, the collaborative offset generation submodule determines the offset direction and offset magnitude based on local portable experience fragments. If the positive effect mapping relationship in the local portable experience fragment indicates that neighboring nodes achieve more stable throughput after reducing their sending rate under similar conditions, the collaborative offset generation submodule determines the offset direction to decrease the rate; if the local portable experience fragment indicates that neighboring nodes increase their effective throughput and latency does not significantly deteriorate after increasing their sending rate under similar conditions, the offset direction is determined to increase the rate; if the positive and negative effects are close, the offset direction is determined to remain the same or make a slight adjustment. The offset magnitude can be jointly determined based on the neighbor's historical effective adjustment magnitude, the local candidate sending rate, and the experience adoption weight. When the experience adoption weight is high, the offset magnitude can be close to the neighbor's historical effective adjustment magnitude; when the experience adoption weight is low, the offset magnitude can be lower than the neighbor's historical effective adjustment magnitude. To avoid the collaborative offset being too large, the server can also set a maximum offset magnitude per cycle to ensure that the collaborative offset does not exceed a preset proportion of the candidate sending rate parameter. For example, if node A's candidate transmission rate parameter is 42Mbps, and neighbor node B achieves a stable effect by reducing its transmission rate by about 3Mbps under similar conditions, and the neighbor's experience adoption weight is relatively high, then the cooperative offset generation submodule can generate a cooperative rate adjustment offset with a direction of decreasing rate and an amplitude of 3Mbps.

[0108] In this embodiment of the invention, the step of dynamically correcting the candidate transmission rate parameters based on the cooperative rate adjustment offset to generate target transmission rate parameters, and configuring the target transmission rate parameters to the transmission protocol stack of the sending node to perform data transmission operations, can be implemented through the following example.

[0109] The coordinated rate adjustment offset is obtained from the multi-agent rate decision module, and the offset direction identifier and offset magnitude value contained in the coordinated rate adjustment offset are parsed.

[0110] The candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period are obtained, and the candidate transmission rate parameters and the offset amplitude value are algebraically superimposed in the direction indicated by the offset direction identifier to generate preliminary corrected transmission rate parameters.

[0111] Obtain the latest observation value of the buffer queue occupancy sequence of the sending node in the current adjustment period, and determine the spare capacity of the buffer queue and the degree of risk of the buffer queue exceeding the limit based on the latest observation value;

[0112] The buffer protection adjustment factor is calculated based on the over-limit risk level of the buffer queue. The buffer protection adjustment factor has an inverse adjustment relationship with the over-limit risk level. When the over-limit risk level indicates that the buffer queue is close to the overflow state, the buffer protection adjustment factor outputs a negative correction amount to reduce the initial correction transmission rate parameter.

[0113] The preliminary corrected transmission rate parameter is superimposed with the buffer protection adjustment factor to generate a buffer-aware corrected transmission rate parameter.

[0114] The latest observation value of the round-trip delay sequence of the message at the sending node in the current adjustment period is obtained. The direction of change of the delay gradient is calculated based on the latest observation value. When the direction of change of the delay gradient indicates that the round-trip delay is in a continuous upward trend, a delay gradient suppression coefficient is generated.

[0115] The buffer-aware correction transmission rate parameter is multiplied and adjusted with the delay gradient suppression coefficient to generate the delay-aware correction transmission rate parameter.

[0116] The maximum transmission unit limit parameter and physical layer link rate limit parameter on the network path associated with the sending node are obtained. The delay-aware correction transmission rate parameter is limited to a closed interval consisting of zero and the physical layer link rate limit parameter to obtain the limited transmission rate parameter. The data packets to be sent are then subjected to packet length constraint processing or fragmentation processing according to the maximum transmission unit limit parameter so that the length of a single data packet to be sent does not exceed the allowed packet length range corresponding to the maximum transmission unit limit parameter.

[0117] The time-sensing corrected transmission rate parameter after the limitation processing is written as the target transmission rate parameter into the transmission rate register in the congestion control module of the transmission protocol stack of the sending node.

[0118] The transmission protocol stack's transmission scheduling module reads the target transmission rate parameter from the transmission rate register, controls the transmission interval between adjacent data packets according to the target transmission rate parameter, and performs data transmission operations at a rate not exceeding the target transmission rate parameter.

[0119] In an exemplary embodiment of the invention, after the multi-agent rate decision module of node A completes the collaborative experience fusion, the server reads the collaborative rate adjustment offset output by the module. This offset includes an offset direction identifier and an offset magnitude value. For example, if the server parses the offset direction identifier as negative and the offset magnitude value as 3 Mbps, it indicates that the effective experience of neighboring nodes in similar network conditions supports node A to further reduce its transmission rate based on its local candidate transmission rate. Subsequently, the server obtains the candidate transmission rate parameter of 42 Mbps generated by node A's local rate adjustment agent within the current adjustment period, and algebraically superimposes the candidate transmission rate with 3 Mbps according to the negative offset direction to generate a preliminary corrected transmission rate parameter of 39 Mbps.

[0120] The server continues to read the latest observations of the buffer queue occupancy sequence within the current adjustment period of node A. Node A's data buffer queue has a maximum capacity of 8MB, and the latest occupancy has reached 6.8MB. Based on this, the server calculates that the remaining free capacity of the buffer queue is only 1.2MB, and determines that the risk of the buffer queue exceeding its limit is high. The server calculates a buffer protection adjustment factor based on this risk level. Since the buffer queue is already close to the high-water mark, the buffer protection adjustment factor outputs a negative correction, for example, -2Mbps. The server superimposes the initial corrected transmission rate parameter of 39Mbps with -2Mbps to generate a buffer-aware corrected transmission rate parameter of 37Mbps, thereby prioritizing the reduction of the risk of further buffer data accumulation.

[0121] In one optional implementation, the server determines a buffer protection adjustment factor based on the buffer queue occupancy rate. The buffer queue occupancy rate is the ratio between the current number of bytes occupied in the buffer queue and the maximum capacity of the buffer queue. When the buffer queue occupancy rate is below a preset safety level, the buffer protection adjustment factor can be zero or a small positive value, indicating that no additional reduction is applied to the initial correction transmission rate parameter. When the buffer queue occupancy rate is between a preset warning level and a preset high level, the buffer protection adjustment factor outputs a small negative correction amount. When the buffer queue occupancy rate is above the preset high level, the buffer protection adjustment factor outputs a larger negative correction amount. For example, if the maximum buffer queue capacity is 8MB and the current occupancy is 6.8MB, the buffer queue occupancy rate is close to the high level, and the server can output a buffer protection adjustment factor of -2Mbps to reduce the initial correction transmission rate parameter; if the current occupancy is only 3MB, the server may not output any additional negative correction amount.

[0122] The server then obtains the latest observations of the round-trip delay sequence of node A's messages and calculates the direction of delay gradient change based on multiple recent sampling points. The round-trip delay of node A continues to rise from 70ms and 82ms to 95ms. The server identifies this continuously increasing trend and generates a delay gradient suppression coefficient, for example, 0.92. The server multiplies the 37Mbps buffer-aware correction transmission rate parameter with 0.92 to obtain a delay-aware correction transmission rate parameter of approximately 34Mbps, thus suppressing the continuously increasing queuing delay.

[0123] In one optional implementation, the server generates a delay gradient suppression coefficient based on the direction of round-trip delay changes across multiple consecutive sampling points. When the round-trip delays at multiple consecutive sampling points all show an upward trend, and the increase exceeds a preset delay change threshold, the server sets the delay gradient suppression coefficient to a value less than 1 to reduce buffer-aware correction of the transmission rate parameter. When the round-trip delay remains stable or shows a downward trend, the server sets the delay gradient suppression coefficient to a value close to 1 to avoid unnecessary suppression of the transmission rate.

[0124] For example, when the round-trip latency of the three most recent sampling points is 70ms, 82ms and 95ms respectively, the server determines that the round-trip latency is continuously increasing and sets the latency gradient suppression coefficient to 0.92; when the round-trip latency of the three most recent sampling points is 70ms, 68ms and 66ms respectively, the server determines that the latency is decreasing and sets the latency gradient suppression coefficient to 1 or a value close to 1.

[0125] The server further reads the maximum transmission unit (MTB) limit parameters and physical layer link rate limit parameters on the network path associated with node A. The physical layer link rate limit for the path where node A is located is 100Mbps. The server limits 34Mbps to a closed interval between 0 and 100Mbps to obtain the clipped transmission rate parameters. Simultaneously, the server performs length constraint processing on the video fragment packets to be sent according to the MTB limit parameters, fragmenting data packets exceeding the allowed packet length range to ensure that the length of a single data packet to be sent does not exceed the maximum allowed packet length of the path.

[0126] After completing the bandwidth limiting and message length processing, the server writes 34 Mbps as the target transmission rate parameter for node A into the transmission rate register of the transport protocol stack congestion control module. The transmission scheduling module of the transport protocol stack reads the target transmission rate parameter from this register and calculates the transmission interval between adjacent data packets based on this rate. The server schedules node A's video data packets according to this transmission interval, ensuring that node A performs data transmission operations at a rate not exceeding 34 Mbps within the current adjustment period. This simultaneously takes into account neighbor cooperation experience, buffer queue protection, latency growth suppression, link upper limit constraints, and message length limitations.

[0127] In this embodiment of the invention, the method further includes:

[0128] By using the collaborative range dynamic adjustment module deployed on each sending node, the rate correction magnitude between the candidate sending rate parameter and the target sending rate parameter of the final application within the current adjustment period is obtained.

[0129] The rate correction magnitude is compared and analyzed with the rate correction magnitude sequence in the historical adjustment period to calculate the fluctuation and variation parameter of the rate correction magnitude. The fluctuation and variation parameter reflects the stability of the influence of the multi-agent rate decision module on the local rate adjustment agent.

[0130] The neighbor message broadcast radius parameter of the sending node is adjusted according to the fluctuation variation degree parameter. When the fluctuation variation degree parameter indicates that the amplitude of collaborative correction is fluctuating, the neighbor message broadcast radius parameter is expanded to include more neighbor sending nodes in collaborative decision-making.

[0131] The neighbor experience reception threshold of the sending node is adjusted according to the fluctuation variation degree parameter. When the fluctuation variation degree parameter indicates that the fluctuation of the cooperative correction amplitude is reduced, the neighbor experience reception threshold is increased to reduce the reception and processing overhead of non-critical experience messages.

[0132] The experience message value evaluation module deployed on each sending node assigns an experience freshness decay weight to each received neighbor rate adjustment experience message. The experience freshness decay weight decays exponentially over time.

[0133] The experience message value assessment module calculates the average throughput improvement of the neighbor throughput change gradient sequence contained in the neighbor rate adjustment experience message, and performs weighted fusion processing on the average throughput improvement and the experience freshness decay weight to generate the experience reference value score of the neighbor rate adjustment experience message.

[0134] Based on the experience reference value score, the multiple neighbor rate adjustment experience messages received in the current adjustment period are sorted in descending order, and the neighbor rate adjustment experience messages with experience reference value scores lower than a preset value threshold are removed from the input of the multi-agent rate decision module.

[0135] By deploying a collaborative frequency adaptive adjustment module on each transmitting node, the fluctuation range of the target transmission rate parameter of the transmitting node is monitored within multiple consecutive adjustment cycles. When the fluctuation range is continuously lower than the preset fluctuation convergence threshold, the calling frequency of the multi-agent rate decision module is reduced.

[0136] By deploying an experience replay cache module on each sending node, the actual rate adjustment action trajectory, the corresponding network state transition sequence, and the corresponding throughput change gradient sequence of the sending node in each adjustment cycle are stored in the local experience replay cache area.

[0137] When the multi-agent rate decision module is not in a call cycle, it randomly samples historical experience fragments from the local experience replay buffer, and performs virtual collaborative deduction processing on the historical experience fragments and the node state representation vectors in the current adjustment cycle, thereby updating the set of experience replay optimization parameters in the rate adjustment strategy function inside the local rate adjustment agent.

[0138] In an embodiment of the invention, for example, after node A completes the data transmission for the current adjustment cycle, the server, through the collaborative range dynamic adjustment module deployed on node A, reads the candidate transmission rate parameter of 42 Mbps generated by the local rate adjustment agent and the target transmission rate parameter of 34 Mbps finally written into the transmission protocol stack, and calculates the rate correction amplitude between the two as 8 Mbps. The server compares this rate correction amplitude with the correction amplitude sequence recorded in the historical adjustment cycles of node A. For example, the correction amplitudes of the previous few cycles were 2 Mbps, 3 Mbps, 7 Mbps, and 8 Mbps, respectively. Based on this, the server calculates the fluctuation and variation parameter of the rate correction amplitude. This parameter is used to indicate whether the influence of the multi-agent rate decision module on the output result of the local rate adjustment agent is stable. When the correction amplitude continuously increases and the change is large, the server determines that the current local agent's independent judgment of the network state is insufficient, and more neighbor experience needs to be introduced to assist in decision-making.

[0139] The server adjusts the neighbor message broadcast radius parameter of node A based on the fluctuation variability parameter. When the cooperative correction amplitude fluctuation of node A increases, the server expands the broadcast radius from nodes within the same rack to nodes under the same edge egress link, enabling node A to receive rate adjustment experience from nodes B, C, and the newly added node D. Simultaneously, the server adjusts the neighbor experience reception threshold based on this fluctuation variability parameter. When the target transmission rate of node A stabilizes for several consecutive cycles and the cooperative correction amplitude fluctuation decreases, the server increases the neighbor experience reception threshold, receiving only neighbor messages with higher signal strength, stronger topological correlation, or higher experience value, thereby reducing the parsing, storage, and computational overhead of non-critical experience messages.

[0140] The server also filters each neighbor rate adjustment experience message through an experience message value assessment module. For experience messages sent by node B, the server assigns an experience freshness decay weight based on the message's generation time; newly generated experience messages have a higher weight, which decreases exponentially after multiple adjustment cycles. The server further reads the neighbor throughput change gradient sequence in the message and calculates the average throughput improvement. For example, if node B's effective throughput stability improves significantly after a slight rate reduction, the average improvement is considered high. The server weights and fuses the average throughput improvement with the experience freshness decay weight to generate an experience reference value score. For experience messages from node D that are too old and show no significant throughput improvement, if the server's calculated score is below a preset value threshold, the message is removed from the input of the multi-agent rate decision module. For experience messages from node B with high scores, the server prioritizes their submission to the collaborative decision-making process after sorting them in descending order.

[0141] The server further monitors the target transmission rate parameter of node A over several consecutive adjustment cycles through the cooperative frequency adaptive adjustment module. The target transmission rate of node A stabilizes between 34Mbps and 35Mbps over several cycles, with the fluctuation range continuously below the preset fluctuation convergence threshold. Based on this, the server reduces the calling frequency of the multi-agent rate decision module, for example, from once per cycle to once every three cycles, to reduce the resource consumption caused by cooperative attention calculation and neighbor experience fusion.

[0142] During periods when the multi-agent rate decision module is not invoked, the server does not cease its learning process. The server uses an experience replay cache module to store the actual rate adjustment action trajectory, corresponding network state transition sequence, and throughput change gradient sequence used by node A in each period in its local experience replay cache. When node A is not in an invocation period, the server randomly samples historical experience fragments from the local experience replay cache and performs virtual co-inference with these historical experience fragments and the current node state representation vector to simulate the throughput changes that may result from using different rate adjustment actions under similar states. Based on the inference results, the server updates the set of experience replay optimization parameters in the rate adjustment strategy function within the local rate adjustment agent, enabling node A to continuously utilize historical experience to optimize subsequent rate adjustment strategies even with reduced calls to the co-inference module.

[0143] In this embodiment of the invention, the method further includes:

[0144] When the sending node detects that the current observed value of the buffer queue occupied sequence exceeds the preset severe congestion alarm threshold, it triggers the emergency rate control mode.

[0145] In the emergency rate control mode, the normal generation process of the candidate transmission rate parameters by the local rate adjustment agent is suspended, and the emergency rate reduction ratio is directly calculated based on the difference between the current observation value of the buffer queue occupancy sequence and the preset severe congestion alarm threshold.

[0146] The emergency transmission rate parameter is generated by multiplying the target transmission rate parameter used by the transmitting node in the previous adjustment cycle by a reduction factor according to the emergency rate reduction ratio.

[0147] During the continuous monitoring period after applying the emergency transmission rate parameter, the recovery status of the message round-trip delay sequence is continuously monitored. When the latest observed value of the message round-trip delay sequence falls back to the preset normal delay range, the emergency rate control mode is gradually exited.

[0148] When the transmitting node detects that the available bandwidth sequence of the channel shows a continuous upward trend for multiple consecutive adjustment cycles, and the current observed value of the buffer queue occupied sequence is lower than the preset buffer idle threshold, the rate detection mode is triggered.

[0149] In the rate detection mode, a preset detection increment amplitude is superimposed on the target transmission rate parameter to generate a detection transmission rate parameter, and the data transmission operation is performed using the detection transmission rate parameter within a preset number of detection duration cycles.

[0150] After the preset number of detection cycles ends, the average throughput value and average round-trip delay value within the detection cycle are obtained, and the average throughput value and average round-trip delay value are compared and analyzed with the corresponding values ​​within the stable period before detection.

[0151] If the increase in the average throughput value relative to the pre-detection value exceeds a preset throughput increase threshold, and the increase in the average round-trip delay value relative to the pre-detection value is lower than a preset delay tolerance limit, then the detection transmission rate parameter is retained as a new baseline transmission rate parameter.

[0152] If the increase in the average throughput value does not exceed the preset throughput increase threshold, or the increase in the average round-trip delay value exceeds the preset delay tolerance limit, then the process will revert to the target transmission rate parameter before the application of the probe transmission rate parameter and reduce the rate probe trigger frequency in the subsequent preset backoff period.

[0153] After experiencing the rate detection mode, the detection result of the rate detection mode and the corresponding network state transition sequence are recorded as a special experience record. The rate detection experience sharing message is broadcast to the neighboring end nodes through the agent message broadcasting module with increased broadcast power. The rate detection experience sharing message contains a detection success identifier or a detection failure identifier.

[0154] In this embodiment of the invention, the method further includes:

[0155] By using the network state fingerprint construction module deployed on each sending node, cluster analysis is performed on the evolution trajectory of the node state representation vector over multiple consecutive adjustment cycles to extract representative state dwell points and state transition paths from the evolution trajectory.

[0156] The representative state dwell points are mapped to prototype state codewords in the network state fingerprint codebook, and directed edges with transition probabilities between the prototype state codewords are constructed according to the state transition path to form a network state transition topology graph.

[0157] At the beginning of each adjustment period, the node state representation vector of the sending node in the previous adjustment period is matched with the prototype state codeword in the network state transition topology to determine the network state codeword identifier of the sending node.

[0158] Based on the directed edges of the jump probability in the network state transition topology graph, starting from the network state codeword identifier of the current state of the sending node, predict the set of candidate network state codewords that may be entered in the next adjustment cycle and the jump probability value corresponding to each candidate network state codeword.

[0159] The candidate network state codeword set and the jump probability value corresponding to each candidate network state codeword are sent to the multi-agent rate decision module, so that when the multi-agent rate decision module performs collaborative experience fusion processing, it will prioritize retrieving neighbor rate adjustment experience fragments associated with the candidate network state codeword set.

[0160] By using the collaborative policy knowledge distillation module deployed on each sending node, the rate adjustment policy function accumulated by the local rate adjustment agent in multiple adjustment cycles is used as the teacher policy, and the initial rate adjustment agent of the newly added sending node in the communication network is used as the student policy.

[0161] Multiple representative node state representation vectors are sampled from the teacher strategy to serve as the anchor state set for knowledge distillation. The anchor state set is then input into the teacher strategy and the student strategy to obtain the probability distribution of teacher actions output by the teacher strategy and the probability distribution of student actions output by the student strategy.

[0162] Calculate the distribution difference parameter between the teacher's action probability distribution and the student's action probability distribution, and iteratively update the action selection parameter set of the student strategy with the goal of minimizing the distribution difference parameter until the distribution difference parameter is lower than a preset distillation convergence threshold;

[0163] The student policy, updated through knowledge distillation, is deployed to the newly added sending node of the communication network as the initial rate regulation policy function of the local rate regulation agent of the newly added sending node of the communication network.

[0164] The local rate adjustment agents of multiple sending nodes collectively report the node state representation vectors generated by themselves within the reporting period to the network management node. The network management node then performs global cluster analysis on the node state representation vectors of the entire network to identify congestion hotspots and bandwidth idle areas across the entire network.

[0165] The system receives global situational partition information issued by the network management node. The global situational partition information includes area identifiers for congestion hotspot areas and area identifiers for bandwidth idle areas. The global situational partition information is used as the basis for the multi-agent rate decision module to adjust the neighbor experience adoption tendency during collaborative experience fusion processing.

[0166] In this embodiment of the invention, for example, after nodes A, B, and C continuously execute multiple rate adjustment cycles, the server reads the node state representation vectors generated in each cycle through the network state fingerprint construction module deployed on each sending node, and performs cluster analysis on the evolution trajectory of these vectors. Node A's state vector repeatedly stays in the region of "decreased bandwidth, increased queue size, and increased latency," and the server extracts this type of state as a representative state dwell point. Node B is consistently in the region of "stable bandwidth, low queue size, and stable latency," and the server extracts it as another type of representative state dwell point. The server also records the transition paths between states; for example, node A gradually transitions from "normal transmission state" to "mild congestion state," and then enters "severe queue backlog state."

[0167] The server maps these representative state dwell points to prototype state codewords in the network state fingerprint codebook. For example, normal transmission state is mapped to codeword S1, light congestion state to codeword S2, and severe congestion state to codeword S3. Based on the state transition paths within consecutive periods, the server constructs directed edges with transition probabilities between prototype state codewords. For instance, the probability of S1 transitioning to S2 is 0.35, the probability of S2 transitioning to S3 is 0.62, and the probability of S2 falling back to S1 is 0.28, thus forming the network state transition topology. At the beginning of each adjustment period, the server performs nearest neighbor matching between the node state representation vector of node A from the previous period and the prototype state codewords in the topology graph to determine that node A is currently in the S2 light congestion state. Subsequently, based on the directed edge with the jump probability starting from S2, the server predicts the set of candidate network state codewords that may be entered in the next adjustment cycle and the corresponding jump probability value, and sends the candidate set to the multi-agent rate decision module, so that it can prioritize the retrieval of neighbor rate adjustment experience fragments related to the congestion deterioration path from S2 to S3 and the recovery path from S2 to S1.

[0168] The server also uses a collaborative policy knowledge distillation module to take the local rate adjustment policy functions formed by nodes A and B after multiple training cycles as the teacher policy, and the initial rate adjustment agent of newly added node E as the student policy. The server selects representative node state representation vectors from the historical states accumulated by the teacher policy to form an anchor state set, such as normal state, mild congestion state, severe congestion state, and bandwidth idle state. The server inputs these anchor states into the teacher policy and student policy respectively to obtain the teacher action probability distribution and the student action probability distribution, and then calculates the distribution difference parameter between the two. The server iteratively updates the action selection parameter set of the student policy with the goal of minimizing this difference until the action probability output by the student policy in the same state is close to that of the teacher policy, and the difference is lower than the preset distillation convergence threshold. The server then deploys the updated student policy to node E as the initial rate adjustment policy function of node E's local rate adjustment agent, so that new nodes do not need to explore rate adjustment rules from scratch.

[0169] At the global management level, the server also centrally sends the node state representation vectors generated by nodes A, B, and C during the reporting period to the network management node. The network management node performs global clustering of the node state representation vectors across the entire network, identifies congestion hotspots near shared egress links and bandwidth idle areas on edge links, and sends global situational partitioning information to the server. After receiving this information, the server writes the congestion hotspot identifier and bandwidth idle area identifier into the multi-agent rate decision module. In subsequent collaborative experience fusion, it increases the adoption tendency for successful rate reduction experiences from the same congestion hotspot and, under appropriate conditions, increases the reference weight for successful rate detection experiences from bandwidth idle areas.

[0170] In this embodiment of the invention, the method further includes:

[0171] By deploying a multi-path transmission management module on each sending node, the path identifier set of multiple concurrent transmission paths currently maintained by the sending node is obtained. These multiple concurrent transmission paths correspond to different receiving nodes or different network interfaces of the same receiving node.

[0172] For each of the multiple concurrent transmission paths, an independent local rate adjustment agent instance and an independent multi-agent rate decision instance are instantiated on the sending node for each transmission path.

[0173] For each local rate adjustment agent instance instantiated for the transmission path, independently collect the channel available bandwidth sequence, buffer queue occupancy sequence and message round-trip delay sequence corresponding to the transmission path, and independently generate candidate transmission rate parameters corresponding to the transmission path.

[0174] For each of the multi-agent rate decision instances instantiated for the transmission path, the neighboring sending node that shares the network link with the transmission path independently receives path-level neighbor rate adjustment experience messages broadcast for the transmission path.

[0175] After generating the target transmission rate parameters for each transmission path, the latest observation values ​​of the round-trip delay sequence of the message for each transmission path are obtained through the multi-path transmission management module. The latest round-trip delay observation values ​​of the message for each transmission path are then input into the path quality evaluation function to generate the path quality score for each transmission path.

[0176] Based on the path quality score of each transmission path, the multi-path transmission management module calculates the path quality score ratio between each transmission path, and allocates the data packets to be sent in the data buffer queue to each transmission path according to the path quality score ratio.

[0177] During the data packet allocation process, a continuous sequence number identifier is maintained for consecutive data packets allocated to the same transmission path, and the number of data packets allocated on each transmission path is checked for proportional consistency with the target transmission rate parameter of each transmission path.

[0178] The scheduling weight of data packets between paths is adjusted according to the result of the proportional consistency check. When the data packet allocation ratio of a certain transmission path exceeds the ratio of the target transmission rate parameter of the transmission path to the sum of the target transmission rates of all paths, the number of data packets allocated to that transmission path is reduced.

[0179] By deploying the path-level collaborative information sharing submodule in the multi-path transmission management module, the node state representation vector and rate adjustment strategy function parameters generated by the local rate adjustment agent instance of each transmission path are shared across the transmission paths, so that the rate adjustment experience trained for one transmission path can be transferred to the rate adjustment strategy function update process of other transmission paths.

[0180] When the multipath transmission management module detects an unrecoverable transmission failure on a certain transmission path, it performs a freeze-suspend operation on the local rate adjustment agent instance and the multi-agent rate decision instance corresponding to the transmission path, and reallocates the data packets that were not successfully sent on the transmission path to the data buffer queues corresponding to the other normally operating transmission paths.

[0181] In this embodiment of the invention, for example, the server reads the set of concurrent transmission path identifiers currently maintained by the sending node through a multi-path transmission management module deployed on the sending node. For instance, if node A simultaneously sends video segment data to both a wired network card and a wireless network card at the same receiving end, the server records the wired path as Path-1 and the wireless link as Path-2. In another service, node A also transmits the same service stream to two receiving nodes respectively, and the server records the corresponding paths as Path-3 and Path-4 respectively. Each path has independent network bandwidth, buffer pressure, and round-trip latency. Therefore, the server does not mix multiple paths into a single unified adjustment object, but instead instantiates independent local rate adjustment agent instances and independent multi-agent rate decision instances for each transmission path on node A.

[0182] For Path-1, the server independently collects the available channel bandwidth sequence, buffer queue occupancy sequence, and packet round-trip time sequence along this path to construct the node state representation vector corresponding to Path-1. The local rate adjustment agent of Path-1 then generates candidate transmission rate parameters, for example, 40Mbps. For Path-2, the server similarly independently collects the three types of sequences for the wireless link. Due to the large fluctuations in wireless link bandwidth and high round-trip time, the local rate adjustment agent of Path-2 generates candidate transmission rate parameters of 18Mbps. The server also receives path-level neighbor rate adjustment experience messages for the multi-agent rate decision instances of Path-1 and Path-2 respectively. Path-1 only receives neighbor node experience shared with the wired egress link, while Path-2 only receives neighbor node experience shared with the wireless access link, thus avoiding interference between experiences from different network paths.

[0183] After generating target transmission rate parameters for each path, the server reads the latest round-trip time (RTT) observations for each path through the multi-path transmission management module and inputs them into the path quality evaluation function. Path-1 has a latest RTT of 28ms and a target transmission rate of 38Mbps, resulting in a higher path quality score. Path-2 has a latest RTT of 75ms and a target transmission rate of 15Mbps, resulting in a lower path quality score. The server calculates the path quality score ratio based on each path's score and allocates the data packets to be sent in the data buffer queue to each path according to this ratio. This ensures that more video segments are sent via Path-1, and a smaller amount of data is sent via Path-2 to maintain link utilization.

[0184] During data packet allocation, the server maintains a sequence number continuity identifier for consecutive data packets allocated to the same path, ensuring that the receiving end can acknowledge and reassemble them in the order within the path. Simultaneously, the server performs a proportional consistency check between the actual number of data packets allocated to each path and the target transmission rate parameters for each path. If Path-2 only accounts for 28% of the total target transmission rate of all paths, but its actual allocated packet ratio reaches 45%, the server immediately reduces the scheduling weight of Path-2 and transfers some subsequent packets to Path-1 to prevent excessive data backlog on low-quality paths.

[0185] The server also shares node state representation vectors and rate adjustment policy function parameters among the paths through a path-level collaborative information sharing submodule. The rate reduction experience learned by Path-2 during wireless link congestion can serve as a reference for Path-1 when similar queue backlogs and latency increases occur; the experience of Path-1 successfully increasing its rate when bandwidth is idle can also be transferred to the policy update process of Path-2. Finally, when the server detects continuous timeouts, long-term missing acknowledgment messages, and is determined to be an unrecoverable transmission failure for Path-2, the server freezes and suspends the local rate adjustment agent instance and multi-agent rate decision instance corresponding to Path-2, and puts the unsuccessfully sent data packets on Path-2 back into the data buffer queue, allocating them to still-operating transmission paths such as Path-1 for continued transmission.

[0186] Furthermore, to illustrate the adaptability of the technical solution of the present invention in different network protocol environments, the following embodiments are also provided.

[0187] In another optional embodiment of the present invention, the method is also applicable to high-speed interconnection networks such as InfiniBand that employ a credit-based flow control mechanism to solve the congestion problem caused by the depletion of credit at the receiving port in a multi-sender communication scenario (many-to-one). Credit represents the available buffer space of the receiving port. When multiple senders simultaneously send data to the same receiving port, the total amount of data arriving may exceed the receiving port's buffer limit, leading to credit depletion, which in turn causes link congestion and indirectly affects the bandwidth of other unrelated communication pairs in the network (such as from port X to port Y). In this embodiment, after generating the target transmission rate parameters, the sending node further performs credit protection adjustment for data transmission through the "authorization scheduling module": First, the client service (data requester) initializes the number of authorization tokens, which is dynamically set according to the receiving port's buffer space. The client service initiates a read request to the storage service. After receiving the read request, the storage end reads data from the disk according to the offset and size of the request, and divides the data into multiple data blocks according to a preset partitioning size, which is adapted to the available buffer space of the receiving port. Before sending each data block, the storage service requests authorization from the client service. The client authorization module determines whether authorization is possible based on currently available authorization tokens: if sufficient tokens are available, it returns authorization permission to the storage end and deducts the corresponding number of tokens from the token pool; otherwise, the authorization request is placed in the waiting authorization queue. Upon receiving authorization, the storage end sends multiple data blocks to the client sequentially, with the total data sent at one time not exceeding the data limit corresponding to the authorization. The client receives the data sequentially, releases the authorization upon completion, returns the corresponding token to the token pool, and wakes up the next authorization request in the waiting queue to continue execution.

[0188] In this embodiment, the number of authorization tokens and the size of a single data block can be dynamically adjusted by the multi-agent rate decision module based on the current network state perception information, thereby achieving synergy with the traditional rate adaptive adjustment mechanism.

[0189] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned distributed transmitter rate adaptive adjustment method utilizing multiple agents. Figure 3 As shown, Figure 3 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, the processor 112, and the communication unit 113 are electrically connected to each other directly or indirectly.

[0190] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A distributed transmitter rate adaptive adjustment method utilizing multiple agents, characterized in that, The method includes: A set of network status awareness information from multiple sending nodes is obtained, wherein the set of network status awareness information includes the channel available bandwidth sequence, buffer queue occupancy sequence, and message round-trip delay sequence collected by each sending node; By deploying local rate adjustment agents at each transmitting node, the available channel bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence are constructed into a node state representation vector. The rate adjustment strategy function inside the local rate adjustment agent is updated based on the node state representation vector to generate candidate transmission rate parameters for the current adjustment period. Receive a neighbor rate adjustment experience message broadcast by a neighboring sending node. The neighbor rate adjustment experience message includes the rate adjustment action trajectory of the neighboring sending node in adjacent historical adjustment periods, the network state transition sequence corresponding to the rate adjustment action trajectory, and the throughput change gradient sequence triggered by the network state transition sequence. The neighbor rate adjustment experience message and the candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period are input into the multi-agent rate decision module. The multi-agent rate decision module performs cooperative experience fusion processing on the rate adjustment action trajectory and the network state transition sequence to obtain the cooperative rate adjustment offset. The candidate transmission rate parameters are dynamically corrected based on the cooperative rate adjustment offset to generate target transmission rate parameters, and the target transmission rate parameters are configured into the transmission protocol stack of the sending node to perform data transmission operations.

2. The method according to claim 1, characterized in that, The acquisition of the network state awareness information set of multiple sending nodes includes: The channel sounding module deployed at each sending node sends a sequence of sounding messages to the receiving node, and calculates the time interval parameter and the arrival rate parameter of the acknowledgment message based on the receiving timestamp and sending timestamp in the acknowledgment message sequence. Input the time interval parameter and the acknowledgment message arrival rate parameter into the bandwidth estimation filter to generate a channel available bandwidth sequence; The current number of bytes occupied in the data buffer queue is read by the buffer queue monitoring module to form a buffer queue occupancy sequence; The round-trip time delay of the message is calculated by the delay measurement module to form a round-trip time delay sequence of the message; The available bandwidth sequence of the channel, the buffer queue occupancy sequence, and the message round-trip delay sequence are processed for abnormal jump points, trend terms are extracted, and outliers are filtered out, and then combined into a normalized three-channel time series data structure to form the network state awareness information set.

3. The method according to claim 1, characterized in that, The process involves using local rate adjustment agents deployed at each transmitting node to construct a node state representation vector from the available channel bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence. The rate adjustment strategy function within the local rate adjustment agent is then updated based on this node state representation vector to generate candidate transmission rate parameters for the current adjustment period. This includes: Extract the channel available bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence of the corresponding sending node within the current adjustment period from the network state awareness information set; The available channel bandwidth sequence, the buffer queue occupancy sequence, and the message round-trip delay sequence are respectively time-domain feature encoded to generate a channel state encoding vector, a buffer state encoding vector, and a delay state encoding vector. The channel state coding vector, the buffer state coding vector, and the delay state coding vector are concatenated to form a node state representation vector. The node state representation vector is input into the rate adjustment strategy function, and candidate transmission rate parameters are generated through the action selection parameter set, and state value estimates are generated through the value estimation parameter set. The rate adjustment strategy function is updated based on the actual throughput value and the state value estimate.

4. The method according to claim 1, characterized in that, The receiving of the neighbor rate adjustment experience message broadcast by the neighbor sending node includes: Each sending node maintains its local list of neighboring nodes; The agent message broadcasting module sends a local rate adjustment experience sharing message to the neighbor sending node in the local neighbor node list. The local rate adjustment experience sharing message includes the rate adjustment action trajectory, network state transition sequence and throughput change gradient sequence. The intelligent agent message receiving module listens for broadcast messages, parses the sending source node identifier, and matches and verifies the sending source node identifier with the local neighbor node list; If the source node identifier exists in the local neighbor node list, or if the source node identifier does not exist in the local neighbor node list and the signal strength of the broadcast message exceeds a preset received strength threshold, the neighbor rate adjustment action trajectory, the neighbor network state transition sequence, and the neighbor throughput change gradient sequence are extracted, processed, and combined and encapsulated into the neighbor rate adjustment experience message.

5. The method according to claim 1, characterized in that, The neighbor rate adjustment experience message and the candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period are input into the multi-agent rate decision module. The multi-agent rate decision module performs cooperative experience fusion processing on the rate adjustment action trajectory and the network state transition sequence to obtain the cooperative rate adjustment offset, including: The neighbor rate adjustment action trajectory contained in the neighbor rate adjustment experience message is input into the experience parsing submodule of the multi-agent rate decision module. The trajectory feature extraction processing is performed on the neighbor rate adjustment action trajectory to extract the rate adjustment direction preference feature and rate adjustment amplitude preference feature of the neighbor sending node in the continuous adjustment period. The neighbor network state transition sequence contained in the neighbor rate adjustment experience message is input into the experience parsing submodule. The state transition pattern mining process is performed on the neighbor network state transition sequence to extract the state transition triggering condition pattern of the neighbor sending node triggering rate adjustment operation under different network state conditions. The neighbor throughput change gradient sequence contained in the neighbor rate adjustment experience message is input into the experience parsing submodule. The gradient effect analysis processing of the neighbor throughput change gradient sequence is performed to extract the positive effect mapping relationship and negative effect mapping relationship between the rate adjustment action trajectory of the neighbor sending node and the throughput change. The candidate transmission rate parameters generated by the local rate adjustment agent within the current adjustment period are obtained, and the candidate transmission rate parameters are input into the local action encoding submodule of the multi-agent rate decision module. The candidate transmission rate parameters are then processed by action intent encoding to generate a local rate adjustment action intent vector. The collaborative attention calculation submodule of the multi-agent rate decision module calculates the action similarity attention weight between the local rate adjustment action intention vector and the rate adjustment direction preference feature, and calculates the amplitude matching degree attention weight between the local rate adjustment action intention vector and the rate adjustment amplitude preference feature. The collaborative attention calculation submodule combines the action similarity attention weight and the amplitude matching attention weight to apply to the positive effect mapping relationship and the negative effect mapping relationship. It performs attention weighting enhancement processing on the positive effect mapping relationship and attention weighting suppression processing on the negative effect mapping relationship to obtain the effect mapping set after attention regulation. The state transition triggering condition pattern is matched with the node state representation vector generated by the local rate regulation agent in the current regulation period, and the similarity distance between the node state representation vector and each state condition template in the state transition triggering condition pattern is calculated. Based on the similarity distance, determine the neighbor state transition triggering condition that best matches the node state representation vector, and extract the attention-modulated effect mapping relationship corresponding to the best-matching neighbor state transition triggering condition as a local transferable experience fragment. The local transferable experience fragment is input into the collaborative offset generation submodule of the multi-agent rate decision module. A positive rate offset proposal is generated based on the positive effect mapping relationship in the local transferable experience fragment, and a negative rate offset suppression is generated based on the negative effect mapping relationship in the local transferable experience fragment. The positive rate offset proposal and the negative rate offset suppression are vector-synthesized to obtain a coordinated rate adjustment offset that includes the offset direction and offset magnitude.

6. The method according to claim 1, characterized in that, The step of dynamically correcting the candidate transmission rate parameters based on the cooperative rate adjustment offset to generate target transmission rate parameters, and configuring the target transmission rate parameters into the transmission protocol stack of the sending node to execute data transmission operations, includes: The coordinated rate adjustment offset is obtained from the multi-agent rate decision module, and the offset direction identifier and offset magnitude value contained in the coordinated rate adjustment offset are parsed. The candidate transmission rate parameters generated by the local rate adjustment agent in the current adjustment period are obtained, and the candidate transmission rate parameters and the offset amplitude value are algebraically superimposed in the direction indicated by the offset direction identifier to generate preliminary corrected transmission rate parameters. Obtain the latest observation value of the buffer queue occupancy sequence of the sending node in the current adjustment period, and determine the spare capacity of the buffer queue and the degree of risk of the buffer queue exceeding the limit based on the latest observation value; The buffer protection adjustment factor is calculated based on the over-limit risk level of the buffer queue. The buffer protection adjustment factor has an inverse adjustment relationship with the over-limit risk level. When the over-limit risk level indicates that the buffer queue is close to the overflow state, the buffer protection adjustment factor outputs a negative correction amount to reduce the initial correction transmission rate parameter. The preliminary corrected transmission rate parameter is superimposed with the buffer protection adjustment factor to generate a buffer-aware corrected transmission rate parameter. The latest observation value of the round-trip delay sequence of the message at the sending node in the current adjustment period is obtained. The direction of change of the delay gradient is calculated based on the latest observation value. When the direction of change of the delay gradient indicates that the round-trip delay is in a continuous upward trend, a delay gradient suppression coefficient is generated. The buffer-aware correction transmission rate parameter is multiplied and adjusted with the delay gradient suppression coefficient to generate the delay-aware correction transmission rate parameter. The maximum transmission unit limit parameter and physical layer link rate limit parameter on the network path associated with the sending node are obtained. The delay-aware correction transmission rate parameter is limited to a closed interval consisting of zero and the physical layer link rate limit parameter to obtain the limited transmission rate parameter. The data packets to be sent are then subjected to packet length constraint processing or fragmentation processing according to the maximum transmission unit limit parameter so that the length of a single data packet to be sent does not exceed the allowed packet length range corresponding to the maximum transmission unit limit parameter. The time-sensing corrected transmission rate parameter after the limitation processing is written as the target transmission rate parameter into the transmission rate register in the congestion control module of the transmission protocol stack of the sending node. The transmission protocol stack's transmission scheduling module reads the target transmission rate parameter from the transmission rate register, controls the transmission interval between adjacent data packets according to the target transmission rate parameter, and performs data transmission operations at a rate not exceeding the target transmission rate parameter.

7. The method according to claim 1, characterized in that, The method further includes: The rate correction magnitude between the candidate transmission rate parameter and the target transmission rate parameter is obtained through the collaborative range dynamic adjustment module, and the fluctuation variation parameter is calculated based on the rate correction magnitude and the rate correction magnitude sequence within the historical adjustment period. Adjust the neighbor message broadcast radius parameter and the neighbor experience reception threshold according to the fluctuation variability parameter; The experience message value assessment module generates an experience reference value score based on the experience freshness decay weight and the gradient sequence of neighbor throughput changes, and removes neighbor rate adjustment experience messages that are below the preset value threshold from the input of the multi-agent rate decision module. When the fluctuation range of the target transmission rate parameter is continuously lower than the preset fluctuation convergence threshold, the calling frequency of the multi-agent rate decision module is reduced. When the multi-agent rate decision module is not in a call cycle, it updates the set of experience replay optimization parameters based on historical experience fragments in the local experience replay cache.

8. The method according to claim 1, characterized in that, The method further includes: The network state fingerprinting module is used to perform cluster analysis on the evolution trajectory of the node state representation vector over multiple consecutive adjustment cycles to form a network state transition topology. At the beginning of each adjustment period, the network state codeword identifier is determined based on the node state representation vector in the previous adjustment period and the network state transition topology, and the candidate network state codeword set and transition probability value are predicted. The candidate network state codeword set and the jump probability value are sent to the multi-agent rate decision module; Through the collaborative strategy knowledge distillation module, the rate regulation policy function accumulated by the local rate regulation agent is used as the teacher policy to iteratively update the student policy of the newly added sending node in the communication network. The network management node performs global clustering analysis on the state representation vectors of all nodes in the network and issues global situational partitioning information, which serves as the basis for the multi-agent rate decision module to adjust the neighbor experience adoption tendency.

9. The method according to claim 1, characterized in that, The method further includes: The multi-path transmission management module obtains a set of path identifiers for multiple concurrent transmission paths; For each transmission path, instantiate an independent local rate adjustment agent instance and a multi-agent rate decision instance on the sending node; For each transmission path, independently collect the available channel bandwidth sequence, buffer queue occupancy sequence, and message round-trip delay sequence, generate corresponding candidate transmission rate parameters, and receive path-level neighbor rate adjustment experience messages; After generating the target transmission rate parameters for each transmission path, a path quality score is generated based on the latest observation of the round-trip delay sequence of the message for each transmission path, and data messages to be sent are allocated according to the path quality score ratio. Adjust the scheduling weight of data packets between paths based on the proportional consistency check results; share cross-path experience among various transmission paths; When an unrecoverable transmission failure occurs on a transmission path, the corresponding instance is frozen and suspended, and the unsuccessfully sent data packets are reallocated.

10. A distributed transmitter rate adaptive adjustment system utilizing multiple agents, characterized in that, Includes at least one service unit; The service unit includes a storage unit and a computing unit; the storage unit is used to store program code; The computing unit is used to run the program code to execute the distributed transmitter rate adaptive adjustment method using multiple agents as described in any one of claims 1 to 9.

Images