Multi-agent dynamic communication optimization method based on energy state and probabilistic graph model

Abstract

The present application relates to the technical field of communication optimization, and relates to a multi-agent dynamic communication optimization method based on energy state and a probabilistic graph model. By constructing a distributed probabilistic graph model, each agent can infer the global system pressure state in a decentralized manner based on local observations and limited information exchange with neighbor nodes, replacing the traditional method of relying on central node aggregation calculation. Each agent models its own request queue length, average response delay, and residual energy as a probabilistic graph node, and iteratively transmits lightweight belief messages between nodes based on the belief propagation algorithm, thereby collaboratively approximating the joint probability distribution of global load and energy. Not only does this completely eliminate the communication bottleneck and single-point failure risk of the central node, but it also significantly improves the perception accuracy and robustness of the system's implicit congestion state through probabilistic inference, enabling the pressure evaluation to upgrade from deterministic exponential alerts to dynamic probability portraits with uncertainty measurement.

Description

Technical Field

[0001] This invention belongs to the field of communication optimization technology and relates to a multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models. Background Technology

[0002] In the fields of distributed collaboration and swarm intelligence, multi-agent systems have become a key platform for achieving parallel processing and large-scale collaboration of complex tasks. As the "neural network" of collaborative interaction among agents, communication efficiency directly determines the system's task completion rate and collaborative accuracy in dynamic environments. Meanwhile, energy supply, as the "energy foundation" for agents to maintain survival and operational logic, is a prerequisite for ensuring the long-term operation of the entire cluster. Especially in resource-constrained heterogeneous network environments, the regulation of communication frequency and the robust maintenance of remaining energy together constitute the underlying dual-support system for agents executing complex logic.

[0003] It is worth noting that there is a deep and complex coupling relationship among communication load, task timeliness, and energy consumption. In the process of multi-agent interaction, communication is not simply the transmission of data bits, but a process involving a dynamic trade-off between "coordination benefits" and "resource costs." When the task load increases sharply, high-frequency channel occupancy will not only cause severe network congestion and retransmission storms, leading to an exponential degradation of information effectiveness, but will also significantly accelerate the energy depletion rate of agents, causing local node failures and shutdowns.

[0004] However, current communication management for multi-agent systems under high load conditions still has significant shortcomings. Existing communication suppression schemes mostly adopt binary strategies based on fixed thresholds, simply dividing emergency and non-emergency states without analyzing the essential value of data payloads from the perspective of information entropy. This results in a large amount of repetitive, low-value redundant semantic information occupying limited bandwidth. The lack of distributed inference capabilities for system pressure distribution and forward-looking strategy evolution leads to the overall cluster exhibiting fragile robustness when facing sudden peak traffic. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides a multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: This invention provides a method for optimizing dynamic communication among multiple agents based on energy state and probabilistic graphical models, which includes the following steps: The task request queue length, remaining energy value, and signal-to-noise ratio data of the current communication link of the intelligent agent are obtained in real time, and the task request queue length, remaining energy value, and signal-to-noise ratio data are combined to generate a local state feature vector. Belief message data is generated based on local state feature vectors, and the belief message data is interacted with neighboring agents within a preset communication range; the received neighbor belief message data is used to iteratively update the locally stored factor graph model, and the global congestion probability distribution data is calculated. In response to a request for task data to be sent, the semantic features of the task data are extracted, and the semantic information entropy and information timeliness values ​​are calculated in combination with historical transmission records; the semantic information entropy and information timeliness values ​​are weighted and fused to generate a communication value index. A Lyapunov optimization model containing queue stability constraints and energy loss penalties is constructed. The local state feature vector, global congestion probability distribution data and communication value index are input into the Lyapunov optimization model to solve for the optimal semantic compression coefficient and transmission backoff probability value of the current time slot. Based on the semantic compression coefficient, hierarchical feature extraction and encoding are performed on the task data to be sent to generate a target compressed data packet; the sending time point is determined based on the sending backoff probability value, and the target compressed data packet is sent to the receiving end.

[0007] Another aspect of the present invention provides a multi-agent dynamic communication optimization device based on an energy state and probabilistic graphical model, including a processor, a memory, and a communication bus; The memory stores a computer-readable program that can be executed by the processor; The communication bus enables communication between the processor and the memory; When the processor executes the computer-readable program, it performs steps in a multi-agent dynamic communication optimization method based on an energy state and probabilistic graphical model as described in any one of the present invention.

[0008] As described above, the multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models provided by this invention has at least the following beneficial effects: This invention provides a multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models. By constructing a distributed probabilistic graphical model, each agent can infer the global system stress state in real time in a decentralized manner based on local observations and limited information exchange with neighboring nodes, replacing the traditional method that relies on centralized node aggregation calculation. Specifically, each agent models its own request queue length, average response latency, and remaining energy as probabilistic graphical nodes, and iteratively transmits lightweight belief messages between nodes based on a belief propagation algorithm, thereby collaboratively approximating the joint probability distribution of global load and energy. This technology not only completely eliminates the communication bottleneck and single-point failure risk of the central node, but also significantly improves the accuracy and robustness of perceiving the implicit congestion state of the system through probabilistic inference, upgrading stress assessment from deterministic exponential alarms to dynamic probabilistic profiling with uncertainty metrics.

[0009] This invention introduces a communication value game mechanism that integrates information entropy and energy potential, and establishes a multi-level adaptive semantic compression coding strategy based on this. Before initiating non-critical communication, each agent first calculates the real-time utility value of the communication using a Lyapunov optimization framework, based on the inferred system pressure probability distribution, the information entropy reduction of the information carried in this communication, and the "energy potential" defined by its own remaining energy. This allows the agent to decide whether to delay, degrade, or send normally. Under high load, the system can achieve a game equilibrium of communication resources by quantifying the value-energy ratio of communication, thereby maximizing the efficiency of critical information flow while avoiding avalanche effects and ensuring the continuous and stable operation of the system under extreme pressure. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0012] Figure 2 This is a schematic diagram of the logical connection for constructing the Lyapunov optimization model in this invention.

[0013] Figure 3 This is a schematic diagram of the device structure connection of the present invention. Detailed Implementation

[0014] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0015] Overview of this application: In existing technologies, communication management in distributed agent systems often relies on static thresholds or fixed policies, making it difficult to balance system load and communication efficiency. Traditional methods, under high system load, can easily lead to network congestion and response delays due to a surge in internal communication requests, triggering a communication avalanche. Existing mechanisms lack real-time awareness of overall system pressure and cannot dynamically adjust communication behavior. Especially under peak load, synchronization and assistance requests between agents may compete for resources, leading to a decline in the performance of critical external services and compromised system stability.

[0016] To address the aforementioned issues, the study discovered a mapping relationship between the system stress index and the communication inhibition probability, enabling dynamic adjustment through the establishment of a stress-inhibition coupling model. The research revealed that non-critical communications are highly sensitive to system stress, while critical communications require priority protection. Therefore, the idea of ​​dynamically switching communication modes based on the stress index was proposed. Further experimental verification incorporated the mapping relationship between the stress index and inhibition probability into the agent's decision-making mechanism, forming an adaptive control system.

[0017] Specifically, each agent monitors its own request queue length, average response latency, and remaining energy in real time, and reports these to the central node or calculates a global system pressure index through distributed consensus. This pressure index is associated with a pre-defined communication suppression probability table, defining suppression probabilities under different pressure levels. Before an agent needs to initiate non-critical communication, it queries the current system pressure index and, combined with the urgency of the communication, obtains a local communication suppression probability. The agent uses this probability to decide whether to delay the communication or downgrade it to a shorter form. Through this probability-based proactive yielding, under high load, the system prioritizes processing external requests while internal communication is orderly downgraded, maintaining system stability and preventing system crashes.

[0018] Compared to existing technologies, traditional methods rely on fixed communication strategies and lack pressure compensation mechanisms, making them prone to avalanche effects under high system loads. This solution innovatively integrates multi-agent state data and establishes a pressure-inhibition coupling model to dynamically regulate communication behavior. Unlike existing static management models, this solution can intelligently switch communication modes based on real-time system pressure and achieve orderly degradation through probabilistic decision-making, significantly improving the system's robustness under peak loads. Through the above technical solution, this application effectively overcomes the system instability problem caused by communication overload, optimizing internal communication efficiency while ensuring critical services. The dynamic inhibition probabilistic mechanism balances communication timeliness and system stability, while the adaptive adjustment function ensures long-term operational reliability.

[0019] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Example 1:

[0020] Please see Figure 1 As shown, a multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models is proposed, which includes the following steps: The task request queue length, remaining energy value, and signal-to-noise ratio data of the current communication link of the intelligent agent are acquired in real time, and the task request queue length, remaining energy value, and signal-to-noise ratio data are combined to generate a local state feature vector.

[0021] Specifically, by activating the hardware status monitoring interface at the agent's underlying layer, the number of pending data packets in the network layer buffer register is read in real time as the task request queue length. The remaining energy value is obtained by reading the real-time voltage and current readings from the power management unit and integrating them. Simultaneously, the physical layer receive signal strength indicator is invoked to obtain the signal power of the current channel. With background noise power ; Next, in order to eliminate the influence of different physical dimensions on subsequent probabilistic graph inference, the original data was subjected to dimensionless and normalized processing, and a normalized task request queue length was constructed. Normalized residual energy value and normalized channel quality factor The formula for calculating the normalized task request queue length is as follows: In the formula The maximum capacity limit of the preset hardware buffer; the formula for calculating the normalized remaining energy value is as follows: In the formula This represents the initial energy value of the agent's battery when it is fully charged; the calculation of the normalized channel quality factor requires first using the formula... Calculate the signal-to-noise ratio in decibels. Then use the formula Perform a linear mapping, where and These represent the highest theoretical signal-to-noise ratio and the lowest demodulation threshold allowed by the communication protocol, respectively. If the calculated result exceeds the [0,1] interval, it is truncated. Finally, the three dimensionless scalars are combined in order to generate the local state feature vector at the current moment. , where T represents the transpose of the vector.

[0022] Belief message data is generated based on the local state feature vector, and the belief message data is interacted with neighboring agents within a preset communication range; the received neighbor belief message data is used to iteratively update the locally stored factor graph model, and the global congestion probability distribution data is calculated.

[0023] In the preferred embodiment of this application, belief message data is generated based on local state feature vectors, and the generation logic is as follows: Obtain the agent's own local state feature vector, which includes the task request queue length, remaining energy value, and signal-to-noise ratio data of the current communication link; The local state feature vector is analyzed, and the task request queue length and remaining energy value are extracted. The task request queue length and remaining energy value are normalized and weighted and summed to calculate the comprehensive load pressure index that reflects the current node's processing capacity. Based on the preset pressure-probability mapping model, the comprehensive load pressure index is converted into a local state probability distribution describing the possibility that a node is currently in different congestion levels. Extract the signal-to-noise ratio (SNR) data from the local state feature vector, and calculate the link confidence weight of the current communication link based on the SNR data; The link confidence weights are used to correct the magnitude of each probability value in the local state probability distribution to generate belief message data.

[0024] Specifically, the normalized task request queue length is extracted from the local state feature vector. With normalized residual energy value The comprehensive load pressure index is calculated using a linear weighted aggregation algorithm. The calculation formula is: ,in The preset task weight coefficient, 1- As the energy weighting coefficient, this formula not only considers the direct pressure from task backlog, but also through ( The item quantifies the potential survival pressure under low power conditions; Next, based on the pre-defined pressure-probability mapping model, Gaussian radial basis functions are used to... Mapping to M discrete congestion level state spaces generates local state probability distributions. The probability density of the k-th state is calculated using the following formula: In the formula σ is the central feature value corresponding to the kth congestion level, for example, 0 for idle and 1 for congested. σ is the bandwidth parameter for the control state discrimination. This step ensures that the sum of the probability distributions is 1 through Softmax normalization. Subsequently, the signal-to-noise ratio data was extracted from the feature vector. Construct a logistic regression function to calculate link confidence weights. The formula is , where λ is the slope parameter and θ is the signal-to-noise ratio threshold offset. This weight reflects the reliability of information transmission in the current communication channel. A value close to 1 indicates an excellent channel, and a value close to 0 indicates an unreliable channel. Finally, the link confidence weights are used to correct the magnitude of the local state probability distribution, generating the final belief message data. The corrected formula is as follows: ,in For a noise vector to follow a uniform distribution (i.e., all elements are 1 / M), the physical meaning of this step is: when the channel quality is poor... When the probability is low, uncertainty (entropy increase) is actively injected into the belief message to reduce the degree of trust of neighboring nodes in the message, thereby avoiding the propagation and amplification of erroneous information in the probability graph.

[0025] It should be added that, in the above description, This represents the local state probability distribution, a probability vector describing the likelihood of the agent node being in different congestion levels. This probability distribution reflects the agent's probability based on the current overall load pressure index. The results were calculated using a pressure-probability mapping model. Each element This represents the probability that the agent is currently in the k-th congestion level; Specifically: This may represent the probability that the agent is in an "idle" state. This may represent the probability that the agent is in a "normal" state. These may represent the probability that the agent is in a "congested" state; the sum of these probabilities is 1, i.e. This ensures that the vector represents a valid probability distribution.

[0026] In the preferred embodiment of this application, the global congestion probability distribution data is calculated, including: Receive neighbor belief message data from neighbor nodes within a preset communication range; input the neighbor belief message data into the locally stored factor graph model, extract the neighbor congestion probability information contained therein, perform message aggregation calculation with the belief message data, and generate node state confidence; Compare the currently calculated node state confidence with the confidence value calculated in the previous iteration. If the difference between the two is within the preset stable error range, the factor graph model is determined to have converged and the iteration stops. If the difference between the two is not within the preset stable error range, the factor graph model is determined to have not yet converged and the next iteration needs to be continued. Extract the confidence scores of the finally converged node states, format and map them to generate global congestion probability distribution data.

[0027] Specifically, the agent continuously listens for and receives neighbor belief messages broadcast by neighbor nodes within a preset communication topology via a wireless communication interface. All received messages are temporarily stored in a local message buffer. Next, the locally stored factor graph model update engine is activated, fusing local observations and neighbor inferences using the Hadamard product. The specific formula for calculating node state confidence is as follows: ,in This represents the kth congestion level (e.g., k=1, 2, 3 correspond to idle, normal, and congested, respectively). The state of the current node calculated in the t-th iteration. The confidence level value, This refers to the local initial state probability calculated based on local sensor data in the previous steps. This represents the feedback from all neighboring nodes in set N(i) regarding their state at the previous time step. The probability product, whose physical meaning lies in using the consensus of neighboring observations to correct local subjective judgments, where Z is the normalization constant, is calculated as follows: This is used to ensure that the sum of the elements of the output vector is strictly equal to 1, thus conforming to the axioms of probability theory; Subsequently, to determine whether the algorithm has reached a steady state, the Euclidean distance drift between the confidence vector of the current iteration and the result of the previous iteration is calculated. The formula is ,like If the error is less than the preset stability threshold, the probabilistic graphical model is considered to have converged, the iteration is immediately stopped, and the current value is set to the minimum. The output is the final global congestion probability distribution data; otherwise, the current calculation result is broadcast to the neighbors as a new belief message to enter the next iteration.

[0028] The interaction of belief message data with neighboring intelligent agents within the preset communication range adopts an event-triggered mechanism based on data divergence, specifically including: Calculate the KL divergence between the belief message data generated at the current time and the belief message data generated at the previous time. The KL divergence value is compared with a preset message triggering threshold; If the KL divergence value is greater than the message triggering threshold, it is determined that the local state has changed significantly, and the broadcast operation of the belief message data is triggered. If the KL divergence value is less than or equal to the message trigger threshold, then the broadcast is suppressed and the belief state of the previous moment is maintained.

[0029] Specifically, the agent's local storage controller is invoked to retrieve historical belief message data successfully broadcast in the previous moment. (This data is the probability distribution vector from the previous time step, and the sum of its elements is 1), and the belief message data just generated at the current time step is obtained. To quantify the information difference between the old and new probability distributions, the processor performs a calculation using the following formula: In the formula, M represents the total number of congestion state levels, and k is the state index. and are the probability values ​​of the current and historical vectors in the k-th dimension, respectively, and ln is the natural logarithm. It is a very small smoothing constant, introduced to prevent mathematical singularities caused by zero denominators or zero logarithmic arguments. The calculated value is... This characterizes the amount of information loss that results from maintaining old beliefs without updating them; Then, the calculated The value is compared with the system's preset message trigger threshold, which is an empirical parameter used to balance communication overhead and the timeliness of status updates. If the comparison result shows If the message trigger threshold is reached, it is determined that the current node's state has undergone a statistically significant change. At this point, the system activates the wireless transmission module and immediately broadcasts the message. And simultaneously update local storage, Replace with This serves as the benchmark for the next comparison; otherwise, it is determined that the current information increment is insufficient to offset the communication energy consumption cost. The system then executes a communication suppression strategy, forcibly shutting down the current broadcast request and maintaining wireless silence, thereby achieving on-demand communication based on information value and significantly reducing the energy consumption of transmitting invalid and redundant data.

[0030] In response to a request for task data to be sent, the semantic features of the task data are extracted, and semantic information entropy and information timeliness values ​​are calculated by combining them with historical transmission records. The semantic information entropy and information timeliness values ​​are then weighted and fused to generate a communication value index, including: For the task data request to be sent, the local pre-built feature extraction algorithm is used to perform dimensionality reduction parsing on the task data and extract high-dimensional semantic feature vectors that represent the core content of the data. Obtain the predicted feature vector output by the locally maintained receiver state prediction model, calculate the Euclidean distance between the high-dimensional semantic feature vector and the predicted feature vector, and determine the calculated distance value as the semantic information entropy, which is used to characterize the amount of new information in the current data relative to the historical transmission records. Read the original time point when the task data request was generated, calculate the difference between the current system time and the original time point, and determine the difference as the information timeliness value; By using preset content weighting coefficients and timeliness weighting coefficients respectively, the semantic information entropy and information timeliness values ​​are weighted and summed to generate a dimensionless communication value index.

[0031] Specifically, a locally pre-built lightweight feature extraction network is used to perform dimensionality reduction analysis on the original high-dimensional data, extracting high-dimensional semantic feature vectors that can represent the core semantic content of the data. (Dimension N), and simultaneously calls the locally maintained receiver state prediction model, using the feature data successfully transmitted in the previous time step as input, to deduce and calculate the expected predicted feature vector of the receiver at the current time step. Next, in order to quantify the novelty of the current data relative to the receiver's existing knowledge, the semantic information entropy between the two vectors is calculated. Its physical essence is the Euclidean distance in the feature space, and the calculation formula is: In the formula and These are the numerical values ​​of the vector's components in the i-th dimension, and the calculation result is... The larger the value, the more difficult it is for the receiving end to predict the current data, meaning it contains a greater amount of information; simultaneously, the original time point when the task data request was generated is read. And obtain the processor's current system clock time. The difference between the two yields a value indicating the timeliness of the data. The unit is milliseconds; Finally, to address the issue that "semantic distance" and "time difference" have inconsistent dimensions and cannot be directly added, a normalized weighted fusion strategy based on the Sigmoid function is adopted to generate the communication value index. The fusion formula is ,in and These are the preset content weight coefficient and timeliness weight coefficient (and they must add up to 1). As a semantic sensitivity modulator, Let e ​​be the time decay constant and ε be the natural constant. This formula maps both indices to the dimensionless interval [0,1] and performs a linear combination, resulting in the final output. The higher the value, the more important the data is, as it contains important information unknown to the receiver and is in a fresh state with high urgency, and should be sent first.

[0032] A Lyapunov optimization model incorporating queue stability constraints and energy loss penalties is constructed. The local state feature vector, global congestion probability distribution data, and communication value index are input into the Lyapunov optimization model to solve for the optimal semantic compression coefficient and transmission backoff probability value of the current time slot.

[0033] like Figure 2 As shown, in the preferred technical solution of this application, a Lyapunov optimization model including queue stability constraints and energy loss penalties is constructed, including: Obtain the agent's own local state feature vector, which includes the task request queue length, remaining energy value, and signal-to-noise ratio data of the current communication link; The local state feature vector is analyzed, the task request queue length is extracted, a Lyapunov scalar function based on the square of the task request queue length is constructed, and the expected difference of the scalar function between the current time slot and the next time slot is calculated to generate a Lyapunov drift term that characterizes the system queue backlog trend. Analyze the global congestion probability distribution data, extract the average congestion probability value of the area where the current node is located, use this value to correct the preset unit bit transmission energy consumption, and construct a congestion penalty factor that reflects the severity of the network environment. Establish a system energy consumption function that includes two variables to be solved: semantic compression coefficient and transmission backoff probability value. Multiply the system energy consumption function by the congestion penalty factor to calculate the expected energy loss cost under the current network congestion state. A communication value index is introduced as a system benefit term, and a drift-penalty-benefit objective function is constructed. This objective function is formed by weighting and summing the Lyapunov drift term, the expected energy loss cost, and the negative communication value index through a preset trade-off control parameter. The trade-off control parameter is used to adjust the priority between queue stability and energy consumption benefit. The state that minimizes the drift-penalty-reward objective function is defined as the optimal solution of the model, thus completing the construction of the Lyapunov optimization model.

[0034] In a preferred embodiment of this application, global congestion probability distribution data is analyzed, the average congestion probability value of the region where the current node is located is extracted, and this value is used to correct the preset unit bit transmission energy consumption to construct a congestion penalty factor, including: Read the global congestion probability distribution data, which is a vector sequence containing probability values ​​for multiple congestion levels; The congestion level weight vector stored locally is called. The elements in this vector correspond one-to-one with the probability values ​​of the congestion level, and the values ​​increase stepwise as the congestion level deepens. The global congestion probability distribution data and the congestion level weight vector are multiplied by a dot product, that is, the probability values ​​of each congestion level are weighted and summed to calculate the average congestion probability value that represents the overall congestion level of the current network. Obtain a preset retransmission loss coefficient, which is used to quantify the proportion of additional energy consumption caused by packet retransmission due to channel congestion. Multiply the average congestion probability value by the retransmission loss coefficient, and add the product result to the value 1 to calculate the dimensionless congestion penalty factor. This factor is used as a multiplier parameter to make magnitude corrections to the unit bit transmission energy consumption under ideal conditions in subsequent steps, so as to reflect the energy consumption cost in the real network environment.

[0035] It should be added that the congestion level weight vector parameter is a constant real number vector, expressed in the form of: , where M is the total number of preset congestion levels.

[0036] Its numerical settings follow a non-linear increasing principle. Its preset logic is based on the "rate of degradation of network service quality as congestion intensifies." This is typically derived through prior offline simulations or actual measurements of average end-to-end latency under different congestion levels. For example, when the state is "idle," Set the baseline value to 1; when the state is "normal", due to a small number of collisions, Setting it to 1.5 (representing a 50% increase in latency); when the status is "congested," latency increases exponentially due to the triggered exponential backoff mechanism. It can be set to 4 or higher.

[0037] retransmission loss factor This is a dimensionless scalar constant, typically ranging from [0, 5]. Its physical meaning is to quantify the "extra energy cost incurred due to data packet retransmission caused by channel collisions." Its pre-defined logic is based on the "expected energy consumption ratio required for a single successful transmission." It is determined according to the retransmission mechanism of the MAC layer protocol used by the agent. The calculation logic is as follows: For example, if the hardware needs to increase transmission power or listen for a longer period of time during retransmission, then The corresponding increase; if the protocol stipulates a maximum number of retransmissions of 3, then the average retransmission energy consumption distribution is obtained through statistical experiments, thereby determining a fixed coefficient value that can reflect the average energy efficiency loss.

[0038] In this embodiment, the congestion level weight vector is pre-stored in a local configuration table. Its preset logic is based on quantifying the impact of different congestion levels on packet forwarding latency. For example, for the three levels of 'idle', 'medium', and 'congested', they are mapped to weights [1.0, 1.8, 4.5], respectively, reflecting the non-linear degradation effect of congestion on network performance. The retransmission loss coefficient is a constant reflecting the hardware energy efficiency characteristics. It is preset based on the ratio of retransmission power consumption to normal transmission power consumption after a channel collision. In this embodiment, through offline calibration experiments, this coefficient is set to 1.25 to characterize the expected energy premium per unit bit of data in a congested environment, thereby ensuring the physical accuracy of the energy loss penalty term in the subsequent Lyapunov optimization model.

[0039] Specifically, the local state feature vector generated in the preceding steps is analyzed, the normalized task request queue length is extracted, and this length is assigned to the queue state variable Q(t) of the current time slot. To establish a mathematical evaluation benchmark for system stability, a Lyapunov scalar function is defined. This is used to characterize the current extrusion potential energy of the system, and a Lyapunov drift term is introduced. To describe the trend of potential energy changing over time; Based on the 'drift-penalty' lemma, by... By performing inequality scaling, we obtain the expression for the drift term actually used in the calculation. ,in Let be the probability value of sending backoff, and be the variable to be solved. That is, the actual transmission probability. This is the maximum theoretical throughput constant of the link. This expression reveals that when the queue Q(t) is large, the transmission backoff probability can be reduced. It can generate a larger negative drift, thereby strongly suppressing queue growth and ensuring system stability; Next, the global congestion probability distribution data is analyzed, and its weighted average is calculated to obtain the average congestion probability value. This value is used to construct a congestion penalty factor. ,in This is the retransmission loss factor, usually taken as 1.5, which represents the rate at which congestion increases energy consumption. Subsequently, the system energy consumption function is established, and its calculation formula is as follows: ,in represents the semantic compression coefficient, and represents the variable to be solved, ranging from 0.1 to 1.0. Smaller values ​​result in higher compression ratios but may increase computational energy consumption. As a baseline transmission energy consumption per unit of data volume, this function accurately characterizes the expected energy loss cost generated by sending data with a specific compression rate and probability under the current congested environment. Finally, the communication value index generated in the preceding steps is introduced. Construct a comprehensive drift-penalty-reward objective function. The formula is In the formula To balance control parameters, the weights between "queue stability (drift term)" and "energy efficiency benefit (penalty benefit term)" are adjusted; the system defines the minimization of this binary function J as the process of finding the optimal solution, i.e., by adjusting... and This minimizes the objective function value, thereby maximizing the low-energy transmission of high-value information while preventing queue overflow.

[0040] In a preferred embodiment of this application, the local state feature vector, global congestion probability distribution data, and communication value index are input into a Lyapunov optimization model to solve for the optimal semantic compression coefficient and transmission backoff probability value of the current time slot, including: A Lyapunov drift penalty objective function is constructed that includes virtual queue stability constraints and energy loss penalties, and the communication value index is introduced into the objective function as a revenue coefficient; The local state feature vector is analyzed to obtain the current queue length and remaining energy. The global congestion probability distribution data is analyzed to obtain the average congestion level of the current network. The current queue length, remaining energy and average congestion level are substituted into the objective function to construct a binary optimization problem about the semantic compression coefficient and the transmission backoff probability value. The gradient descent method is used to iteratively solve the binary optimization problem, and the values ​​of the control variables that minimize the objective function are calculated. Boundary constraint mapping is applied to the solved control variable values ​​to restrict them to a preset effective range, and the optimal semantic compression coefficient and transmission backoff probability value for the current time slot are output respectively.

[0041] In a preferred embodiment of this application, the gradient descent method is used to iteratively solve the binary optimization problem, calculating the values ​​of the control variables that minimize the objective function, including: Read the preset initial iteration point values ​​from the local memory, and initialize the semantic compression coefficient and the transmission backoff probability value respectively, using them as the benchmark input variables for the first round of iteration; Based on the constructed drift-penalty-reward objective function, partial derivative operations are performed on the semantic compression coefficient and the transmission backoff probability value respectively. The partial derivative values ​​of the objective function at the current variable value point are calculated, and the two partial derivative values ​​are combined to form the gradient vector indicating the descent direction of the function value. Obtain the preset learning rate step size, which is used to control the magnitude of each iteration update. Calculate the product of the gradient vector and the learning rate step size. Subtract the corresponding component in the product from the current semantic compression coefficient and the send backoff probability value to generate the updated control variable value. This calculation process follows the principle of backpropagation to gradually approach the minimum point of the function. Calculate the Euclidean distance difference between the updated control variable value and the value of the previous iteration. Compare this difference with a preset convergence tolerance threshold. If the difference is less than the threshold, the iteration is considered to have converged, and the current value is output as the minimized control variable value. Otherwise, the updated value is fed back to the partial differential operation step, and the next round of gradient calculation and variable update is performed until the convergence condition is met.

[0042] Specifically, based on the drift-penalty-reward objective function, the initial iteration point value is read from the local register. and the preset learning rate step size And the convergence tolerance threshold ε; then enter the gradient descent iterative loop, for the binary optimization problem, calculate the partial derivatives with respect to the two decision variables to construct the gradient vector, where the semantic compression coefficient is used as the reference. The formula for calculating the partial derivative is as follows: This equation shows that the gradient direction of the compression coefficient is positively driven by the current transmission probability and energy consumption cost, for the transmission backoff probability value. The formula for calculating the partial derivative is as follows: This formula reveals that the gradient of the backoff probability is jointly constrained by the queue backlog Q(t), current energy consumption, and communication value. Next, the variable values ​​are updated using the backpropagation principle, and the update formula is as follows: This means subtracting the product of the learning rate and the corresponding gradient from the current variable value to approximate the extreme point along the direction of the steepest descent of the function. Due to the hard constraints of the physical system, a boundary constraint mapping operation must be performed after each update, that is, using the clipping function clip(·) to clip the updated value. Forced to be limited to the range [0.1, 1.0] (to avoid over-compression leading to distortion or coefficients greater than 1), Restricted to the interval [0,1], mathematically expressed as ; Finally, the Euclidean distance difference between the current iteration result and the previous iteration result is calculated. ,like If the value is less than the preset convergence tolerance threshold ε, the algorithm is considered to have converged, and the current value is output as the optimal solution; otherwise, the new value is carried over to the next iteration.

[0043] Based on the semantic compression coefficient, hierarchical feature extraction and encoding are performed on the task data to be sent to generate a target compressed data packet; the sending time point is determined based on the sending backoff probability value, and the target compressed data packet is sent to the receiving end.

[0044] In the preferred embodiment of this application, the logic for generating the target compressed data packet is as follows: Read the optimal semantic compression coefficient and use the linear mapping algorithm to convert the optimal semantic compression coefficient into the corresponding feature retention ratio and quantization bit depth. The feature retention ratio is used to determine the sparsity of subsequent feature extraction, and the quantization bit depth is used to determine the accuracy of data discretization. The task data to be sent is input into a local pre-built hierarchical semantic extraction network. The network outputs a high-dimensional feature tensor containing multiple abstraction levels. The high-dimensional feature tensor is filtered at the channel level according to the feature retention ratio. Core semantic feature channels with a total activation value greater than a preset threshold are retained, and redundant background noise channels are removed. The floating-point values ​​in the retained core semantic feature channels are discretized and quantized using quantization bit depth, and then converted into a fixed-length binary feature code stream. The binary feature stream is encapsulated and concatenated with a metadata header containing the current compression parameters to generate a structured target compressed data packet.

[0045] Specifically, the logic for generating the target compressed data packet first reads the optimal semantic compression coefficients output by the preceding Lyapunov optimization step. This coefficient, serving as the sole control input, is decomposed into two physical layer execution parameters using a linear mapping algorithm: feature retention ratio. With quantization bit depth The feature retention ratio is directly anchored to This is used to define the sparsity of semantic information, while the quantization bit depth is based on the formula... Perform the calculation, where and These are the highest and lowest precisions supported by the hardware codec, respectively. The round function returns a rounded integer in a specified format, which will be rounded to the nearest integer. Next, the raw task data to be sent is input into a locally pre-built hierarchical semantic extraction network. The intermediate layer of the network outputs a high-dimensional feature tensor with dimensions C×H×W, where C is the total number of channels. Subsequently, the system calculates the L1 norm of the activation matrix of each channel to characterize its information carrying capacity, and sorts the channels by importance, retaining only the top K core semantic feature channels. The remaining channels are discarded as redundant background noise, thus achieving primary compression in the feature dimension; H represents the height of the feature map. W represents the width of the feature map; Subsequently, the floating-point feature data within the retained core channels were discretized and quantized, and the calculated values ​​were used to perform the quantization process. Mapping consecutive floating-point values ​​to fixed-length integer indices, the quantization formula is as follows: ,in The original floating-point value. and This process converts data into a binary feature stream, defining the dynamic range boundary of the current feature channel. Finally, this binary feature stream is compared with the current... Bit-level concatenation of the metadata headers in the tensor dimension is performed to encapsulate and generate a structured target compressed data packet.

[0046] In a preferred embodiment of this application, determining the transmission time point based on the transmission backoff probability value includes: Read the transmission backoff probability value, call the local random number generator to generate a uniformly distributed random value between 0 and 1, compare the random value with the transmission backoff probability value, perform Bernoulli transmission decision, if the random value is greater than the transmission backoff probability value, it is determined that the current time meets the transmission condition, and the current time slot is marked as the immediate transmission time point; If the random value is less than or equal to the transmission backoff probability value, then an active backoff strategy is executed. The preset maximum contention window length is obtained, the product of the window length and the transmission backoff probability value is calculated and rounded up to obtain the number of backoff time slots, and the Nth time slot after the current time is marked as the delayed transmission time point, where N is equal to the number of backoff time slots. When the predetermined transmission time is reached, the wireless communication module is awakened to perform physical carrier sensing on the channel. If the channel is confirmed to be idle, the target compressed data packet is loaded into the radio frequency transmission buffer, converted into an electromagnetic wave signal by the baseband modulation circuit, and transmitted to the receiving end through the radio frequency antenna, thereby completing the data transmission process based on optimized probability control.

[0047] Specifically, the process of determining the transmission time point based on the transmission backoff probability value and sending the target compressed data packet to the receiving end first involves the communication control unit reading the transmission backoff probability value output by the preceding Lyapunov optimization step. The local hardware random number generator is invoked to generate a random value Rand that follows a uniform distribution in the range [0,1]. Then, the Bernoulli transmission decision is executed, with the logical discriminant being: If the inequality holds (i.e., Result is true), it means that the current random event falls within the transmission probability interval (because the actual transmission probability is...). The system determines that the immediate access conditions are met and marks the time slot corresponding to the current machine clock as the immediate transmission time point. Conversely, if This indicates that the system needs to perform proactive yielding to alleviate network congestion. In this case, the system reads the maximum contention window length preset by the MAC layer protocol. Using the formula The number of backoff slots was calculated, among which For the rounding up operation, the physical meaning of this formula is to linearly map the backoff probability to a specific waiting time; the higher the probability, the longer the wait, thus delaying the transmission time. ,in The physical duration is for a single time slot; when the system clock reaches a certain value... When the channel is idle, the radio frequency front-end is woken up to perform physical carrier sensing and detect whether the channel energy value is lower than the preset idle threshold. If the channel is idle, the target compressed data packet generated in the previous step is read from the buffer, digitally modulated and converted into an analog waveform through the baseband circuit, and finally transmitted to the receiving end through the power amplifier and antenna to complete a physical transmission precisely controlled by the probability model. Example 2:

[0048] like Figure 3 As shown, a multi-agent dynamic communication optimization device based on an energy state and probabilistic graphical model includes a processor, a memory, and a communication bus. The memory stores a computer-readable program that can be executed by the processor; The communication bus enables communication between the processor and the memory; When the processor executes the computer-readable program, it performs steps in a multi-agent dynamic communication optimization method based on an energy state and probabilistic graphical model as described in any one of the present invention.

[0049] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0050] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0051] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0053] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models, characterized in that, The method includes the following steps: The system acquires the agent's local state feature vector in real time; generates belief message data based on the local state feature vector and interacts with neighboring agents within a preset communication range; iteratively updates the locally stored factor graph model using the received neighbor belief message data and calculates the global congestion probability distribution data. In response to a request for task data to be sent, the semantic features of the task data are extracted, and the semantic information entropy and information timeliness values ​​are calculated in combination with historical transmission records; the semantic information entropy and information timeliness values ​​are weighted and fused to generate a communication value index. A Lyapunov optimization model containing queue stability constraints and energy loss penalties is constructed. The local state feature vector, global congestion probability distribution data and communication value index are input into the Lyapunov optimization model to solve for the optimal semantic compression coefficient and transmission backoff probability value of the current time slot. Based on the semantic compression coefficient, hierarchical feature extraction and encoding are performed on the task data to be sent to generate a target compressed data packet; the sending time point is determined based on the sending backoff probability value, and the target compressed data packet is sent to the receiving end.

2. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 1, characterized in that, Belief message data is generated based on local state feature vectors. The generation logic is as follows: Obtain the agent's own local state feature vector, which includes the task request queue length, remaining energy value, and signal-to-noise ratio data of the current communication link; The local state feature vector is analyzed, and the task request queue length and remaining energy value are extracted. The task request queue length and remaining energy value are normalized and weighted and summed to calculate the comprehensive load pressure index that reflects the current node's processing capacity. Based on the preset pressure-probability mapping model, the comprehensive load pressure index is converted into a local state probability distribution describing the possibility that a node is currently in different congestion levels. Extract the signal-to-noise ratio (SNR) data from the local state feature vector, and calculate the link confidence weight of the current communication link based on the SNR data; The link confidence weights are used to correct the magnitude of each probability value in the local state probability distribution to generate belief message data.

3. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 2, characterized in that, The global congestion probability distribution data was calculated, including: Receive neighbor belief message data from neighbor nodes within a preset communication range; input the neighbor belief message data into the locally stored factor graph model, extract the neighbor congestion probability information contained therein, perform message aggregation calculation with the belief message data, and generate node state confidence; Compare the currently calculated node state confidence with the confidence value calculated in the previous iteration. If the difference between the two is within the preset stable error range, the factor graph model is determined to have converged and the iteration is stopped. Extract the confidence scores of the finally converged node states, format and map them to generate global congestion probability distribution data.

4. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 1, characterized in that, Construct a Lyapunov optimization model that includes queue stability constraints and energy loss penalties, including: Obtain the agent's own local state feature vector, which includes the task request queue length, remaining energy value, and signal-to-noise ratio data of the current communication link; The local state feature vector is parsed, the task request queue length is extracted, a Lyapunov scalar function based on the square of the task request queue length is constructed, and the expected difference of the scalar function between the current time slot and the next time slot is calculated to generate the Lyapunov drift term. Analyze the global congestion probability distribution data, extract the average congestion probability value of the area where the current node is located, use this value to correct the preset unit bit transmission energy consumption, and construct the congestion penalty factor. Establish a system energy consumption function that includes two variables to be solved: semantic compression coefficient and transmission backoff probability value. Multiply the system energy consumption function by the congestion penalty factor to calculate the expected energy loss cost under the current network congestion state. The communication value index is introduced as a system revenue term, and a drift-penalty-revenue objective function is constructed. This objective function is formed by weighting and summing the Lyapunov drift term, the expected energy loss cost, and the negative communication value index through preset trade-off control parameters. The state that minimizes the drift-penalty-reward objective function is defined as the optimal solution of the model, thus completing the construction of the Lyapunov optimization model.

5. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 4, characterized in that, By inputting the local state feature vector, global congestion probability distribution data, and communication value index into the Lyapunov optimization model, the optimal semantic compression coefficient and transmission backoff probability value for the current time slot are obtained, including: A Lyapunov drift penalty objective function is constructed that includes virtual queue stability constraints and energy loss penalties, and the communication value index is introduced as a revenue coefficient into the drift-penalty-revenue objective function; The local state feature vector is analyzed to obtain the current queue length and remaining energy. The global congestion probability distribution data is analyzed to obtain the average congestion level of the current network. The current queue length, remaining energy and average congestion level are substituted into the drift-penalty-reward objective function to construct a binary optimization problem about the semantic compression coefficient and the sending backoff probability value. The gradient descent method is used to iteratively solve the binary optimization problem, and the values ​​of the control variables that minimize the objective function are calculated. Boundary constraint mapping is applied to the solved control variable values, and the optimal semantic compression coefficient and transmission backoff probability value are output as the current time slot, respectively.

6. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 4, characterized in that, Analyze the global congestion probability distribution data, extract the average congestion probability value of the region where the current node is located, use this value to correct the preset unit bit transmission energy consumption, and construct a congestion penalty factor, including: Read the global congestion probability distribution data, which is a vector sequence containing probability values ​​for multiple congestion levels; The congestion level weight vector stored locally is called. The elements in this vector correspond one-to-one with the congestion level probability values. The global congestion probability distribution data and the congestion level weight vector are multiplied by a dot product to calculate the average congestion probability value that represents the overall congestion level of the current network. Obtain the preset retransmission loss coefficient, multiply the average congestion probability value by the retransmission loss coefficient, and add the product result to the value 1 to calculate the dimensionless congestion penalty factor.

7. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 5, characterized in that, The binary optimization problem is solved iteratively using the gradient descent method, calculating the values ​​of the control variables that minimize the objective function, including: Read the preset initial iteration point values ​​from the local memory, and initialize the semantic compression coefficient and the transmission backoff probability value respectively, using them as the benchmark input variables for the first round of iteration; Based on the constructed drift-penalty-reward objective function, partial derivative operations are performed on the semantic compression coefficient and the transmission backoff probability value respectively. The partial derivative values ​​of the objective function at the current variable value point are calculated, and the two partial derivative values ​​are combined to form the gradient vector indicating the descent direction of the function value. Obtain the preset learning rate step size, calculate the product of the gradient vector and the learning rate step size, and subtract the corresponding component in the product from the current semantic compression coefficient and the transmission backoff probability value to generate the updated control variable value. Calculate the Euclidean distance difference between the updated control variable value and the value of the previous iteration. Compare this difference with a preset convergence tolerance threshold. If the difference is less than the threshold, the iteration is considered to have converged, and the current value is output as the minimized control variable value. Otherwise, the updated value is fed back to the partial differential operation step, and the next round of gradient calculation and variable update is performed until the convergence condition is met.

8. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 1, characterized in that, The logic for generating the target compressed data packet is as follows: Read the optimal semantic compression coefficient and use a linear mapping algorithm to convert the optimal semantic compression coefficient into the corresponding feature retention ratio and quantization bit depth; The task data to be sent is input into a local pre-built hierarchical semantic extraction network. The high-dimensional feature tensor is filtered at the channel level according to the feature retention ratio, and the core semantic feature channels with a total activation value greater than a preset threshold are retained. The floating-point values ​​in the retained core semantic feature channels are discretized and quantized using quantization bit depth, and then converted into a fixed-length binary feature code stream. The binary feature stream is encapsulated and concatenated with a metadata header containing the current compression parameters to generate a structured target compressed data packet.

9. The multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models according to claim 8, characterized in that, The transmission time point is determined based on the transmission backoff probability value, including: Read the transmission backoff probability value, call the local random number generator to generate a uniformly distributed random value between 0 and 1, compare the random value with the transmission backoff probability value, perform Bernoulli transmission decision, if the random value is greater than the transmission backoff probability value, it is determined that the current time meets the transmission condition, and the current time slot is marked as the immediate transmission time point; If the random value is less than or equal to the transmission backoff probability value, then an active backoff strategy is executed. The preset maximum contention window length is obtained, the product of the window length and the transmission backoff probability value is calculated and rounded up to obtain the number of backoff time slots, and the Nth time slot after the current time is marked as the delayed transmission time point, where N is equal to the number of backoff time slots. When the designated sending time arrives, the target compressed data packet is sent.

10. A multi-agent dynamic communication optimization device based on energy state and probabilistic graphical models, characterized in that, Includes processor, memory, and communication bus; The memory stores a computer-readable program that can be executed by the processor; The communication bus enables communication between the processor and the memory; When the processor executes the computer-readable program, it performs steps in the multi-agent dynamic communication optimization method based on energy state and probabilistic graphical models as described in any one of claims 1 to 9.

Images