Center-free distributed Gossip communication method

Abstract

The invention discloses a centerless distributed Gossip communication method, which is used for solving the problem that a traditional Push-Sum algorithm needs global random communication and is poor in engineering practicability. The method comprises the following steps: initializing local data and weight of each node; in each round of communication, the node receives data shares sent by all neighbors in the last round and divides the current total data of the node into k + 1 parts according to an equal proportion, k is the number of the neighbors, one part is reserved locally, and the other parts are sent to the neighbor nodes respectively. According to the method, the communication range is limited to neighbor nodes, mass conservation of data and weights is kept, efficient convergence to a global average value in an actual network topology can be achieved, and the method has high engineering applicability and higher convergence speed.

Description

Technical Field

[0001] This invention belongs to the field of multi-target tracking communication and relates to a decentralized distributed Gossip communication method. It is an engineering improvement of the push-sum communication algorithm, which is based on the uniform gossip communication method, so that it can be applied to communication in actual network topologies. Background Technology

[0002] Multi-target tracking technology is a well-established and highly regarded branch of information fusion, with a history spanning over seventy years. In the early 1970s, scholar Bar-Shalom combined Kalman filtering with data association techniques, achieving a key breakthrough in multi-target tracking. Since then, this technology has continued to attract in-depth research from academia and industry.

[0003] In practical tasks, the information provided by a single sensor is limited and often insufficient to handle complex multi-target tracking scenarios. Multi-sensor systems, through collaborative work, can acquire more comprehensive and accurate target information, thereby effectively improving the accuracy and robustness of the tracking system. In the early stages of multi-target tracking technology development, multi-sensor fusion generally adopted a centralized architecture, mainly including centralized, centralized-distributed, and hybrid architectures. Centralized architectures can achieve high-quality fusion results, but the fusion center bears a heavy computational load. Centralized-distributed architectures configure independent computing units at each node, allowing for preprocessing of sensor data, which reduces some of the burden on the fusion center but also leads to some information loss. Hybrid architectures flexibly combine the characteristics of both to balance their advantages. In recent years, with the rapid development of network communication and multi-agent theory, decentralized distributed fusion architectures have gradually gained widespread attention. This architecture eliminates the single fusion center, with each node having equal status. It can process data locally and also transmit and fuse information with others, thus exhibiting outstanding performance in terms of system scalability and robustness.

[0004] In decentralized distributed converged architectures, a crucial challenge is ensuring that each node obtains consistent multi-objective state estimation results; this is also known as consensus estimation. Especially in large network topologies with a vast number of nodes, any reliance on a central device for coordination limits system scalability. Gossip-based communication methods are emerging as an important communication paradigm in decentralized distributed converged architectures. In these methods, each node contacts one or a few nodes (usually randomly selected) in each round and exchanges information with them. The dynamic process of information propagation resembles the propagation of rumors, offering high fault tolerance and self-stability.

[0005] Compared to methods where all nodes in a network topology communicate simultaneously, such as flooding or diffusion mechanisms, gossip communication incurs only moderate communication overhead and can maintain overall operational stability even in cases of link instability or node failure. Furthermore, although gossip communication methods typically only guarantee convergence probabilistically, in large-scale network topologies, systems often focus more on the overall aggregation result of the network rather than the instantaneous state of individual nodes. Considering the generally limited communication bandwidth and energy resources in decentralized distributed networks, the aggregation computation process usually requires communication messages to be as concise as possible. Any communication protocol that attempts to centrally collect all local data at a single node may lead to communication bottlenecks or message implosion, thus affecting the overall system efficiency. It is against this backdrop that gossip communication methods have gradually become an important technical means in decentralized distributed convergence systems.

[0006] This invention focuses on the gossip communication problem in decentralized distributed networks. In this type of network topology, it is generally desired that through local information exchange between nodes, the network will eventually converge globally to the average of the initial data of each node, i.e., achieve average consensus. In existing research, uniform gossip is a commonly used theoretical communication model, which assumes that each node can uniformly and randomly select a node from all nodes in the global network to communicate in each round of communication. Based on this, a push-sum communication algorithm is proposed to calculate the average in decentralized distributed networks. This algorithm has good convergence analysis in mathematical theory and has been proven to converge to the correct average result in a probabilistic sense.

[0007] However, it should be noted that the communication assumptions upon which uniform gossip relies are often difficult to satisfy in real-world engineering systems. Specifically, the push-sum communication algorithm requires nodes to be able to arbitrarily choose communication targets within the global network, which implicitly assumes that nodes possess reachability to the entire network topology or global addressing capabilities. However, in practical multi-sensor networks, multi-target tracking systems, and multi-agent cooperative systems, nodes typically can only communicate with their physical or logical neighbors, and communication links are strictly limited by factors such as topology, communication radius, bandwidth, and energy constraints. Therefore, the assumption that nodes can randomly select any node in the global network for communication is often difficult to achieve in engineering practice, and is even unrealistic in most application scenarios.

[0008] Based on the above analysis, this invention addresses the limitations of the push-sum communication algorithm based on the uniform gossip communication model in practical applications by proposing an engineering-level improvement scheme. This scheme, while maintaining the core idea and mass conservation properties of the push-sum algorithm, relaxes the assumption of globally random communication, extending it to communication scenarios that conform to actual network topology constraints. This allows this type of gossip communication method to be better applied to practical decentralized distributed multi-target tracking systems. Summary of the Invention

[0009] The purpose of this invention is to improve the push-sum communication algorithm based on the uniform gossip communication method in a decentralized distributed network. It proposes a Gossip communication method for decentralized distributed networks, which improves the original push-sum communication algorithm while maintaining the inherent advantages of the gossip communication method, such as high fault tolerance and high scalability.

[0010] The technical solution adopted is as follows:

[0011] In a first aspect, the present invention provides a method for decentralized distributed Gossip communication, comprising:

[0012] Step S1: In the decentralized distributed network topology, each node device initializes its local data and weights. Each node device has at least one neighbor node and has the basic state to participate in Push-Sum communication.

[0013] Step S2: At the start of a communication round, each node device receives the communication data sent by all its neighboring nodes in the previous round.

[0014] Step S3, the sending phase: Each node device, based on its own local data allocation share and the total received communication data, sends the remaining allocated data to all neighboring nodes proportionally.

[0015] Preferably, the initialization operation in step S1 specifically involves:

[0016] Local measurements of each node device i (i.e., the initial local data, also known as observations), its weights Always set to 1, the initial total data is Node devices In the The communication data received within a communication round is represented as follows: , This represents all the data received by node device i from all its neighboring nodes. This represents all the weights sent by all neighboring nodes to node device i.

[0017] Preferably, step S3 specifically involves:

[0018] Node devices Total data including local data and received communication data The data is divided into corresponding shares, and while retaining the corresponding shares locally, the corresponding shares are also sent to all neighboring nodes:

[0019]

[0020]

[0021]

[0022]

[0023] in This indicates that a corresponding share of the data is retained locally. Represents the local splitting factor, 0 < <1, express The weight, This represents the share data that neighbor node j receives from node device i. , This represents the set of neighboring nodes of node device i. express The weight, Represents the splitting factor of neighbor node j, 0 < <1.

[0024] Preferably, the splitting factor satisfies the following condition:

[0025]

[0026] Preferably, the splitting factor satisfies the following condition:

[0027]

[0028] Preferably, the decentralized distributed network topology where the node device is located is a connected graph, and as the communication rounds iterate, the local data of each node device in the network topology tends to the global average value.

[0029] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described thereon.

[0030] Thirdly, the present invention provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described thereon.

[0031] The beneficial effects of this invention are at least as follows:

[0032] This invention limits nodes to communicating only with their neighboring nodes and introduces a share allocation mechanism based on neighbor sets. This allows it to be directly applied to topology-constrained real-world network environments, thereby overcoming the dependence of the original push-sum communication algorithm on global random communication and significantly enhancing the engineering feasibility of the method.

[0033] This invention, while broadening the communication range, retains the core "data-weight" bivariate structure and mass conservation property of the Push-Sum algorithm. By setting a splitting factor, it ensures that the sum of data and the sum of weights remain constant in the network during each communication round. Under a connected graph topology, the estimated values ​​of each node can still converge to the global average, mathematically guaranteeing correctness and convergence reliability.

[0034] This invention distributes data simultaneously to all neighboring nodes in each transmission phase, employing an equal-split strategy. This approach allows information to spread in parallel along multiple paths within the network, improving the coverage speed and uniformity of information propagation, thereby achieving faster convergence within the same number of communication rounds.

[0035] Since each node operates solely based on its local neighbor information, without a central coordinating node, the system exhibits typical advantages of a decentralized distributed system. Even if some nodes or links fail, the remaining nodes can continue executing the algorithm through neighbor communication, maintaining the inherent high fault tolerance and self-organizing capabilities of Gossip communication. Furthermore, nodes only need to maintain their neighbor sets, without requiring global topology information, facilitating dynamic network expansion and large-scale deployment.

[0036] This invention involves only a limited number of communications between local neighbors in each round, and the message content consists only of data and weight share values. It has low communication overhead and small bandwidth consumption, making it suitable for distributed network environments where energy and communication resources are limited. Attached Figure Description

[0037] To more clearly illustrate the technical solution of the present invention, the drawings used in 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.

[0038] Figure 1This is a flowchart of a decentralized distributed Gossip communication method provided in an embodiment of the present invention.

[0039] Figure 2 It is a network topology diagram oriented towards a decentralized distributed system.

[0040] Figure 3 This is an experimental simulation diagram of the Gossip communication method for a decentralized distributed system provided in this embodiment of the invention. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0042] To clearly describe the improved push-sum communication algorithm of this invention, the uniform gossip communication method and the original push-sum communication algorithm based on it are introduced here, and a mathematical proof is given that it can converge to the global average.

[0043] In decentralized distributed network topologies, the gossip communication model is becoming a popular communication paradigm due to its advantages such as high fault tolerance and high scalability. Uniform gossip is one such model. In the uniform gossip communication method, each node in the network topology can randomly select a node in the global network topology to communicate with. The two communicating nodes can share data information with each other. After several rounds of communication, each node in the network topology can reach a data consensus with a high probability, converging to the global average value.

[0044] The Push-sum communication algorithm is a specific communication algorithm based on the uniform gossip communication method. Its specific process is as follows:

[0045] As Figure 2 As shown, a decentralized distributed network has several nodes, each with equal status and possessing the functions of sensing and detection, communication and transmission, and fusion computing. For clarity, the network will be modeled using mathematical notation. Assume there are a total of [number missing] nodes in the decentralized distributed network topology. Each node. During the initialization phase, each node in the network topology... Has local values and weight The weight value is always kept at 1. Node The goal is to calculate the aggregated value after communication, which in this case is the average. Node Use locally This represents the total amount of data received, where the subscripts indicate the communication round number and the node index, respectively. For nodes... In the initialization phase, the local data of a node can be represented as:

[0046]

[0047] In the In each communication round, the node First, receive the data sent to it in the previous communication round, then update the local data.

[0048]

[0049]

[0050] in, Indicates the previous round to the node The set of nodes that send information. This represents the set of data sent.

[0051] Then the node A node is randomly selected from the global network topology according to a uniform distribution for communication. The local data information, including the value and weight, is divided into two parts and sent to the selected communication node. The sent data can be represented as:

[0052]

[0053]

[0054] The node's estimate of the global average can be expressed as:

[0055] With each iteration of the communication rounds, it can be mathematically proven that all... It can converge to with a high probability ,in Let represent the index of all nodes in the network topology. The proof that it converges to a uniform (i.e., global average) is as follows:

[0056] In any communication round Within this system, the sum of data and its weights always satisfy the law of conservation of mass, that is:

[0057]

[0058]

[0059] In each communication round, the node The local data will be divided equally, with half kept locally and the other half sent to the communication nodes. As the above formula shows, this means that the node... The quality remains consistent before and after transmission; the old value is simply rewritten as the sum of all the data shares received.

[0060] The update after each communication round can be represented as:

[0061]

[0062] in, It is a column random matrix that satisfies Because the first matrix Column corresponding node The output data, nodes Divide all the data into two, so the column sum is 1. Based on the properties of column random matrices, it can be proven that sum is conserved throughout the process:

[0063] therefore,

[0064]

[0065] In summary:

[0066]

[0067] Similarly, the conservation of weights can also be proven:

[0068]

[0069]

[0070] This embodiment provides a decentralized distributed gossip communication method, which improves upon the original push_sum communication algorithm. It ensures that each node in the network topology can only communicate with its neighbors, making the algorithm more suitable for real-world applications. Furthermore, within each communication round, each node divides its data into k+1 equal parts, where k is the number of its neighbors, effectively changing from single-target transmission to multi-target transmission.

[0071] To describe this more accurately, we model the network topology: a network is a connected graph. ,node The neighbor set is The specific implementation steps of the improved push-sum communication method include:

[0072] Step (1): In the decentralized distributed network topology, each node device initializes its local data and weights. Each node device has at least one neighbor node and has the basic state to participate in Push-Sum communication.

[0073] Step (2): At the start of a communication round, each node device receives the communication data sent by all its neighboring nodes in the previous round.

[0074] Step (3), the sending phase, each node device, based on its own local data allocation share and the total data of the received communication data, sends the remaining allocated data to all neighboring nodes in proportion.

[0075] Step (4): Under the premise of a connected graph, as the communication rounds iterate, the local data of each node device in the network topology will tend to the global average value.

[0076] More specifically, step (1) is as follows:

[0077] In a decentralized distributed network topology, each node A local data initialization must be performed. This step aims to ensure that each node has state variables in a consistent format that can participate in distributed averaging. Each node The data variables that need to be initialized are: the node's local measurement values. This is the data that participates in the global average across the entire network; each node also maintains its own weights locally. The weight is always set to 1 to ensure that the average converges. Therefore, the initial sum of the data is... .node All neighbor nodes can be represented by a set This indicates that the node In the All data received within a communication round can be represented as . This represents all the data received by node device i from all its neighboring nodes. This represents all the weights received by node i from all its neighboring nodes. At this point, each node has the basic state to participate in Push-Sum communication, and the network topology is ready to enter iterative communication.

[0078] More specifically, step (2) is:

[0079] In the At the start of a round of communication, the node The previous round of transmissions from all neighbors must be aggregated for the node. It will come from all its neighboring nodes Received the data they sent in the previous round These data are the split shares. Within the push-sum communication algorithm, nodes... It doesn't distinguish where the data comes from; it simply sums all of it.

[0080]

[0081]

[0082] After this phase ends, the node Having the data status after this round of integration .

[0083] More specifically, step (3) is as follows:

[0084] After receiving data from all neighboring nodes, the node A decision needs to be made on how to split this data into multiple shares for transmission. The key aspect of this invention lies in the node. Instead of sending a copy to just one neighbor node, this invention sends a copy to each neighbor node while reserving a share for the local node. Therefore, this invention introduces a splitting factor. ,in The splitting factor satisfies:

[0085]

[0086] Splitting factors ensure that no node loses all weights in each round, thus avoiding a weight of 0, and also ensure that the data quality is conserved throughout the process, since all shares add up to 1.

[0087] In this invention, the splitting factor is:

[0088] That is, all the data is equally distributed among the node itself and all its neighboring nodes.

[0089] During the data transmission phase of each communication round, the node... Total data including local data and received communication data , This represents the total data of node device i, including local data and received communication data, during communication phase t. To correspond to the weights, the above data is divided into corresponding shares and sent: the corresponding share is retained locally and also sent to all neighboring nodes.

[0090]

[0091]

[0092]

[0093]

[0094] in This indicates that a corresponding share of the data is retained locally. Represents the local splitting factor, 0 < <1, express The weight, This represents the share data that neighbor node j receives from node device i. , This represents the set of neighboring nodes of node device i. express The weight, Represents the splitting factor of neighbor node j, 0 < <1;

[0095] More specifically, step (4) is:

[0096] Assuming network topology connectivity and data quality conservation, continuous communication iterations will eventually lead to a more uniform distribution of the contribution vectors of all nodes. The estimated value for each node can be expressed as... They will tend to be consistent and converge to the global average:

[0097]

[0098] As the number of communication rounds increases, data gradually mixes within the network topology. A larger share of message transmission leads to faster data diffusion, but also increases the bandwidth requirements of the network topology. Furthermore, the more nodes and the sparser the network topology graph, the more communication rounds are needed to converge to consistency. Ultimately, due to the conservation of data quality, the convergence result will be correct and will converge to the global average.

[0099] In the simulation, this invention constructs a decentralized distributed network topology graph, which is a randomly connected undirected graph with a total of 10 nodes. This embodiment uses a modified push-sum communication algorithm, meaning that each node can only communicate with its neighbors. As... Figure 3 As shown, all nodes have different initial values ​​(observations) during initialization. As the number of communication rounds increases, the data curves of all nodes gradually converge, eventually approaching the same horizontal line. The simulation data is shown in the table below (only the communication data of the first 10 rounds are shown in the table).

[0100] Table 1 Simulation Data

[0101] Iteration Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 8 Node 9 Node 10 1 7.81 6.91 13.73 10.83 8.00 11.67 12.35 7.42 10.62 12.86 2 10.07 10.13 12.97 8.48 8.99 9.53 11.33 10.72 11.57 9.87 3 9.99 9.40 11.42 9.78 9.77 10.56 11.12 9.65 10.82 10.57 4 10.37 10.02 11.04 9.61 9.93 10.26 10.79 10.14 10.83 10.17 5 10.34 10.00 10.66 9.92 10.12 10.40 10.66 10.04 10.63 10.26 6 10.38 10.14 10.53 9.99 10.18 10.34 10.54 10.16 10.56 10.22 7 10.36 10.17 10.43 10.09 10.23 10.35 10.47 10.18 10.48 10.24 8 10.35 10.22 10.38 10.15 10.25 10.34 10.42 10.22 10.43 10.25 9 10.34 10.24 10.35 10.19 10.27 10.33 10.38 10.24 10.40 10.26 10 10.33 10.26 10.33 10.22 10.28 10.32 10.36 10.26 10.37 10.27

[0102] The data above clearly shows that all nodes initially have completely different observations, but as the number of communication rounds increases, they gradually converge to the same value, and precisely and asymptotically approximate the true average value of 10.3. The data supports the conclusion that the improved parallel neighbor segmentation Push-Sum method of this invention still maintains the mass conservation property and enables the estimated values ​​of all nodes in the network to converge to the global average value, thus verifying the correctness and feasibility of the proposed engineering improvement method.

[0103] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A Gossip communication method for decentralized distributed systems, characterized in that, The method includes: Step S1: In the decentralized distributed network topology, each node device initializes its local data and weights. Each node device has at least one neighbor node and has the basic state to participate in Push-Sum communication. Step S2: At the start of a communication round, each node device receives the communication data sent by all its neighboring nodes in the previous round. Step S3, the sending phase: Each node device, based on its own local data allocation share and the total received communication data, sends the remaining allocated data to all neighboring nodes proportionally.

2. The method according to claim 1, characterized in that, The initialization operation in step S1 is as follows: Local measurements of each node device i Its weight Always set to 1, the initial total data is Node devices In the The communication data received within a communication round is represented as follows: , This represents all the data received by node device i from all its neighboring nodes. This represents all the weights sent by all neighboring nodes to node device i.

3. The method according to claim 1, characterized in that, Step S3 is as follows: Node devices Total data including local data and received communication data The data is divided into corresponding shares, and while retaining the corresponding shares locally, the corresponding shares are also sent to all neighboring nodes: ； ； ； ； in This indicates that a corresponding share of the data is retained locally. Represents the local splitting factor, 0 < <1, express The weight, This represents the share data that neighbor node j receives from node device i. , This represents the set of neighboring nodes of node device i. express The weight, Represents the splitting factor of neighbor node j, 0 < <1.

4. The method according to claim 3, characterized in that, The splitting factor satisfies the following conditions: 。 5. The method according to claim 4, characterized in that, The splitting factor satisfies the following conditions: 。 6. The method according to claim 4, characterized in that, The decentralized distributed network topology where the node devices reside is a connected graph. As communication rounds iterate, the local data of each node device in the network topology tends to the global average value.

7. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-6.

8. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-6.

Images