A method for Boltzmann-based model neural network prediction
By layering the network topology of the Boltzmann machine model and applying a log server, the problem of low processing efficiency for large log data volumes in large systems was solved, realizing efficient data processing and the application of artificial intelligence in the field of big data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA TELECOM DIGITAL INTELLIGENCE TECH CO LTD
- Filing Date
- 2023-09-15
- Publication Date
- 2026-06-05
AI Technical Summary
The existing Boltzmann machine model network topology lacks a clear hierarchy, resulting in low efficiency when processing massive log data from large systems.
By deploying a log server, the network topology of the Boltzmann machine model is processed in layers. The logs of the log server are used to mine the relationships, generate the first and second matrices, and put them into the Markov chain for prediction. Asynchronous random updates are performed in combination with the state update rules of the neuron nodes.
When processing massive log data from large systems, it improves the timeliness of data processing, highlights the role of artificial intelligence in big data processing, and effectively makes up for the shortcomings of the Boltzmann machine network topology.
Smart Images
Figure CN117195952B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of neural network prediction, and in particular relates to a method for prediction based on a Boltzmann model neural network. Background Technology
[0002] The Boltzmann machine model was proposed by Hinton et al. in 1984. In fact, applying a random perturbation mechanism to the Hopfield model can form a Boltzmann machine model. The Boltzmann machine introduces the Boltzmann distribution probability from statistical physics and employs the so-called simulated annealing algorithm, enabling the model to escape local minima during the optimization process and thus find the optimal solution globally. For solving optimization problems, the Boltzmann machine's operation is essentially an energy optimization process. Each point in the solution space represents a solution, and different solutions have different cost function (i.e., energy function) values. Optimization is about finding the minimum (or maximum) solution of the cost function in the solution space. Applying the simulated annealing method to the Boltzmann machine's analysis process yields the globally optimal energy solution. However, compared to general feedforward networks, existing Boltzmann machine network topologies lack a clear hierarchical structure. Summary of the Invention
[0003] To address the above problems, this invention provides the following solution: a method for prediction using a Boltzmann-based model neural network, comprising:
[0004] A stochastic neural network with symmetric connection weights is created using a Boltzmann machine model, and a log server is deployed on the stochastic neural network.
[0005] The network topology of the Boltzmann machine is layered using the logs of the log server to obtain a first matrix and a second matrix.
[0006] The first and second matrices are respectively placed into the Markov chain of the Boltzmann machine state transition for prediction. If the difference between the predicted values is less than 10% weighted average, the matrix prediction value is the final prediction value; if the difference between the predicted values is greater than 10% weighted average, the matrix data with the higher proportion of core business data is taken as the final prediction value.
[0007] Preferably, the neuron node of the random neural network includes a first state where the neuron outputs v=0 and a second state where the neuron outputs v=1;
[0008] When the activation function value of a neuron changes, it causes a state update of the neuron node. This state update is asynchronous and random among the neuron nodes.
[0009] Preferably, when any neuron node i updates its state, the probability that the next state is 1 is:
[0010]
[0011] Where T represents the temperature parameter of the network, and takes a positive value;
[0012] A i The activation function for node i is expressed by the formula:
[0013] ;
[0014] The probability that the next state is 0 is:
[0015]
[0016] Among them, w ij v represents the connection weight between the i-th neuron and the j-th neuron. j θ is the input to neuron j. i This represents the neuron threshold.
[0017] Preferably, the hierarchical processing of the Boltzmann machine's network topology using the logs of the log server includes:
[0018] By leveraging the relationships between application data from the log server, the application service topology is mined and correlated with the network topology to obtain the first matrix.
[0019] Preferably, the application service topology is used to record the relationships between different servers and different types of indicators; it is also used to actively / passively monitor important information flows / connections / links and obtain network resource status.
[0020] Preferably, the category indicators include services, applications, and boards;
[0021] The board includes a server CPU, memory, and disk.
[0022] Preferably, the hierarchical processing of the Boltzmann machine's network topology using the logs of the log server further includes,
[0023] Mine a data set based on the application service topology from the logs of each service to obtain a topology data set;
[0024] The operation and maintenance status of the network units is obtained based on the topology data set, and a second matrix is generated based on the operation and maintenance status of the network units.
[0025] Preferably, the topology data set is used to record monitoring data and operational health status of servers and databases, boards and middleware, server boards and boards, and boards and boards at different time dimensions.
[0026] Preferably, the first matrix and the second matrix are respectively put into the Markov chain of the Boltzmann machine state transition for prediction, specifically through the one-step transition probability matrix of the Markov chain.
[0027] The formula for the one-step transition probability matrix of the Markov chain is as follows:
[0028]
[0029] In the formula, consider a set of states S0, S1, ..., Sm1 of a given system. These states have known transition probabilities p (representing the probability of state S1 appearing after state S1) and appear one after another. The state changes of the system are represented by an m×m matrix composed of p, ∑pB=1, p1≥0; i, j=1,2,...,m-1.
[0030] Compared with the prior art, the present invention has the following advantages and technical effects:
[0031] The Boltzmann-based neural network prediction method provided in this invention can prevent situations where the daily log data volume can reach tens of tB (TBs) in large systems providing global services by deploying log servers. The innovative use of hierarchical log processing compensates for the lack of clear hierarchy in the Boltzmann machine's network topology, ensuring both timeliness of data processing and highlighting the role of artificial intelligence in big data processing when handling massive amounts of data in a network. Attached Figure Description
[0032] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0033] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;
[0034] Figure 2 This is a schematic diagram of the structure of a three-node Boltzmann machine according to an embodiment of the present invention. Detailed Implementation
[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0036] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0037] like Figure 1 As shown, the present invention provides a method for prediction using a Boltzmann-based model neural network, comprising the following steps:
[0038] Step 1: First, a stochastic neural network with symmetric connection weights is created using the Boltzmann machine model, and a log server is deployed on this stochastic neural network. Since the Boltzmann machine network topology does not have a clear hierarchy compared to a typical feedforward network, deploying a log server can prevent situations where the daily log data volume can reach tens of terabytes (TBs) in some large systems providing global services. The received data is then analyzed in two layers, as follows:
[0039] The first layer utilizes the correlation relationships of application data from log services to mine the application service topology and correlates it with the network topology to obtain the first matrix. The application service topology records the relationships between different servers and three types of metrics: business, application, and network card. Simultaneously, it actively / passively monitors important information flows / connections / links, striving to accurately obtain network resource status while reducing network monitoring overhead. The network devices consist of multiple network cards, each including a server CPU, memory, and disk.
[0040] The second layer extracts a data set based on the application service topology relationships from the logs of each service; this is called the topology data set. The topology data set records monitoring data and operational health status across different time dimensions, including between servers and databases, boards and middleware, and between servers and boards. Then, based on the topology data set, the operational status of the network units is obtained, and a second matrix is generated according to this status.
[0041] Further optimization of the scheme, a detailed description of the technical principles of the Boltzmann machine:
[0042] A Boltzmann machine is a stochastic neural network with symmetric connection weights. Each neuron has two states: its output v is either v=0 or v=1, hence it is called a binary neuron. When the activation function value of a neuron changes, it causes a node state update. This update is asynchronous and random among nodes. When any node i updates its state, the probability that the next state is 1 is...
[0043]
[0044] Where T represents the temperature parameter of the network, and takes a positive value;
[0045] A i Let represent the activation function of node i, and let represent the sigmoid function, which is the Boltzmann probability function. The formula is:
[0046]
[0047] The probability that the next state is 0 is:
[0048]
[0049] Generally, when the activation function When the value is greater than 0, the probability of the next state being 1 is... (1) The probability p(0) is greater than the probability that the next state is 0, and as A i As the value increases, p(1) also increases. Similarly, when When <0, (0)> (1), and with The decrease in value increases the probability of the state being 0. (0) Increase.
[0050] Furthermore, the curvature of the probability distribution curve is related to the temperature T; the higher the temperature, the flatter the curve, and the easier the state change. Conversely, the lower the temperature, the steeper the curve, and the more difficult the state change. Especially when the temperature approaches 0 (T→0), the probability distribution curve approximates a unit step function. The activation function properties of the unit are essentially described by this probability function, since p(T=0) is essentially equivalent to a threshold function. In this case, the Boltzmann machine is equivalent to a discrete Hopfield network.
[0051] The following example illustrates the impact of temperature changes on network state transition relationships.
[0052] Suppose a three-node Boltzmann machine, such as Figure 2 As shown. The network parameters are:
[0053] = =0, = =-0.7,
[0054]
[0055] Determine the state transition relationships of the network at temperatures T=0.5 and T=1.
[0056] First, calculate the activation function values of each node unit in a given state. (Based on the state...) Taking this as an example, the state update probability of each node in each state is calculated sequentially when the temperature T=0.5 and T=1. The results are shown in Table 1. The transition probability of each state is calculated.
[0057] Table 1
[0058]
[0059] Each state can transition to one of the other three corresponding states. Because each node has the same probability of state change at any given time, the probability of transitioning from state (011) to (111) is p1(1) / 3. In summary, if a state is caused by the activation change of the i-th node, then when v=1 in the state, the total probability of reaching this state is p(1) / 3; when v=0 in the state, the probability of reaching this state is p(0) / 3.
[0060] The state transition probability can be expressed by a unified formula as:
[0061]
[0062] The probability that the state remains unchanged can be calculated using the following formula.
[0063]
[0064] Based on the two formulas above, the probability of each state transitioning to other states at different temperatures can be determined. It is evident that the introduction of temperature alters the transition probability distribution of the (011) state in the original HNN, allowing it not only to transition to lower energy states but also to higher energy states. At higher temperatures, the probability of transitioning to other states approaches 1 / 6.
[0065] Step 2: Obtain the logs from the previous layering step and generate two matrices. These matrices are then fed into a Boltzmann machine state transition Markov chain for prediction. If the difference between the predicted values is less than 10% and weighted averaged, the two matrices are used as the final predicted value. Otherwise, the matrix with the higher proportion of core business data is used as the final predicted value.
[0066] Describing the state transition relationships of a network graphically is complex and impractical. This embodiment uses Markov chains from stochastic processes to express this relationship. A Markov chain is a Markov process where both time and state are discrete. Consider a set of states S0, S1, ..., S2 of a given system. m-1 These states appear one after another with known transition probabilities p (representing the probability that state S1 will occur after state S1). The state changes of the system can be represented by an m×m matrix of p:
[0067]
[0068] In the formula, ∑p ij =1, p ij ≥0; i, j=0,1,2,...,m-1, this is a non-negative element matrix in which the sum of the elements in each row is 1, called the one-step transition probability matrix of the Markov chain.
[0069] The Boltzmann-based neural network prediction method provided in this invention can prevent situations where the daily log data volume can reach tens of tB (TBs) in large systems providing global services by deploying log servers. The innovative use of hierarchical log processing compensates for the lack of clear hierarchy in the Boltzmann machine's network topology, ensuring both timeliness of data processing and highlighting the role of artificial intelligence in big data processing when handling massive amounts of data in a network.
[0070] The above are merely preferred embodiments 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.
Claims
1. A method for prediction using a Boltzmann-based model neural network, characterized in that, include: A stochastic neural network with symmetric connection weights is created using a Boltzmann machine model, and a log server is deployed on the stochastic neural network. The network topology of the Boltzmann machine is layered using the logs of the log server to obtain a first matrix and a second matrix. The first and second matrices are respectively placed into the Markov chain of the Boltzmann machine state transition for prediction. If the difference between the predicted values is less than 10% weighted average, the matrix prediction value is the final prediction value; if the difference between the predicted values is greater than 10% weighted average, the matrix data with the higher proportion of core business data is taken as the final prediction value. The layering of the Boltzmann machine's network topology using the logs from the log server includes, By leveraging the relationships between application data from the log server, the application service topology is mined and correlated with the network topology to obtain the first matrix. The application service topology is used to record the relationships between different servers and different types of indicators; it is also used to actively / passively monitor important information flows / connections / links and obtain network resource status. Using the logs of the log server to perform layered processing of the Boltzmann machine's network topology also includes, Mine a data set based on the application service topology from the logs of each service to obtain a topology data set; The operation and maintenance status of the network units is obtained based on the topology data set, and a second matrix is generated based on the operation and maintenance status of the network units; The topology data set is used to record monitoring data and operational health status of servers and databases, boards and middleware, servers and boards, and boards at different time dimensions.
2. The method for prediction based on a Boltzmann-based model neural network according to claim 1, characterized in that, The neurons of the random neural network include a first state where the neuron outputs v=0 and a second state where the neuron outputs v=1. When the activation function value of a neuron changes, it causes a state update of the neuron node. This state update is asynchronous and random among the neuron nodes.
3. The method for prediction based on a Boltzmann-based model neural network according to claim 2, characterized in that, When any neuron node i updates its state, the probability that the next state is 1 is: Where T represents the temperature parameter of the network, and takes a positive value; A i The activation function for node i is expressed by the formula: ； The probability that the next state is 0 is: Among them, w ij v represents the connection weight between the i-th neuron and the j-th neuron. j θ is the input to neuron j. i This represents the neuron threshold.
4. The method for prediction based on a Boltzmann-based model neural network according to claim 1, characterized in that, The categories of indicators include business, application, and board; The board includes a server CPU, memory, and disk.
5. The method for prediction based on a Boltzmann-based model neural network according to claim 1, characterized in that, The first and second matrices are respectively put into the Markov chain of the Boltzmann machine state transition for prediction, specifically through the one-step transition probability matrix of the Markov chain. The formula for the one-step transition probability matrix of the Markov chain is as follows: In the formula, consider a set of states S0, S1, ..., S2 of a given system. m1 These states appear one after another with known transition probabilities p, where p represents the probability of state S1 occurring after state S1. The state changes of the system are represented by an m×m matrix of p, ∑p ij =1, p ij ≥0;i,j=0,1,2,...,m-1.