Intention-driven self-intelligence network on-demand service method and system

Abstract

The invention discloses an intention-driven self-intelligence network on-demand service method and system. The method comprises the following steps: acquiring a user intention; generating a configuration generation large model, and translating a user intention into a configuration command template sequence which is not filled with parameters; obtaining a network configuration document, generating a large model correction template sequence through configuration, and filling values for semantic-level parameters; setting a plurality of service quality indexes, generating a service quality index weight matrix by the configuration generation large model according to the intention of a user, then constructing an expert model for each service quality index, calculating a resource allocation weight vector of the index-level parameter under the corresponding service quality index by the expert model, and calculating the resource allocation weight vector of the index-level parameter; and finally, according to the weight matrix and the resource allocation weight vector of the expert model, determining index-level parameter values, and filling the index-level parameter values into the template sequence. And outputting the template sequences to network equipment one by one for execution. The invention relates to the field of information communication networks, can accurately translate user intentions into configuration commands from the semantic level, and dynamically generates the configuration commands meeting user requirements.

Description

Technical Field

[0001] The present invention relates to an intention-driven self-intelligent network on-demand service method and system, and relates to the field of information communication networks. Background Art

[0002] In the grand vision of 6G networks, meeting highly flexible and diverse business needs and application scenarios has become an inevitable requirement for the evolution of future wireless networks. With the development of network intelligence, the concept of self-intelligent networks has emerged. Self-intelligent networks use artificial intelligence, machine learning and automation technologies to achieve the ability of "self-configuration, self-optimization, self-repair and self-protection". It can dynamically adjust and optimize network operations to adapt to changing needs and conditions without human intervention. In addition, self-intelligent networks also introduce the concept of "intent", which is an abstract form of goals, needs or expectations expressed by users or systems at a high level. By analyzing intent, self-intelligent networks can efficiently regulate network operations, thereby significantly improving network efficiency, reliability and responsiveness, while reducing operating costs and complexity.

[0003] However, the key to achieving this goal is to realize the intent translation function in the self-intelligent network. Intent translation is the process of parsing the high-level goals of users or businesses into specific technical implementations. Through intent translation, abstract intent can be converted into actual network configuration and operation, thereby effectively regulating network resources and meeting diverse business needs. Existing solutions generally train deep learning models, input intent requirements into the model for translation, generate network configuration strategies, and optimize the strategies in combination with resource allocation algorithms and issue them for execution.

[0004] Due to the limitations of natural language processing technology in the past, intent translation is often not accurate enough, especially in generating configuration commands based on user intent at the semantic level. Usually, intent can only be mapped to a fixed translation template. This problem directly limits the application of self-intelligent networks. Traditional deep learning models usually lack the ability to understand and generate long texts and can only complete relatively simple tasks such as entity extraction and template matching. This limitation makes it impossible to meet the requirements of self-intelligent networks for fine-grained intent understanding, configuration generation, and resource allocation.

[0005] In the process of intent translation, the manual of the network equipment is not effectively utilized, which may lead to erroneous results due to the complex relationship between the parameters in the configuration command and their corresponding values. For example, the network resource status does not match the value range specified in the manual. In addition, in the process of generating configuration commands through intent translation, the scheduling of resources must take into account both the status of the underlying network and user needs. Specifically, on the one hand, it is necessary to ensure that the issued commands can be safely executed under the current network conditions, and on the other hand, it is necessary to meet the user's intent needs as much as possible to avoid the allocated resources deviating too much from the expected value, so as to maximize the network benefits.

[0006] The existing patent CN202410586617.X (application name: A device and method for translating application intentions in an intent-based network, applicant: Guizhou University, application date: 2024.05.13) discloses a device and method for translating application intentions in an intent-based network, which involves the field of a new generation of artificial intelligence, including an intention collection module, an intention translation module, a knowledge base and a network virtual request interface; the intention collection module collects application information in the business to achieve a complete expression of business intentions in the IBN; the intention translation module converts the collected business intention information into specific network virtual requests, and uses machine learning technology to explore the deep-level correlation between business intentions and network virtual requests; the network virtual request interface is a bridge between the system and the network slice, which obtains information from the network slice and sends the generated network virtual request to the network; the knowledge base is used to store data information collected by the intention collection module and the network virtual request interface. Specifically, this technical solution collects the application intention data of users in the network and trains this data using an enhanced Transformer algorithm. Through this enhanced Transformer algorithm, the potential rules between user intentions and network virtual requests can be deeply explored. Subsequently, the user's intentions are converted into specific network virtual requests, and the slice information corresponding to the request is obtained from the network slice manager to fill the network virtual request. Finally, the generated network virtual request will be sent to the network for actual execution. On the one hand, this technical solution matches the intention to a network virtual request with a fixed template through the Transformer algorithm, but fails to truly understand the user's intention at the semantic level, so it cannot dynamically generate configuration commands that can meet the user's needs, nor can it ensure that the generated network virtual request complies with the issued network equipment; on the other hand, it only obtains information from the network slice to fill the request as a result of resource allocation, without fully considering the real-time status of the underlying network, and therefore cannot ensure that the intention requirements are met to the greatest extent.

[0007] Therefore, how to accurately translate user intent into configuration commands from a semantic level, so as to dynamically generate configuration commands that meet user needs and determine the rationality and executability of the commands has become a technical issue that technical personnel have focused on. Summary of the invention

[0008] In view of this, the purpose of the present invention is to provide an intent-driven self-intelligent network on-demand service method and system, which can accurately translate user intent into configuration commands from a semantic level, thereby dynamically generating configuration commands that meet user needs and determining the rationality and executability of the commands.

[0009] In order to achieve the above object, the present invention provides an intention-driven self-intelligent network on-demand service method, comprising:

[0010] Step 1: Collect input user intent, which is a configuration request or business requirement expressed in natural language;

[0011] Step 2: Generate a configuration generation model according to the user's intention, and then use the configuration generation model to translate the user's intention and output a configuration command template sequence without filled parameters, wherein the configuration command template sequence includes semantic level parameters and indicator level parameters;

[0012] Step 3: Obtain the user intent and the network configuration document of each semantic-level parameter in the configuration command template sequence, and then input the user intent, the configuration command template sequence, the user intent and the network configuration documents of all semantic-level parameters into the configuration generation model. The configuration generation model modifies the configuration command template sequence and fills in the values ​​of all semantic-level parameters, and finally outputs the configuration command template sequence after the semantic-level parameters are filled.

[0013] Step 4: Set a number of service quality indicators. The configuration generation model generates a service quality indicator weight matrix according to the user's intention. Then, an expert model is constructed and trained for each service quality indicator. Each expert model calculates the resource allocation weight vector of each indicator-level parameter in the configuration command template sequence under the corresponding service quality indicator. Finally, according to the service quality indicator weight matrix and the resource allocation weight vector of each expert model, the value of each indicator-level parameter is determined and filled into the configuration command template sequence.

[0014] Step 5: Output the configuration command template sequence one by one to the network device for execution.

[0015] In order to achieve the above object, the present invention also provides an intention-driven self-intelligent network on-demand service system, comprising:

[0016] An intention input device collects input user intentions, where user intentions are configuration requests or business requirements expressed in natural language;

[0017] An intention translation device generates a configuration generation model according to user intention, and the configuration generation model translates the user intention into a configuration command template sequence without filled parameters, the configuration command template sequence contains semantic level parameters and indicator level parameters, and then obtains the user intention and the network configuration document of each semantic level parameter in the configuration command template sequence, inputs the user intention, the configuration command template sequence, the user intention and the network configuration document of all semantic level parameters into the configuration generation model, and the configuration generation model modifies the configuration command template sequence and fills in values ​​for all semantic level parameters, thereby obtaining the configuration command template sequence after the semantic level parameters are filled;

[0018] The resource scheduling device sets a number of service quality indicators, and generates a service quality indicator weight matrix according to the user's intention by the configuration generation model, and then constructs and trains an expert model for each service quality indicator, and each expert model calculates the resource allocation weight vector of each indicator-level parameter in the configuration command template sequence under the corresponding service quality indicator, and finally determines the value of each indicator-level parameter according to the service quality indicator weight matrix and the resource allocation weight vector of each expert model, and fills it into the configuration command template sequence;

[0019] The network status sensing device runs inside the network controller and collects network data in real time;

[0020] The configuration execution device outputs the configuration command template sequence filled with semantic-level parameters and indicator-level parameters one by one to the network device for execution.

[0021] In order to achieve the above object, the present invention further provides a computing device, comprising:

[0022] Memory and processor;

[0023] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the intention-driven self-intelligent network on-demand service method are implemented.

[0024] In order to achieve the above-mentioned object, the present invention also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the intention-driven self-intelligent network on-demand service method.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention makes full use of the network equipment manual, performs fine-tuning of all parameters of the large model to adapt it to the data in the network equipment manual, thereby generating a configuration generation large model, and the configuration generation large model can automatically generate a configuration command template sequence according to the user's intention, accurately translate the user's intention into the configuration command from the semantic level, and ensure the rationality and executability of the command, thereby realizing the organic combination of intention conversion and resource allocation, and significantly improving the automation and intelligence level of the self-intelligent network; the present invention also combines with external knowledge bases (such as equipment manuals, user examples, etc.), makes full use of configuration information such as network equipment manuals, corrects unreasonable template sequences, and fills in appropriate values ​​for the semantic-level parameters in the configuration command template sequence, thereby reducing the problem of large model hallucinations. The invention also designs a hybrid expert model method based on a large model, which uses a pre-trained hybrid expert model, combined with a large model and real-time network status, to orchestrate the resources required for the execution of the intent from multiple service quality perspectives (such as packet loss rate and latency), and allocate network resources (such as bandwidth and CPU resources), thereby assisting in determining the appropriate values ​​of the indicator-level parameters in the configuration command template sequence, and realizing efficient orchestration of network resources. It not only makes the generated configuration commands more in line with user needs at the semantic level and improves user satisfaction, but also ensures the efficient use of the overall network resources, effectively simplifies the network management process, improves management efficiency, and fully guarantees the consistency between user intent and configuration commands in terms of functional requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The present invention is a flowchart of an intention-driven self-intelligent network on-demand service method shown in an exemplary embodiment of the present invention.

[0027] Figure 2 An exemplary embodiment of the present invention is shown Figure 1 Specific process flow chart of step 4.

[0028] Figure 3 It is a structural diagram of an intention-driven self-intelligent network on-demand service system shown in an exemplary embodiment of the present invention.

[0029] Figure 4 It is a schematic diagram of the structure of a computer device according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the present invention provides an intention-driven self-intelligent network on-demand service method, including:

[0032] Step 1: Collect input user intent, which is a configuration request or business requirement expressed in natural language;

[0033] Step 2: Generate a configuration generation model based on the user's intent, and then use the configuration generation model to translate the user's intent and output a configuration command template sequence without filled parameters, where the configuration command template sequence contains semantic-level parameters and indicator-level parameters. The semantic-level parameters are non-network resource parameters, such as user IP, service type, etc., and the indicator-level parameters are network resource parameters, such as bandwidth, CPU resources, etc.;

[0034] Step 3: Obtain the user intent and the network configuration document of each semantic-level parameter in the configuration command template sequence, and then input the user intent, the configuration command template sequence, the user intent and the network configuration documents of all semantic-level parameters into the configuration generation model. The configuration generation model modifies the configuration command template sequence and fills in the values ​​of all semantic-level parameters, and finally outputs the configuration command template sequence after the semantic-level parameters are filled.

[0035] Step 4: Set a number of service quality indicators. The configuration generation model generates a service quality indicator weight matrix according to the user's intention. Then, an expert model is constructed and trained for each service quality indicator. Each expert model calculates the resource allocation weight vector of each indicator-level parameter in the configuration command template sequence under the corresponding service quality indicator. Finally, according to the service quality indicator weight matrix and the resource allocation weight vector of each expert model, the value of each indicator-level parameter is determined and filled into the configuration command template sequence.

[0036] Step 5: Output the configuration command template sequence one by one to the network device for execution.

[0037] In step 2, generating a configuration to generate a large model may further include:

[0038] Obtain the device manual corresponding to the user intent, preprocess the device manual, parse and obtain information in the device manual, including the mapping relationship between user intent and configuration command template sequence, description of configuration commands and parameters, and the relationship between configuration commands and parameters, construct a network domain knowledge dataset in the format of question and answer, and then build a configuration generation model based on the Transformer architecture, and use the network domain knowledge dataset to fine-tune the configuration generation model.

[0039] The configuration generation model can be built based on the Transformer architecture, and its core self-attention mechanism can effectively understand and extract key information from user intent. The large model is pre-trained on a wide range of general text data to ensure that it has strong language understanding and generation capabilities. In this way, it can capture complex semantic relationships and be applied to a variety of natural language processing tasks. On this basis, the large model is fine-tuned using the constructed domain knowledge dataset, so that the large model can increase its understanding of network configuration commands and improve its ability to translate intent tasks, so as to understand user intent from a semantic level and formulate corresponding configuration command template sequences. In order to avoid the catastrophic forgetting problem caused by the fine-tuning process, the full parameter fine-tuning training method can be used to update all the weights of the large model. Compared with the traditional LoRA (Low-Rank Adaptation) fine-tuning method, this method can improve the large model's ability to understand intent diversity and greatly reduce the generation of large model hallucination problems. The fine-tuned large model is deployed online to form a configuration generation model.

[0040] In step 3, obtaining the user intent and configuring the network configuration document for each semantic level parameter in the command template sequence may further include:

[0041] Network configuration documents such as device manuals and user examples are collected, and the network configuration documents are segmented and vectorized to build an external knowledge base in the network field. Then, the FAISS similarity vector retrieval algorithm is adopted, that is, the user intention or each semantic-level parameter is initially vectorized, and then the similarity is calculated one by one with all the segmented vector documents in the external knowledge base in the network field. Finally, several vector documents with high similarity are selected as the network configuration documents for user intention or each semantic-level parameter.

[0042] In step 3, the user intent, the configuration command template sequence, the user intent and the network configuration document of all semantic-level parameters are input into the configuration generation model, the configuration generation model modifies the configuration command template sequence, and fills in the values ​​of all semantic-level parameters, and finally outputs the configuration command template sequence after the semantic-level parameters are filled, which may further include:

[0043] The user intention, configuration command template sequence, user intention and network configuration documents of all semantic-level parameters are combined into prompt words, which are input into the configuration generation model. The configuration generation model corrects the unreasonable parts in the configuration command template sequence based on the input network configuration document, such as missing parameters, wrong sequence order, etc. Then, the user intention and network configuration documents of all semantic-level parameters are combined into prompt words, which are further input into the configuration generation model. The configuration generation model takes values ​​and fills in the semantic-level parameters in the configuration command template sequence.

[0044] The present invention can also use a hybrid expert model algorithm based on a large model to complete the allocation of network resources and assist in completing the value selection of indicator-level parameters in the configuration command. Specifically, the user's needs for different service quality indicators are first analyzed through a weight analysis algorithm, and then in the expert model allocation algorithm, the expert model trained for different service quality indicators will allocate the network resources required by the user based on the current network information, and each expert model will output an independent allocation result, thereby calculating a unique resource allocation plan, and synchronously filling in the corresponding indicator-level parameters in the configuration command template sequence.

[0045] like Figure 2 As shown, Figure 1 Step 4 may further include:

[0046] Step 41: Set a number of service quality indicators, and obtain multiple resource allocation values ​​corresponding to each indicator-level parameter in the configuration command template sequence, input the advantages and disadvantages of all service quality indicators and user intentions into the configuration generation model, and calculate the weight of each service quality indicator by the configuration generation model to form a service quality indicator weight matrix: W = {w1, w2, ..., w N}, w1, w2, …, w N are the weights of the 1st, 2nd, …, Nth service quality indicators, respectively, and N is the total number of service quality indicators;

[0047] Traditional hybrid expert models usually use specially trained gating mechanisms to analyze intent inputs and assign corresponding weights to multiple expert models, thereby forming a weight matrix. In contrast, the present invention relies on the general task understanding capabilities of the large model itself, and combines the network domain knowledge injected after fine-tuning to use the configuration generation large model as a gating mechanism. In terms of specific implementation, the advantages and disadvantages of different service quality indicators and user intent are constructed into prompts and input into the configuration generation large model. The configuration generation large model uses its powerful semantic understanding capabilities to assign appropriate weights to each service quality indicator to correspond to different expert models, thereby constructing a service quality indicator weight matrix;

[0048] Step 42: construct and train an expert model for each service quality indicator. The expert model calculates the different resource allocation value weights of each indicator-level parameter in the configuration command template sequence under its corresponding service quality indicator according to the targeted service quality indicator and the current network status, and forms a resource allocation weight vector for each indicator-level parameter under its corresponding service quality indicator: is the resource allocation weight vector of the ith indicator-level parameter under the nth service quality indicator, are the 1st, 2nd, …, Mth values ​​of the i-th indicator-level parameter under the nth service quality indicator. i The weight of resource allocation value, Mi is the total number of resource allocation values ​​corresponding to the i-th indicator-level parameter;

[0049] Step 43: Calculate the resource allocation weight vector of each indicator-level parameter by matrix operation according to the service quality indicator weight matrix and the resource allocation weight vector of each indicator-level parameter under all service quality indicators: is the resource allocation weight vector of the i-th indicator-level parameter, The i-th index level parameter corresponds to the 1st, 2nd, ..., Mth i The weight of the resource allocation value, They are the resource allocation weight vectors of the i-th indicator-level parameter under the 1st, 2nd, ..., Nth service quality indicators, and then the resource allocation value corresponding to the maximum value is selected from the weights of all resource allocation values ​​corresponding to each indicator-level parameter as the value of the indicator-level parameter and filled into the configuration command template sequence.

[0050] Expert models refer to machine learning or artificial intelligence models that have high expertise and capabilities in specific fields or tasks. These models can provide highly accurate and specialized solutions in a specific field through special training processes and optimization. The present invention can train expert models based on deep reinforcement learning models, construct an optimization objective function for its service quality indicators, and use it as a reward function for training. Real-time collected network data is used as training input, such as network topology, switch flow table, switch CPU resources, bandwidth and latency, etc., and select from multiple resource allocation value options corresponding to each indicator-level parameter. After multiple rounds of reinforcement learning training, the resource allocation weight vector of each optimized indicator-level parameter under its corresponding service quality indicator is finally generated. When training the expert model, the topological network of the network resources corresponding to each indicator-level parameter can be set as: G = {D, L}, where D represents the topological network link set and the number of links is S D , L represents the set of network devices in the topology network, and the number of devices is S L , each network device l(l∈L) is allocated a certain network resource B l The topological network allocates network resources B in the device to meet the user service demand u (u∈U). u To strive to meet all user service needs U.

[0051] The state space, action space, and reward function in the Markov decision process for training the expert model can be defined as follows:

[0052] State space S u :S u Indicates user service demand u from source network device n uReach the target network device d u The resource status of the network device passed by is composed of the remaining network resources in network device 1 (that is, the idle resources that can be allocated by the current network device) and B l , Source network equipment n u and target network device d u composition, in, They are 1st, 2nd, ..., S L The remaining network resources in the network devices

[0053] Action space A: the set of network resource allocation values ​​that can be set in the current topology network, expressed as A=B k =

[0054] {b1,b2,…,b k}, where b1, b2, …, b k They are the network resource allocation values ​​that each network device can choose.

[0055] Reward function r u (S u ,A): The reward function is used to reflect the optimization goal of each expert model, that is, to maximize the corresponding service quality indicator benefits, so the reward function is expressed as Among them, β1 represents the service fee paid by the user that can be obtained by meeting the service quality index, and its value can be a preset unit fixed value. β2 is the cost coefficient, which represents the equipment loss generated when the topology network allocates network resources, and its value can be a fixed parameter.

[0056] For example, if it is set to 1, Denotes the service quality indicator t n The calculation method of B i , S i yes The two input parameters of , different service quality indicators have different calculation methods, and a suitable calculation method can be selected according to the service quality indicator, for example, the corresponding service quality indicator can be calculated according to network data such as speed transmission rate and interruption probability. The present invention does not specifically limit this method. The utility function M(·) is calculated as follows: t n,max ,t n,min They are respectively the maximum and minimum acceptable service quality indicators that the current topology network can provide to users.

[0057] The present invention can train and generate expert models based on the Deep Q-Network (DQN) algorithm. DQN is particularly good at dealing with problems in high-dimensional state spaces. Its core idea is to use deep neural networks to approximate the Q function, thereby avoiding the problem that the Q table in traditional Q learning cannot handle large state spaces. However, there are still problems such as slow model convergence speed and unstable convergence results. In order to better enable the model to explore the network state space, explore the edge cases in the space as much as possible, and improve the stability of model learning, the present invention adds a noise layer to DQN to replace the common ∈-greedy strategy in DQN, and introduces noise random variables ∈ and ∈ in the parameters of the behavior network and target network of DQN. - , and the noise parameters σ and σ - Therefore, the objective function formula of DQN is replaced by:

[0058] Among them, θ represents the model parameters of the behavior network, and θ - represents the model parameters of the target network, S u+1 is the new state of the network, a u is a possible action in the current state, r u is the reward, D represents the data distribution in the experience replay memory, a * is a possible action in the new state, γ is the discount factor, and its value can be set according to actual business needs. Q(·) is the default output value of the DQN model, which is usually a multidimensional vector. E(·) represents the expected value or mean of the variables in the brackets.

[0059] The training process of the expert model is as follows:

[0060] Step 1: Initialize the experience replay memory, the model parameters θ of the behavior network, and the parameters θ in the target network - ,For each reinforcement learning training iteration, the network resources of the devices in the topology network are reinitialized and the current state space S is obtained u , including the remaining network resources B in each network device l , source node n u and the target node d u ;

[0061] Step 2: Perform reinforcement learning training iterations, wherein the specific steps of each training iteration further include: first select and execute an action a from the action space A u , that is, the network resources allocated to the user, the topological network performs the allocation of network resources, and then calculates the reward r based on the allocation result u (S u ,A), and then obtain the next new network state S u+1, and store the conversion process (S u ,a u ,r u ,S u+1 ) into the experience playback memory. When the number of transformation processes stored in the memory space reaches the preset number, a number of transformation samples (S j ,a j ,r j ,S j+1 ) and set the target value:

[0062] y=r j +γmax a*∈A Q(S j+1 ,a * ,∈ - θ - ,σ - ), r j 、a j 、s j 、s j+1 They are the reward of the jth conversion sample extracted, an action performed in the current state, the current network state, and the next new network state. Then, the model parameter θ of the behavior network is adjusted by gradient descent to minimize the difference between the prediction and the target, and the state is changed to the new state. At this point, a round of iteration is completed. After a fixed iteration step, the target network changes the model parameter θ - Replace with θ.

[0063] To explain the method of the present invention more clearly, the following is an implementation process of a specific embodiment of the present invention. When receiving the user's intention: ensuring that user A can watch 4K video smoothly from 6 to 8 p.m., the processing process of the embodiment is as follows:

[0064] Step s1 (corresponding to Figure 1 Step 2) Configure the generated large model and output the following configuration command template sequence according to the user's intention:

[0065] "define operation:for endpoint()add middlebox()al low service()setbandwidth()set time_range()"

[0066] "define algor itm:set PucchAlgoSwitch()set DynAdjVoltSwitch()";

[0067] Step s2 (corresponding to Figure 1Step 3) Obtain the following relevant network configuration documents: 1) To optimize 4K video, you generally need to turn on the PUCCH algorithm switch, dynamic voltage regulation switch, and load balancing algorithm switch. 2) To smoothly watch 4K video on the Internet, you need to use the middleware video streaming and 4K video service. 3) The bandwidth used by 4K video generally ranges from 50-100Mbps. 4) User A's IP address is 192.168.0.1;

[0068] Step s3 (corresponding to Figure 1 Step 3) Based on the retrieved network configuration document, it is recombined with the user's natural language intention and the configuration command template sequence as prompt words and input into the configuration generation model, so as to modify the unreasonable part of the configuration command template sequence (such as the lack of load balancing algorithm switch parameters) and fill in the semantic level parameter values:

[0069] "define operation:for endpoint('192.168.0.1')add middlebox('videostreaming')al low service('4k video')set bandwidth()set time_range('18:00-20:00')"

[0070] "define algor itm:set PucchAlgoSwitch('ON')set DynAdjVoltSwitch('ON')set MlbAlgoSwitch('ON')";

[0071] Step s4 (corresponding to Figure 1 Step 4), combine the user intention and the characteristic description of the network's service quality indicators (service quality indicators include transmission rate, packet loss rate and bit error rate) as prompt words and input them into the configuration generation model to generate a service quality indicator weight matrix: W = [0.5, 0.3, 0.2] (where the weight of the transmission rate is 0.5, the weight of the packet loss rate is 0.3, and the weight of the bit error rate is 0.2);

[0072] Step s5 (corresponding to Figure 1 Step 4) Obtain the network topology and bandwidth of each device to form a state vector as the input of the three expert models. The resource allocation values ​​corresponding to the index-level parameter bandwidth include: 40Mps, 60Mps, 80Mps, 100Mps. The three expert models calculate a resource allocation weight vector under the corresponding service quality index for the index-level parameter bandwidth: 1) Transmission rate expert model: 2) Packet loss rate expert model: 3) Bit Error Rate Expert Model:

[0073] Then calculate the resource allocation weight vector of the indicator-level parameter bandwidth: The bandwidth resource allocation value of 60 Mbps corresponding to the maximum weight of 0.3967 is taken as the bandwidth allocated to the user;

[0074] Step s6 (corresponding to Figure 1 Step 4) Fill in the index-level parameter bandwidth in the configuration command template sequence:

[0075] "define operat ion:for endpoint('192.168.0.1')add midd lebox('videostreaming')al low service('4k video')set bandwidth('60','mbps')set t ime_range('18:00-20:00')"

[0076] "define a lgor ithm:set PucchAlgoSwitch('ON')set DynAdjVo ltSwitch('ON')set M l bAlgoSwitch('ON')"

[0077] Step s7 (corresponding to Figure 1 Step 5) Output the configuration command template sequence one by one to the actual network device for execution.

[0078] Step s8: The configuration command is successfully executed and the success information is fed back to the user.

[0079] like Figure 3 As shown, the present invention provides an intention-driven self-intelligent network on-demand service system, including:

[0080] An intention input device collects input user intentions, where user intentions are configuration requests or business requirements expressed in natural language;

[0081] An intention translation device generates a configuration generation model according to user intention, and the configuration generation model translates the user intention into a configuration command template sequence without filled parameters, the configuration command template sequence contains semantic level parameters and indicator level parameters, and then obtains the user intention and the network configuration document of each semantic level parameter in the configuration command template sequence, inputs the user intention, the configuration command template sequence, the user intention and the network configuration document of all semantic level parameters into the configuration generation model, and the configuration generation model modifies the configuration command template sequence and fills in values ​​for all semantic level parameters, thereby obtaining the configuration command template sequence after the semantic level parameters are filled;

[0082] The resource scheduling device sets a number of service quality indicators, and generates a service quality indicator weight matrix according to the user's intention by the configuration generation model, and then constructs and trains an expert model for each service quality indicator, and each expert model calculates the resource allocation weight vector of each indicator-level parameter in the configuration command template sequence under the corresponding service quality indicator, and finally determines the value of each indicator-level parameter according to the service quality indicator weight matrix and the resource allocation weight vector of each expert model, and fills it into the configuration command template sequence;

[0083] The network status sensing device runs inside the network controller and collects network data in real time;

[0084] The configuration execution device outputs the configuration command template sequence filled with semantic-level parameters and indicator-level parameters one by one to the network device for execution.

[0085] See also Figure 4 , Figure 4 4 is a block diagram of a computing device 400 shown in an exemplary embodiment of this specification. The components of the computing device 400 include but are not limited to a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and the database 450 is used to store data.

[0086] The computing device 400 also includes an access device 440 that enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of network interface (e.g., a network interface card (NIC)) of wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and the like.

[0087] In one embodiment of the present specification, the above components of the computing device 400 and Figure 4 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 4 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0088] The computing device 400 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a PC. The computing device 400 may also be a mobile or stationary server or a cloud server, etc.

[0089] The processor 420 is used to execute the following computer executable instructions, which, when executed by the processor, implement the steps of the above-mentioned intention-driven self-intelligent network on-demand service method.

[0090] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned intention-driven self-intelligent network on-demand service method belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be referred to the description of the technical scheme of the above-mentioned intention-driven self-intelligent network on-demand service method.

[0091] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned intention-driven self-intelligent network on-demand service method.

[0092] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium belongs to the same concept as the above-mentioned intention-driven self-intelligent network on-demand service method. The details not described in detail in the technical scheme of the storage medium can be referred to the description of the technical scheme of the above-mentioned intention-driven self-intelligent network on-demand service method or system.

[0093] An embodiment of the present specification further provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above-mentioned intention-driven autonomous network on-demand service method.

[0094] The above is a schematic scheme of a computer program of this embodiment. It should be noted that the technical scheme of the computer program and the technical scheme of the above-mentioned intention-driven self-intelligent network on-demand service method belong to the same concept, and the details not described in detail in the technical scheme of the computer program can be referred to the description of the technical scheme of the above-mentioned intention-driven self-intelligent network on-demand service method or system.

[0095] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0096] The computer instructions include computer program codes, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0097] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0098] 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 in the scope of protection of the present invention.

Claims

1. An intent-driven self-intelligent network on-demand service method, characterized in that: Included are: Step 1: Collect input user intent, which is a configuration request or business requirement expressed in natural language; Step 2: Generate a configuration generation model according to the user's intention, and then use the configuration generation model to translate the user's intention and output a configuration command template sequence without filled parameters, wherein the configuration command template sequence includes semantic level parameters and indicator level parameters; Step 3: Obtain the user intent and the network configuration document of each semantic-level parameter in the configuration command template sequence, and then input the user intent, the configuration command template sequence, the user intent and the network configuration documents of all semantic-level parameters into the configuration generation model. The configuration generation model modifies the configuration command template sequence and fills in the values ​​of all semantic-level parameters, and finally outputs the configuration command template sequence after the semantic-level parameters are filled. Step 4: Set a number of service quality indicators. The configuration generation model generates a service quality indicator weight matrix according to the user's intention. Then, an expert model is constructed and trained for each service quality indicator. Each expert model calculates the resource allocation weight vector of each indicator-level parameter in the configuration command template sequence under the corresponding service quality indicator. Finally, according to the service quality indicator weight matrix and the resource allocation weight vector of each expert model, the value of each indicator-level parameter is determined and filled into the configuration command template sequence. Step 5: Output the configuration command template sequence one by one to the network device for execution.

2. The method according to claim 1, characterized in that In step 2, generating a configuration to generate a large model further includes: Obtain the device manual corresponding to the user intent, preprocess the device manual, parse the device manual to obtain the mapping relationship between the user intent and the configuration command template sequence, the description of the configuration command and parameters, and the relationship information between the configuration command and parameters, construct a network domain knowledge dataset in the format of question and answer, and then build a large configuration generation model based on the Transformer architecture, and use the network domain knowledge dataset to fine-tune the large configuration generation model.

3. The method according to claim 1, characterized in that In step 3, obtaining the network configuration document of each semantic-level parameter in the user intent and configuration command template sequence further includes: Equipment manuals and user example network configuration documents are collected, and the network configuration documents are segmented and vectorized to build an external knowledge base in the network field. Then, the FAISS similarity vector retrieval algorithm is adopted, that is, the user intention or each semantic-level parameter is initially vectorized, and then the similarity is calculated one by one with all the segmented vector documents in the external knowledge base in the network field. Finally, several vector documents with high similarity are selected as the network configuration documents of the user intention or each semantic-level parameter.

4. The method according to claim 1, characterized in that: In step 3, the user intent, the configuration command template sequence, the user intent and the network configuration document of all semantic-level parameters are input into the configuration generation model, the configuration generation model modifies the configuration command template sequence, and fills in the values ​​of all semantic-level parameters, and finally outputs the configuration command template sequence after the semantic-level parameters are filled, further including: The user intention, the configuration command template sequence, the user intention and the network configuration documents of all semantic-level parameters are combined into prompt words, and the configuration is input into a large model. The configuration generation large model modifies the configuration command template sequence based on the input network configuration document. Then the user intention and the network configuration documents of all semantic-level parameters are combined into prompt words, and the configuration generation large model is input into the large model. The configuration generation large model takes values ​​and fills in the semantic-level parameters in the configuration command template sequence.

5. The method according to claim 1, characterized in that Step 4 further includes: Step 41: Set a number of service quality indicators, and obtain multiple resource allocation values ​​corresponding to each indicator-level parameter in the configuration command template sequence, input the advantages and disadvantages of all service quality indicators and user intentions into the configuration generation model, and calculate the weight of each service quality indicator by the configuration generation model to form a service quality indicator weight matrix: W={w1,w2,...,w N }, w1, w2, …, w N are the weights of the 1st, 2nd, …, Nth service quality indicators, respectively, and N is the total number of service quality indicators; Step 42: construct and train an expert model for each service quality indicator. The expert model calculates the different resource allocation value weights of each indicator-level parameter in the configuration command template sequence under its corresponding service quality indicator according to the targeted service quality indicator and the current network status, and forms a resource allocation weight vector for each indicator-level parameter under its corresponding service quality indicator: is the resource allocation weight vector of the ith indicator-level parameter under the nth service quality indicator, are the 1st, 2nd, …, Mth values ​​of the i-th indicator-level parameter under the nth service quality indicator. i The weight of resource allocation value, M i is the total number of resource allocation values ​​corresponding to the i-th indicator-level parameter; Step 43: Calculate the resource allocation weight vector of each indicator-level parameter by matrix operation according to the service quality indicator weight matrix and the resource allocation weight vector of each indicator-level parameter under all service quality indicators: is the resource allocation weight vector of the i-th indicator-level parameter, The i-th index level parameter corresponds to the 1st, 2nd, ..., Mth i The weight of the resource allocation value, They are the resource allocation weight vectors of the i-th indicator-level parameter under the 1st, 2nd, ..., Nth service quality indicators, and then the resource allocation value corresponding to the maximum value is selected from the weights of all resource allocation values ​​corresponding to each indicator-level parameter as the value of the indicator-level parameter and filled into the configuration command template sequence.

6. The method according to claim 1, characterized in that The expert model is trained based on the deep reinforcement learning model. When training the expert model, the topological network of the network resources corresponding to each indicator-level parameter is set as: G = {D, L}, where D represents the topological network link set and the number of links is S D , L represents the set of network devices in the topology network, and the number of devices is S L , each network device l is allocated network resources B l , l∈L, the topology network allocates network resources B in the device to serve the user service demand u u To meet all user service needs U, u∈U.

7. The method according to claim 6, characterized in that The state space, action space, and reward function in the Markov decision process for training the expert model are defined as follows: State space S u :S u Indicates user service demand u from source network device n u Reach the target network device d u The resource status of the network devices passed through is composed of the remaining network resources B in network device l l , Source network equipment n u and target network device d u composition, Among them, B1, B2, ..., They are 1st, 2nd, ..., S L The remaining network resources in the network devices Action space A: the set of network resource allocation values ​​that can be set in the current topology network, expressed as A=B k = {b1,b2,…,b k }, where b1, b2, …, b k They are the network resource allocation values ​​that each network device can choose. Reward function r u (S u ,A): The reward function is used to reflect the optimization goal of each expert model, that is, to maximize the corresponding service quality indicator benefits. The reward function is expressed as Among them, β1 represents the service fee paid by users that can be obtained by meeting the service quality index, β2 is the cost coefficient, which represents the equipment loss generated when the topology network allocates network resources, Represents the service quality indicator t n The calculation method of the utility function M(·) is as follows: t n,max ,t n,min They are respectively the maximum and minimum acceptable service quality indicators that the current topology network can provide to users.

8. The method according to claim 7, characterized in that The expert model is trained based on the DQN algorithm. A noise layer is added to DQN, and noise random variables ∈ and ∈ are introduced into the parameters of the DQN behavior network and target network. - , and the noise parameters σ and σ - To enhance, therefore, the objective function formula of DQN is set to: Among them, θ represents the model parameters of the behavior network, and θ - represents the model parameters of the target network, S u+1 is the new state of the network, a u is a possible action in the current state, r u is the reward, D represents the data distribution in the experience replay memory, a * is a possible action in the new state, γ is the discount factor, Q(·) is the default output value of the DQN model, and E(·) represents the expected value or mean of the variables in brackets. The training process of the expert model is as follows: Step 1: Initialize the experience replay memory, the model parameters θ of the behavior network, and the parameters θ in the target network - ,For each reinforcement learning training iteration, the network resources of the devices in the topology network are reinitialized and the current state space S is obtained u , including the remaining network resources B in each network device l , source node n u and target node d u ; Step 2: Perform reinforcement learning training iterations, wherein the specific steps of each training iteration further include: first select and execute an action a from the action space A u , that is, the network resources allocated to the user, the topological network performs the allocation of network resources, and then calculates the reward r based on the allocation result u (S u ,A), and then obtain the next new network state S u+1 , and store the conversion process (S u ,a u ,r u ,S u+1 ) into the experience playback memory. When the number of transformation processes stored in the memory space reaches the preset number, a number of transformation samples (S j ,a j ,r j ,S j+1 ) and set the target value: y=r j +γmax a*∈A Q(S j+1 ,a * ,∈ - θ - ,σ - ), r j 、a j 、s j 、s j+1 They are the reward of the jth conversion sample extracted, an action performed in the current state, the current network state, and the next new network state. Then, the model parameter θ of the behavior network is adjusted by gradient descent to minimize the difference between the prediction and the target, and the state is changed to the new state. At this point, a round of iteration is completed. After a fixed iteration step, the target network changes the model parameter θ - Replace with θ.

9. An intention-driven self-intelligent network on-demand service system, characterized in that: Included are: An intention input device collects input user intentions, where user intentions are configuration requests or business requirements expressed in natural language; An intention translation device generates a configuration generation model according to user intention, and the configuration generation model translates the user intention into a configuration command template sequence without filled parameters, the configuration command template sequence contains semantic level parameters and indicator level parameters, and then obtains the user intention and the network configuration document of each semantic level parameter in the configuration command template sequence, inputs the user intention, the configuration command template sequence, the user intention and the network configuration document of all semantic level parameters into the configuration generation model, and the configuration generation model modifies the configuration command template sequence and fills in values ​​for all semantic level parameters, thereby obtaining the configuration command template sequence after the semantic level parameters are filled; The resource scheduling device sets a number of service quality indicators, and generates a service quality indicator weight matrix according to the user's intention by the configuration generation model, and then constructs and trains an expert model for each service quality indicator, and each expert model calculates the resource allocation weight vector of each indicator-level parameter in the configuration command template sequence under the corresponding service quality indicator, and finally determines the value of each indicator-level parameter according to the service quality indicator weight matrix and the resource allocation weight vector of each expert model, and fills it into the configuration command template sequence; The network status sensing device runs inside the network controller and collects network data in real time; The configuration execution device outputs the configuration command template sequence filled with semantic-level parameters and indicator-level parameters one by one to the network device for execution.

10. A computing device, characterized in that: include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the intention-driven self-intelligent network on-demand service method described in any one of claims 1-8 are implemented.

11. A computer-readable storage medium, characterized in that: It stores computer executable instructions, which, when executed by a processor, implement the steps of the intention-driven self-intelligent network on-demand service method described in any one of claims 1 to 8.

Images