Optimization method and optimization device for large model and electronic equipment
By constructing a hierarchical heuristic knowledge graph and adjusting feature weights, the problem of insufficient processing capabilities of sparse key information in the big model is solved, and the accuracy and robustness of business processing are improved.
Patent Information
- Application Number
- CN202411272546.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-12-17
AI Technical Summary
In complex large models, sparse key information is difficult to be correctly identified and processed, which affects the accuracy and robustness of business processing and leads to reduced processing capabilities.
By building a hierarchical heuristic knowledge graph, identify and define sparse key information, search target features in the target big model, and adjust feature weights to enhance the processing power of the big model.
It improves the business processing capabilities of the large model when facing sparse key information, and improves the processing capabilities and accuracy of target business.
Smart Images
Figure CN119150969A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of smart home / intelligent family technology, and more specifically, to an optimization method and optimization device and electronic device for a large model. Background Art
[0002] At present, smart home systems have been widely used in various aspects such as home security, environmental control, and health management. With the increase in the number of home appliances in smart home systems and the complexity and diversification of business scenarios, the big models used by smart home systems will become more complex.
[0003] In a complex large model, there is often some key information that is not easily perceived but has a significant impact on the performance of the large model (these key information are referred to as sparse key information in this application). If the large model cannot correctly identify, understand and process these sparse key information, it will affect the accuracy and robustness of the large model in processing the business related to the sparse key information, resulting in a reduction in the ability of the large model to process this business.
[0004] Therefore, how to improve the business processing capabilities of large models when faced with sparse key information has become a technical problem that needs to be solved urgently.
[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present application, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention
[0006] In order to provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. The summary is not an extensive review, nor is it intended to identify key / critical components or delineate the scope of protection of these embodiments, but rather serves as a prelude to the detailed description that follows.
[0007] The disclosed embodiments provide an optimization method, an optimization device, and an electronic device for a large model, which can improve the business processing capability of a large model when facing sparse key information.
[0008] In some embodiments, an optimization method for a large model includes: obtaining business data of a target business and constructing a hierarchical heuristic knowledge graph based on the business data; wherein the target business is a business that requires enhanced processing capabilities in the target large model to be optimized; receiving setting instructions input by a user to define sparse key information in the hierarchical heuristic knowledge graph; based on the hierarchical heuristic knowledge graph, searching for target features corresponding to the sparse key information in the target large model; and adjusting the weights of the reference target features in the target large model to enhance the target large model's processing capabilities for the target business.
[0009] Optionally, a hierarchical heuristic knowledge graph is constructed based on the business data, including: identifying and marking multiple entities in the business data, and the relationships between the multiple entities; obtaining heuristic information corresponding to the target item business, and determining business rules based on the heuristic information; based on the relationships between the multiple entities and the business rules, the business data is constructed into a knowledge graph with a hierarchical structure to obtain a hierarchical heuristic knowledge graph.
[0010] Optionally, based on the hierarchical heuristic knowledge graph, target features corresponding to the sparse key information in the target large model are searched, including: constructing a heuristic search strategy according to the hierarchical heuristic knowledge graph and the architectural parameters of the target large model; using the heuristic search strategy to search for the sparse key information in the target large model to determine the target features corresponding to the sparse key information.
[0011] Optionally, a heuristic search strategy is constructed based on the hierarchical heuristic knowledge graph and the architectural parameters of the target large model, including: determining the search boundary based on the architectural parameters of the target large model; determining the heuristic function and evaluation indicators based on the hierarchical heuristic knowledge graph; and constructing a heuristic search strategy based on the search boundary, heuristic function and evaluation indicators.
[0012] Optionally, adjusting the weight of the reference target feature in the target macromodel includes: determining an objective function of the reference target feature in the target macromodel; and adjusting the influence weight of the target feature on the objective function according to a pre-set adjustment strategy.
[0013] Optionally, after adjusting the weight of the reference target feature in the target macro model according to the influence weight of the target feature on the target function according to a preset adjustment strategy, the optimization method further includes: removing business processing function features other than the target feature function in the target macro model.
[0014] Optionally, after adjusting the weights of reference target features in the target big model, the optimization method also includes: receiving a business processing request related to the target item business; running the adjusted target big model to process the business processing request to obtain a business processing result; optimizing the hierarchical heuristic knowledge graph according to the business processing result, and / or optimizing the heuristic search strategy for searching sparse key information in the target big model.
[0015] In some embodiments, an optimization device for a large model includes: a construction module, configured to obtain business data of a target item business and construct a hierarchical heuristic knowledge graph based on the business data; wherein the target item business is a business that requires enhanced processing capabilities in a target large model to be optimized; a definition module, configured to receive setting instructions input by a user and define sparse key information in the hierarchical heuristic knowledge graph; a search module, configured to search for target features corresponding to the sparse key information in the target large model based on the hierarchical heuristic knowledge graph; and an adjustment module, configured to adjust the weights of reference target features in the target large model to enhance the processing capabilities of the target large model for the target item business.
[0016] In some embodiments, an optimization device for a large model includes a processor and a memory storing program instructions, and the processor is configured to execute the optimization method for a large model as described above when running the program instructions.
[0017] In some embodiments, an electronic device includes: a device body; and an optimization device for a large model as described above, installed on the device body.
[0018] The optimization method, optimization device, and electronic device for a large model provided by the embodiments of the present disclosure can achieve the following technical effects:
[0019] In the disclosed embodiment, a hierarchical structure and heuristic information knowledge graph, i.e., a hierarchical heuristic knowledge graph, can be constructed based on the business data of the target item business that needs to enhance the processing capability in the target big model to be optimized. After the construction is completed, the sparse key information in the hierarchical knowledge graph can be defined according to the setting instructions input by the user. After the definition is completed, the target features corresponding to the sparse key information in the target big model can be searched based on the hierarchical heuristic knowledge graph, and the weights of the reference target features in the target big model can be adjusted. In this way, the target big model can understand and discover the sparse key information at a deeper level when processing the target item business. Therefore, the disclosed embodiment can improve the business processing capability of the big model when facing sparse key information, and can improve the processing capability of the target item business.
[0020] The above general description and the following description are exemplary and explanatory only and are not intended to limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0022] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0023] Figure 1 It is a schematic diagram of the hardware environment of the optimization method for a large model provided by an embodiment of the present disclosure;
[0024] Figure 2 is a schematic diagram of an electronic device provided by an embodiment of the present disclosure;
[0025] Figure 3 is a schematic diagram of an optimization method for a large model provided by an embodiment of the present disclosure;
[0026] Figure 4 is a schematic diagram of another optimization method for a large model provided by an embodiment of the present disclosure;
[0027] Figure 5 is a schematic diagram of an optimization device for a large model provided by an embodiment of the present disclosure;
[0028] Figure 6 It is a schematic diagram of another optimization device for large models provided by an embodiment of the present disclosure. DETAILED DESCRIPTION
[0029] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.
[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0031] According to one aspect of the embodiments of the present disclosure, a method for optimizing a large model is provided. The optimization method can be widely used to optimize a large model arranged in a server of a whole-house intelligent digital control system such as a smart home system, a smart home system, an intelligent home device ecosystem, and an intelligence house ecosystem.
[0032] Optionally, the whole-house intelligent digital control system includes multiple household appliances, including: smart air conditioners, smart range hoods, smart refrigerators, smart ovens, smart stoves, smart washing machines, smart water heaters, smart washing equipment, smart dishwashers, smart projection equipment, smart TVs, smart clothes drying racks, smart curtains, smart audio and video, smart sockets, smart speakers, smart speakers, smart fresh air equipment, smart kitchen and bathroom equipment, smart bathroom equipment, smart sweeping robots, smart window cleaning robots, smart mopping robots, smart air purification equipment, smart steamers, smart microwave ovens, smart kitchen treasures, smart purifiers, smart water dispensers and smart door locks, but not limited to these.
[0033] In some embodiments, the optimization method for large models can be applied to Figure 1 The hardware environment is composed of the electronic device 100 and the server 200 of the whole house intelligent digital control system (for example, the server of the smart home system). Figure 1 As shown, the electronic device 100 is connected to the server 200 via a network, so that the electronic device 100 can optimize the large model deployed in the server 200.
[0034] The above network may include but is not limited to at least one of the following: wired network, wireless network. The above wired network may include but is not limited to at least one of the following: wide area network, metropolitan area network, local area network, and the above wireless network may include but is not limited to at least one of the following: WIFI (Wireless Fidelity), Bluetooth.
[0035] In some embodiments, the electronic device 100 may Figure 2 As shown, it includes a device body 110 and an optimization apparatus 500 (600) for a large model.
[0036] Specifically, the optimization device 500 (600) for the large model is installed on the equipment body.
[0037] Optionally, the optimization device 600 for the large model includes a processor, which can obtain the business data of the target item business that needs to enhance the processing capability in the target large model to be optimized, and can construct a hierarchical heuristic knowledge graph based on the business data. The processor can define the sparse key information in the hierarchical heuristic knowledge graph based on the setting instructions received from the user. After the definition is completed, the processor can search for the target features corresponding to the sparse key information in the target large model based on the hierarchical heuristic knowledge graph, and adjust the weight of the reference target features in the target large model.
[0038] In combination with the above electronic device, the embodiment of the present disclosure provides an optimization method for a large model, such as Figure 3 As shown, the optimization method includes:
[0039] S301, the processor obtains the business data of the target business and constructs a hierarchical heuristic knowledge graph based on the business data.
[0040] The target item business is the business that needs enhanced processing capability in the target large model to be optimized.
[0041] Specifically, the target large model represents the large model to be optimized in the smart home system.
[0042] It can be understood that the large model deployed in the server of the smart home system may be able to handle multiple services, such as question-answering interaction services, fault prediction services, and control optimization services. It may also be used to process only one service, such as question-answering interaction services. The above-mentioned target services represent services with weak service processing capabilities that require large models to enhance processing capabilities.
[0043] Specifically, the acquired service data includes historical data of processing the target service in the smart home system, and data related to the target service in the Internet.
[0044] Specifically, a hierarchical heuristic knowledge graph refers to a knowledge graph with heuristic information that is constructed using a hierarchical structure. In a hierarchical heuristic knowledge graph, entities and relationships are organized into different levels to reflect the hierarchy, classification, and mutual relationship between different entities. Heuristic information is used to guide the construction process of the hierarchical knowledge graph, including identifying the relationship between entities and optimizing the graph structure.
[0045] Specifically, based on the acquired business data, multiple entities and the relationships between multiple entities can be determined. Therefore, the processor can construct a hierarchical heuristic knowledge graph based on the business data.
[0046] Optionally, before constructing a hierarchical heuristic knowledge graph based on the business data, the optimization method also includes: preprocessing the business data.
[0047] Specifically, by preprocessing the business data, noise data, duplicate data and irrelevant data in the business data can be removed, and the business data can be converted into a standardized data format. This is conducive to improving the accuracy of the hierarchical heuristic knowledge graph constructed subsequently.
[0048] Specifically, the methods of preprocessing business data include but are not limited to: data cleaning, disassembly of complex structure data, etc.
[0049] S302, the processor receives a setting instruction input by a user and defines sparse key information in a hierarchical heuristic knowledge graph.
[0050] Specifically, sparse key information is key information that is not easily perceived in a large model but has an important impact on the performance of the large model.
[0051] Specifically, users can analyze business data or use simulation tools to analyze business data to determine which key information in the business data is not easily perceived but has an important impact on the performance of the large model. After determining the sparse key information, users can enter the corresponding setting instructions to define and mark the corresponding information in the hierarchical heuristic knowledge graph.
[0052] Optionally, the business data is analyzed to determine sparse key information, which can be analyzed in combination with the sensitivity of the hierarchical heuristic knowledge graph.
[0053] Specifically, combined with the sensitivity of the hierarchical heuristic knowledge graph, we can use the hierarchical heuristic knowledge graph as a basis to simulate the failure or abnormal performance of each node using simulation tools to observe the impact on other nodes or the whole. Nodes that have a large radiation impact or have a large impact on certain key nodes should be identified as key information even if the number of state changes is not large.
[0054] Specifically, the steps for analyzing the sensitivity of the hierarchical heuristic knowledge graph are as follows:
[0055] First, a scoring system is created to score each node based on factors such as its position in the hierarchical heuristic knowledge graph, the number of edges associated with it, and the attributes of the edges.
[0056] Second, for each node, simulate their failure or abnormality based on known data or professional knowledge. For example, when an air conditioner fails, simulate a series of related events that may cause room temperature rise, power consumption change, etc.
[0057] Third, by simulating node failures or abnormalities and combining them with the above-mentioned scoring system, we observe how changes in the current node affect other nodes in the hierarchical heuristic knowledge graph.
[0058] Fourth, through the above simulation and observation, an "impact score" is defined for each possible failure or abnormal situation, where the larger the range of the node radiation, the greater the weight score needs to be given.
[0059] Fifth, the event information of nodes that do not occur frequently but have high scores when simulating failures or abnormalities is identified as sparse key information.
[0060] For example, for air conditioners in the field of smart home appliances, under normal weather conditions, changes in the working state of the air conditioner do not necessarily have a significant impact on the indoor temperature, and failures are relatively rare. However, if there is a weather condition of heat waves and extreme heat that poses a huge threat to human comfort and health, the failure of the air conditioner will have a serious impact. In this case, the failure of the air conditioner in heat waves and extreme heat weather conditions can be used as sparse key information.
[0061] Optionally, the business data is analyzed to determine sparse key information, which may be analyzed using a PageRank algorithm.
[0062] Specifically, the steps for analysis using the PageRank algorithm are as follows:
[0063] First, the PageRank algorithm assigns an initial ranking value to each node in the hierarchical heuristic knowledge graph.
[0064] Second, the PageRank algorithm formula is used to update the ranking score of each node. The basic idea of the PageRank algorithm is that a node with a high ranking will pass a higher ranking value to other nodes connected to it. The formula is as follows:
[0065] PR(A)=(1-d)+d(PR(T1) / C(T1)+...+PR(Tn) / C(Tn));
[0066] Among them, PR(A) is the PageRank value of node A to be calculated, PR(Ti) is the PageRank value of all nodes Ti connected to node A; C(Ti) is the number of connections contained in node Ti, and d is the damping coefficient, the default value is generally 0.85.
[0067] Third, perform reciprocating iterative calculations. When the scores of all current nodes in a certain iteration are less than the preset threshold compared with the previous iteration or the preset maximum number of iterations is reached, the algorithm ends.
[0068] Fourth, the event information of nodes whose ranking scores exceed a threshold is confirmed as sparse key information.
[0069] S303, the processor searches for target features corresponding to the sparse key information in the target large model based on the hierarchical heuristic knowledge graph.
[0070] Specifically, the target large model is a deep learning model obtained through large-scale data training, which has strong semantic understanding and reasoning capabilities. Based on the hierarchical heuristic knowledge graph that defines sparse key information, the target large model can use its semantic understanding ability to perform logical reasoning and knowledge reasoning on the sparse key information. Therefore, the target features corresponding to these sparse key information can be searched in the internal representation space of the target large model.
[0071] Specifically, since the hierarchical structure can gradually narrow the search scope and avoid aimless searches in large-scale data, searching for target features corresponding to sparse key information in the target large model based on the hierarchical heuristic knowledge graph helps improve search efficiency.
[0072] Specifically, since the target large model is trained through deep learning technology and has strong semantic understanding capabilities, it can enhance the semantic understanding capabilities of the target large model by combining the structured knowledge representation of the hierarchical heuristic knowledge graph, enabling it to more accurately identify and understand sparse key information, which helps to improve the accuracy of the search.
[0073] S304, the processor adjusts the weight of the reference target feature in the target large model to enhance the processing capability of the target large model for the target item business.
[0074] Specifically, in the target large model, different features have different importance for specific businesses or tasks. Some features may be highly relevant to the target business and have a significant impact on the prediction or decision results of the target large model, while other features may be relatively minor. Therefore, by adjusting the weights of the reference features, the target large model can be made to focus more on these key information corresponding to the features when making decisions. Therefore, the processor can enhance the processing capability of the target large model for the target business by adjusting the weights of the reference target features in the target large model.
[0075] Optionally, adjusting the weight of the reference target feature in the target macromodel includes: determining an objective function of the reference target feature in the target macromodel; and adjusting the influence weight of the target feature on the objective function according to a pre-set adjustment strategy.
[0076] Specifically, the target function is a function that performs a specific function with reference to the target feature, for example, a function that determines whether each household appliance in the smart home system has a fault based on the collected operating parameters of the smart home system.
[0077] Specifically, by adjusting the weight of the impact of the target feature on the target function according to the preset adjustment strategy, the target function can be made more targeted when performing a specific function. This is conducive to improving the ability of the target large model to process the target business.
[0078] For example, the fault diagnosis model in the smart home system is used as a target large model. When performing fault diagnosis, the various sensors collected are data collected from multiple angles. However, the working status or performance index data of some key equipment (such as power supply, main control board) has a greater impact on the operation of the smart home system, but the probability of problems is not high. This information can be identified as sparse key information. In this case, the loss function in the fault diagnosis model can be designed to adjust the weights of the operating parameters of the power supply and the main control board, so that the fault diagnosis model pays more attention to the operating parameters of the power supply and the main control board, and gives a strong representation learning promotion. In this way, the fault diagnosis ability of the fault diagnosis model can be enhanced.
[0079] Optionally, the influence weight of the target feature on the objective function is adjusted according to a pre-set adjustment strategy, including: assigning a higher weight to the target feature in the objective function; or, formulating an optimization strategy for the target feature and making the target large model learn the optimization strategy.
[0080] Optionally, after adjusting the influence weight of the target feature on the target function according to a preset adjustment strategy, the optimization method further includes: removing business processing functions other than the target function in the target large model.
[0081] Specifically, after adjusting the weight of the impact of the target feature on the target function according to the preset adjustment strategy, the business processing functions other than the target function in the target large model are removed, so that the target large model can become a dedicated model only for processing the target feature. In this way, through the optimization method provided by the embodiment of the present disclosure, a large model specifically used for processing sparse key information can be constructed. When this large model is applied to the smart home system, the business processing capability of the smart home system when facing the target business can be enhanced.
[0082] In the disclosed embodiment, a hierarchical structure and heuristic information knowledge graph, i.e., a hierarchical heuristic knowledge graph, can be constructed based on the business data of the target item business that needs to enhance the processing capability in the target big model to be optimized. After the construction is completed, the sparse key information in the hierarchical knowledge graph can be defined according to the setting instructions input by the user. After the definition is completed, the target features corresponding to the sparse key information in the target big model can be searched based on the hierarchical heuristic knowledge graph, and the weights of the reference target features in the target big model can be adjusted. In this way, the target big model can understand and discover the sparse key information at a deeper level when processing the target item business. Therefore, the disclosed embodiment can improve the business processing capability of the big model when facing sparse key information, and can improve the processing capability of the target item business.
[0083] In some embodiments, a hierarchical heuristic knowledge graph is constructed based on business data, including: identifying business data to mark multiple entities in the business data, and the relationships between the multiple entities; obtaining heuristic information corresponding to the target item business, and determining business rules based on the heuristic information; based on the relationships between the multiple entities and the business rules, the business data is constructed into a knowledge graph with a hierarchical structure to obtain a hierarchical heuristic knowledge graph.
[0084] Specifically, natural language processing (NLP) technology can be used to extract entities with specific meanings from business data, such as device names, event names, and user names. Entities can also be extracted from business data in combination with specific rules and dictionaries in the business field.
[0085] For example, if the business data is "temperature has an impact on air conditioning efficacy", the identified entity is "air conditioning". If the business data is "number of people has an impact on room lighting requirements", the identified entities are "people" and "lighting".
[0086] Specifically, the mutual relationship between multiple entities can be determined by analyzing the interaction and connection between the entities. Specifically, the mutual relationship includes but is not limited to "belong to", "include" and "associated with".
[0087] For example, taking the business data of "the number of people has an impact on the lighting demand of the room" as an example, it can be determined that the entity "personnel" and the entity "lighting" have an associated relationship.
[0088] Specifically, the heuristic information corresponding to the target business includes expert knowledge and empirical rules in the business field. Based on these data, the specific relationship between entities, the logic of data processing and the basis for decision making can be further clarified. Therefore, the business rules of the target business can be determined based on the heuristic information.
[0089] Specifically, based on the business rules of the target business and the relationships between multiple entities, the importance of each entity in multiple dimensions can be clarified, and the hierarchical structure of the knowledge graph can be constructed based on the importance. The entities are taken as nodes of the knowledge graph, and the relationships between entities are taken as edges connecting the nodes and added to the hierarchical structure to obtain a hierarchical heuristic knowledge graph.
[0090] Specifically, the above importance can be determined based on the system level, technical complexity and other parameters to which the entity belongs in each dimension. For example, if the importance of the air conditioner in both the environmental comfort dimension and the energy consumption dimension is relatively high, while the importance of the lighting lamp is relatively high only in the environmental brightness, the air conditioner can be assigned a higher level than the lighting lamp.
[0091] It should be noted that the structure and content of the constructed hierarchical heuristic knowledge graph can be adjusted and optimized when business needs change or new business data is entered.
[0092] In some embodiments, based on a hierarchical heuristic knowledge graph, target features corresponding to sparse key information in a target large model are searched, including: constructing a heuristic search strategy according to the hierarchical heuristic knowledge graph and the architectural parameters of the target large model; using the heuristic search strategy to search for sparse key information in the target large model to determine the target features corresponding to the sparse key information.
[0093] Specifically, according to the hierarchical structure of the hierarchical heuristic knowledge graph and the architectural parameters of the target large model, a series of heuristic rules can be determined to guide the search process. For example, according to the hierarchical structure of the hierarchical heuristic knowledge graph, the level where the entities or relationships directly connected to the sparse key information are located can be determined, and these levels can be searched first; according to the architectural parameters of the target large model, parameters such as the search direction and search boundary can be determined. Therefore, according to the hierarchical heuristic knowledge graph and the architectural parameters of the target large model, a heuristic search strategy can be constructed.
[0094] Specifically, after constructing a heuristic search strategy, the hidden layer of the target large model is traversed according to the heuristic search strategy, the feature embedding matrix is queried, and the attention mechanism is applied to find features corresponding to the sparse key information. During the search process, it is necessary to continuously evaluate the degree of match between the found features and the sparse key information, and the features with the highest matching degree are ranked as the target features. Among them, similarity metrics, classification accuracy, regression error and other indicators can be used to evaluate the degree of match between the found features and the sparse key information.
[0095] Optionally, a heuristic search strategy is constructed based on the hierarchical heuristic knowledge graph and the architectural parameters of the target large model, including: determining the search boundary based on the architectural parameters of the target large model; determining the heuristic function and evaluation indicators based on the hierarchical heuristic knowledge graph; and constructing a heuristic search strategy based on the search boundary, heuristic function and evaluation indicators.
[0096] Specifically, it is understandable that searching the entire space inside the large model will waste a lot of time and have low search efficiency. Therefore, in the embodiment of the present disclosure, it is necessary to determine the search boundary of the target large model according to the architecture parameters of the target large model.
[0097] Specifically, what kind of key information is sparse key information can be determined based on the hierarchical heuristic knowledge graph. Therefore, the evaluation index can be determined based on the hierarchical heuristic knowledge graph.
[0098] Specifically, the gradient of each search step or the probability prediction of achieving the goal can be determined according to the hierarchical heuristic knowledge graph, and the direction of the search can be determined. Therefore, according to the hierarchical heuristic knowledge graph, a heuristic function for guiding the search process can be determined.
[0099] Specifically, after determining the search boundary and evaluation index, the heuristic function can use simulated annealing, genetic algorithm, ant colony optimization, Bayesian optimization and other algorithms to search the space divided by the search boundary in the target large model. Therefore, according to the search boundary, heuristic function and evaluation index, a heuristic search strategy can be constructed.
[0100] In the disclosed embodiment, when searching for target features corresponding to sparse key information in the target large model, a heuristic search strategy is first constructed based on the hierarchical heuristic knowledge graph and the architecture parameters of the target large model, and then the target large model is searched according to the heuristic search strategy. In this way, the efficiency and accuracy of searching for target features are improved.
[0101] The present disclosure provides another optimization method for large models, such as Figure 4 As shown, the optimization method includes:
[0102] S401, the processor obtains the business data of the target business and constructs a hierarchical heuristic knowledge graph based on the business data.
[0103] The target item business is the business that needs enhanced processing capability in the target large model to be optimized.
[0104] S402, the processor receives a setting instruction input by a user and defines sparse key information in a hierarchical heuristic knowledge graph.
[0105] S403, the processor searches for target features corresponding to the sparse key information in the target large model based on the hierarchical heuristic knowledge graph.
[0106] S404, the processor adjusts the weight of the reference target feature in the target large model to enhance the processing capability of the target large model for the target item business.
[0107] S405: The processor receives a business processing request related to the target business.
[0108] Specifically, the business processing request is a request from a user for testing the target business processing capability of the adjusted target large model.
[0109] S406: The processor runs the adjusted target large model to process the business processing request and obtains the business processing result.
[0110] S407, the processor optimizes the hierarchical heuristic knowledge graph according to the business processing results, and / or optimizes the heuristic search strategy for sparse key information in the search target large model.
[0111] Specifically, based on the business processing results, it can be determined whether the business processing capabilities of the adjusted target large model when facing sparse key information have been enhanced, whether the enhanced effect can meet the usage requirements, and the shortcomings. Based on the shortcomings, the unreasonable structure in the hierarchical heuristic knowledge graph can be analyzed, or the unreasonable design in the heuristic search strategy for searching the target large model can be analyzed. Therefore, the processor can optimize the hierarchical heuristic knowledge graph and the heuristic search strategy based on the business processing results.
[0112] Specifically, after optimizing the hierarchical heuristic knowledge graph and heuristic search strategy, the optimized hierarchical heuristic knowledge graph and heuristic search strategy will be used to re-optimize the target large model.
[0113] In the disclosed embodiment, after adjusting the weight of the reference target feature in the target large model, the adjusted target large model is tested, and the target large model is further optimized according to the test result. In this way, the business processing capability of the target large model when facing sparse key information can be further improved.
[0114] Combination Figure 5As shown, the embodiment of the present disclosure provides an optimization device 500 for a large model, including: a construction module 501, a definition module 502, a search module 503 and an adjustment module 504. The construction module 501 is configured to obtain the business data of the target item business, and to construct a hierarchical heuristic knowledge graph based on the business data; wherein the target item business is a business that needs to enhance the processing capability in the target large model to be optimized. The definition module 502 is configured to receive the setting instructions input by the user, and define the sparse key information in the hierarchical heuristic knowledge graph. The search module 503 is configured to search for the target features corresponding to the sparse key information in the target large model based on the hierarchical heuristic knowledge graph. The adjustment module 504 is configured to adjust the weight of the reference target features in the target large model to enhance the processing capability of the target large model for the target item business.
[0115] Combination Figure 6 As shown, the embodiment of the present disclosure provides an optimization device 600 for a large model, including: a processor (processor) 601 and a memory (memory) 602. Optionally, the optimization device 600 for a large model may also include a communication interface (Communication Interface) 603 and a bus 604. Among them, the processor 601, the communication interface 603, and the memory 602 can communicate with each other through the bus 604. The communication interface 603 can be used for information transmission. The processor 601 can call the logic instructions in the memory 602 to execute the optimization method for the large model in the above embodiment.
[0116] In addition, the logic instructions in the memory 602 described above may be implemented in the form of software functional units and when sold or used as independent products, may be stored in a computer-readable storage medium.
[0117] The memory 602 is a computer-readable storage medium that can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor 601 executes the function application and data processing by running the program instructions / modules stored in the memory 602, that is, the optimization method for the large model in the above embodiment is implemented.
[0118] The memory 602 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory 602 may include a high-speed random access memory and may also include a non-volatile memory.
[0119] An embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above-mentioned optimization method for a large model.
[0120] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.
Claims
1. An optimization method for large models, It is characterized in that include: Obtain the business data of the target business and build a hierarchical heuristic knowledge graph based on the business data; The target item business is the business that needs enhanced processing capability in the target large model to be optimized; Receive the setting instruction input by the user, define the sparse key information in the hierarchical heuristic knowledge graph; search for the target features corresponding to the sparse key information in the target large model based on the hierarchical heuristic knowledge graph; Adjust the weights of the reference target features in the target big model to enhance the target big model's processing capabilities for the target item business.
2. The optimization method for a large model according to claim 1, It is characterized in that Construct a hierarchical heuristic knowledge graph based on business data, including: Identify multiple entities in business data and the relationships between multiple entities; Obtaining heuristic information corresponding to the target item business, and determining business rules based on the heuristic information; According to the mutual relationships and business rules among multiple entities, the business data is constructed into a knowledge graph with a hierarchical structure to obtain a hierarchical heuristic knowledge graph.
3. The optimization method for a large model according to claim 1, It is characterized in that Based on the hierarchical heuristic knowledge graph, we search for target features corresponding to sparse key information in the target large model, including: Construct a heuristic search strategy based on the hierarchical heuristic knowledge graph and the architecture parameters of the target large model; A heuristic search strategy is adopted to search for sparse key information in the target large model to determine the target features corresponding to the sparse key information.
4. The optimization method for a large model according to claim 3, It is characterized in that According to the hierarchical heuristic knowledge graph and the architecture parameters of the target large model, a heuristic search strategy is constructed, including: Determine the search boundary based on the architecture parameters of the target large model; Determine the heuristic function and evaluation index based on the hierarchical heuristic knowledge graph; Construct a heuristic search strategy based on search boundaries, heuristic functions and evaluation indicators.
5. The optimization method for a large model according to any one of claims 1 to 4, It is characterized in that Adjust the weights of reference target features in the target macro model, including: Determine the objective function of the reference target feature in the target macro model; Adjust the impact weight of the target feature on the objective function according to the pre-set adjustment strategy.
6. The optimization method for a large model according to claim 5, It is characterized in that After adjusting the influence weight of the target feature on the target function according to a preset adjustment strategy, the optimization method further includes: removing the business processing functions other than the target function in the target large model.
7. The optimization method for a large model according to any one of claims 1 to 4, It is characterized in that After adjusting the weights of the reference target features in the target macro model, the optimization method further includes: Receive business processing requests related to target business items; Run the adjusted target large model to process the business processing request and obtain the business processing result; Optimize the hierarchical heuristic knowledge graph according to the business processing results, and / or optimize the heuristic search strategy for sparse key information in the search target large model.
8. An optimization device for large models, It is characterized in that include: A construction module is configured to obtain business data of a target business and construct a hierarchical heuristic knowledge graph according to the business data; wherein the target business is a business that needs to enhance processing capability in the target large model to be optimized; A definition module is configured to receive a setting instruction input by a user and define sparse key information in a hierarchical heuristic knowledge graph; A search module is configured to search for target features corresponding to sparse key information in a target large model based on a hierarchical heuristic knowledge graph; The adjustment module is configured to adjust the weight of the reference target feature in the target large model to enhance the processing capability of the target large model for the target item business.
9. An optimization device for a large model, comprising a processor and a memory storing program instructions, It is characterized in that The processor is configured to execute the optimization method for a large model according to any one of claims 1 to 7 when running the program instructions.
10. An electronic device, It is characterized in that include: Equipment body; The optimization device for large models as described in claim 8 or 9 is installed on the equipment body.