Self-adaptive question-answering system and method based on knowledge distillation and multi-modal dynamic fusion

Abstract

The invention discloses an adaptive question-answering system based on knowledge distillation and multi-modal dynamic fusion, and the system comprises a knowledge distillation module which is used for migrating knowledge of a teacher model pre-trained on corpora in the communication field to a lightweight student model, achieving model compression through optimizing a distillation loss function, and obtaining a multi-modal dynamic fusion model; the loss function comprises a soft label output by the teacher model and a KL divergence constraint output by the student model; the multi-modal knowledge fusion module comprises a feature extraction unit, a self-adaptive weighting unit and an attention fusion unit; the self-adaptive inference engine comprises a semantic analysis unit; according to the cross-modal reasoning method and system, semantic alignment of equipment parameters, protocol texts and topological graphs is achieved through the multi-modal dynamic fusion technology, and the cross-modal reasoning accuracy is improved; compared with an original model, the lightweight student model has the advantage that the reasoning speed is increased in a protocol analysis task.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing and artificial intelligence, and specifically relates to an adaptive question-answering system and method based on knowledge distillation and multi-modal dynamic fusion. Background Art

[0002] The present invention belongs to the technical field of natural language processing and artificial intelligence, and specifically relates to a communication knowledge system construction and adaptive question-answering method based on knowledge distillation and multi-modal dynamic fusion. Aiming at the core problems such as the dispersion of professional knowledge in the communication field (isolated multi-source data such as device parameters, protocol rules, and historical work orders), semantic fragmentation of multi-modal data (text, images, structured tables), and broken reasoning logic chains in complex scenarios, through the collaborative optimization of lightweight model compression driven by knowledge distillation, multi-modal dynamic weight fusion, and adaptive inference path generation technologies, a lightweight knowledge system for communication device operation and maintenance is constructed, and high-precision and low-latency domain-adaptive question answering is achieved. By migrating the knowledge of large models to lightweight inference engines and combining cross-modal semantic alignment and dynamic conflict resolution mechanisms, this method breaks through the limitations of traditional question-answering systems in scenarios such as communication protocol matching, device fault diagnosis, and risk assessment, such as lagging knowledge updates, inefficient utilization of multi-source data, and insufficient complex logical reasoning capabilities. It is particularly suitable for industrial scenarios with high professionalism and strong real-time requirements such as 5G base station operation and maintenance, core network configuration optimization, and communication protocol version iteration. Summary of the Invention

[0003] To achieve the above object, the technical solution of the present invention is as follows: To solve the problems of dispersed knowledge in the communication field, shallow semantic understanding, and insufficient dynamic scenario adaptation ability in the prior art, the present invention proposes a communication knowledge system construction and adaptive question-answering method based on knowledge distillation and multi-modal dynamic fusion, aiming to achieve unified management, deep reasoning, and dynamic update of multi-source heterogeneous knowledge, and improve the question-answering accuracy and adaptability in complex scenarios.

[0004] The present invention provides an adaptive question-answering system and method based on knowledge distillation and multi-modal dynamic fusion. The system includes: A knowledge distillation module for migrating the knowledge of a teacher model pre-trained on communication field corpora to a lightweight student model, and realizing model compression by optimizing the distillation loss function. The loss function includes the KL divergence constraint between the soft labels output by the teacher model and the output of the student model; A multi-modal knowledge fusion module, including a feature extraction unit that uses a Transformer model to extract semantic vectors of text, device structure diagrams, and parameter tables respectively; an adaptive weighting unit that dynamically adjusts the fusion weights of the text, image, and table modalities based on the question type; and an attention fusion unit that realizes cross-modal semantic interaction through a multi-head attention mechanism; An adaptive inference engine performs the following operations: a semantic parsing unit that uses a Bi-LSTM-CRF model to identify communication entities and intentions in a question; a dynamic path generation unit that matches knowledge base entries according to the question semantics and generates an inference link with weights; a conflict resolution unit that uses a student model to perform logical verification on sub-task answers and outputs them sorted based on an evidence priority algorithm.

[0005] Preferably, the knowledge distillation module includes: In the pre-training stage of the teacher model, training corpus is constructed using communication device manuals, protocol RFC documents, and operation and maintenance work order data; A progressive distillation strategy that first transfers the parameters of the semantic understanding layer and then transfers the parameters of the task inference layer; The distillation loss function is defined as: , where and is dynamically adjusted according to the communication task type. L_task represents the loss function of the student model on the original target task, and L_KL represents the loss function of teacher knowledge transfer. By dynamically adjusting, the accuracy of the student model's answer to the task goal is ensured, and it forms a two-objective optimization with L_KL to prevent it from simply mimicking the teacher model.

[0006] Preferably, in the multi-modal knowledge fusion module: The adaptive weighting unit adopts a gating mechanism, and the gating value g is calculated from the cosine similarity between the question type encoding vector and the modal feature; The attention fusion unit performs hierarchical fusion: first, local attention within the modality is performed, and then global cross-modal attention is performed.

[0007] Preferably, the dynamic path generation unit includes: A path selector based on reinforcement learning, and the reward function is defined as R = γ * answer accuracy + (1 - γ) * inference latency, where γ is an adjustable weight; A path weight update mechanism that automatically reduces the path weight of obsolete protocol rules when a device protocol version change is detected.

[0008] Preferably, the conflict resolution unit performs: Contradiction detection based on the ontology of the communication field, and the ontology defines the mutually exclusive relationships of device parameters, protocol versions, and physical layer specifications; Multi-evidence source credibility evaluation, defining the weight of the device manual w1 > the weight of the protocol document w2 > the weight of the historical work order w3; If the conflict persists, trigger an artificial review mark and feedback it to the knowledge base update module.

[0009] Preferably, the adaptive question-answering method based on knowledge distillation and multi-modal dynamic fusion implemented by the system includes the following steps: The knowledge system construction stage includes: Multi-modal data preprocessing and feature extraction, cleaning the original corpus in the communication field and performing structured parsing, extracting key entities such as device parameters, protocol rules, and fault codes; using deep learning models to generate semantic vectors for text, images, and tables respectively; Knowledge distillation-driven lightweight model training, transferring the knowledge of the pre-trained GPT-3.5 teacher model in the communication field to the TinyBERT-6L student model through the KL divergence loss function, and adopting a dynamic temperature coefficient adjustment strategy to optimize protocol parsing and fault diagnosis capabilities; Multi-modal dynamic fusion and knowledge base construction, designing a gating weight allocation module to dynamically adjust the fusion weights of text, images, and tables, generating a unified multi-dimensional communication knowledge graph through the multi-head attention mechanism, and storing it in a graph database; Incremental knowledge update and version compatibility, when a new device protocol or version upgrade occurs, injecting new knowledge using the few-shot distillation technique, combined with the historical rollback mechanism and protocol dependency graph; The adaptive question answering and reasoning stage includes: Semantic parsing and intent recognition, identifying communication entities in the user's question through the Bi-LSTM-CRF model, and using the lightweight student model to parse the question intent; Dynamic inference link generation, constructing an inference path generator based on reinforcement learning, designing a reward function weighted by the accuracy-delay ratio, dynamically selecting the optimal knowledge base entries, and decomposing subtasks; Conflict resolution and answer generation, using the ontology library in the communication field to verify logical consistency, resolving conflicts according to the priority of device manual > protocol RFC > historical work order, and outputting a structured answer containing the inference path and confidence score.

[0010] Preferably, the knowledge distillation training includes: Transferring the semantic understanding layer parameters of the teacher model; distilling the parameters of the task inference layer, retaining the protocol parsing and fault diagnosis logic; dynamically adjusting the temperature coefficient T to balance the soft label distribution and hard label supervision.

[0011] Preferably, the calculation method of the gating weight allocation module is: Inputting the question semantic vector and each modal feature vector; generating modal weights through a fully connected layer and the Sigmoid function; fusing multi-modal features weighted by the weights, and outputting a unified knowledge representation.

[0012] Preferably, the reinforcement learning reward function is designed as: Reward value R = accuracy weight × prediction accuracy + delay weight × (1 - delay ratio).

[0013] The method includes a knowledge system construction stage and an adaptive question - answering reasoning stage, which are synergistically optimized through technologies such as dynamic association of knowledge graphs, multi - dimensional similarity matching, and generation of adaptive reasoning links. The method specifically includes the following steps: Knowledge system construction stage: Step 1: Multi - modal data pre - processing and feature extraction.

[0014] Perform text cleaning and structured parsing on the original corpus in the communication field (equipment manuals, protocol documents, operation and maintenance logs), extract key entities such as equipment parameters, protocol rules, and fault codes; use deep - learning models (BERT / VisionTransformer) to generate semantic vectors of text, images, and tables respectively.

[0015] Step 2: Training of lightweight models driven by knowledge distillation.

[0016] Transfer the knowledge of the pre - trained GPT - 3.5 teacher model in the communication field to the TinyBERT - 6L student model through the KL - divergence loss function, and adopt a dynamic temperature coefficient adjustment strategy to optimize protocol parsing and fault diagnosis capabilities.

[0017] Step 3: Multi - modal dynamic fusion and knowledge - base construction.

[0018] Design a gating weight assignment module to dynamically adjust the fusion weights of text, images, and tables, generate a unified multi - dimensional communication knowledge graph through the multi - head attention mechanism, and store it in a graph database.

[0019] Step 4: Incremental knowledge update and version compatibility.

[0020] When a new device protocol or version upgrade occurs, use the few - shot distillation technology to inject new knowledge, and combine the historical rollback mechanism and protocol dependency graph to avoid catastrophic forgetting.

[0021] Adaptive question - answering reasoning stage: Step A: Semantic parsing and intention recognition. Identify communication entities (equipment models, protocol versions) in the user's question through the Bi - LSTM - CRF model, and use the lightweight student model to parse the question intention (diagnosis, configuration, risk assessment).

[0022] Step B: Generation of dynamic reasoning links. Build a reasoning path generator based on reinforcement learning, design a reward function weighted by accuracy and latency ratio, and dynamically select the optimal knowledge - base entries and decompose subtasks.

[0023] Step C: Conflict resolution and answer generation. Use the ontology library in the communication field to verify logical consistency, resolve conflicts according to the priority of equipment manuals > protocol RFCs > historical work orders, and output a structured answer containing the reasoning path and confidence score.

[0024] Preferably, the knowledge distillation training in the second step is specifically as follows: Stage 1: Transfer the semantic understanding layer parameters of the teacher model to optimize the text feature extraction ability; Stage 2: Distill the parameters of the task inference layer to retain the protocol parsing and fault diagnosis logic; Stage 3: Dynamically adjust the temperature coefficient T to balance the soft label distribution and hard label supervision.

[0025] Preferably, the calculation method of the gating weight allocation module in the third step is as follows: Input the problem semantic vector and each modality feature vector; Generate modality weights through a fully connected layer and the Sigmoid function; Weight and fuse multi-modal features according to the weights to output a unified knowledge representation.

[0026] Preferably, the reinforcement learning reward function in step B is designed as: Reward value R = accuracy weight × prediction accuracy + latency weight × (1 - latency ratio); Dynamically adjust the weight coefficient to prioritize the accuracy of high-precision scenarios.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: The multi-modal dynamic fusion technology in the present invention realizes the semantic alignment of device parameters, protocol texts, and topology diagrams, improving the cross-modal inference accuracy; The lightweight student model improves the inference speed and reduces the GPU video memory occupancy in the protocol parsing task compared with the original model; The inference path generation technology driven by reinforcement learning shortens the answer generation time and reduces the false alarm rate in the 5G base station fault diagnosis; In the core network configuration optimization task, it improves the accuracy of the solution recommendation and reduces resource conflicts; It improves the success rate of multi-vendor device compatibility verification and supports extended scenarios such as 5G network slice management. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the overall system flow of the present invention; Figure 2 is the core process of multi-modal data processing of the present invention; Figure 3 is the knowledge distillation training model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] The present invention will be further clarified below in conjunction with the drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0030] Embodiment: The knowledge system construction process of this system includes four core modules: multi-modal data processing, lightweight model training, dynamic fusion strategy, and update mechanism. The modules cooperate and optimize through feature vector interaction to ensure the efficiency and scalability of the system.

[0031] Multi-modal Data Preprocessing and Feature Extraction. The system uses heterogeneous data processing technology to structurally process multi-source information. The specific process is as follows: Text Feature Extraction: The system uses a pre-trained language model to parse unstructured text, perform entity relationship extraction tasks, generate triple knowledge units, and store them in a graph database to form a semantic network. This process lays the foundation for subsequent knowledge reasoning and task execution.

[0032] Image Feature Encoding: Extract the spatial association features of the device topology map through a vision Transformer model to generate retrievable visual semantic vectors. These visual semantic vectors can provide the system with rich image understanding capabilities and improve the accuracy of question answering.

[0033] Structured Data Modeling: Apply a graph neural network to analyze the dependency relationships between parameters, construct a parameter association matrix, and store it in an in-memory database. This can effectively capture complex structured data features and support the subsequent reasoning process.

[0034] Knowledge Distillation Model Training. To achieve efficient knowledge transfer, the system adopts a two-stage distillation strategy for model training: Semantic Alignment Stage: In this stage, the parameters of the teacher model are kept fixed, and the semantic encoding outputs of the student model are aligned through the mean squared error loss function. This process helps the student model gradually learn and absorb the knowledge of the teacher model, ensuring the consistency of semantic understanding.

[0035] Task Transfer Stage: Introduce a temperature adjustment mechanism to optimize the KL divergence loss function and dynamically control the intensity of knowledge transfer. During training, an exponential decay strategy is used to adjust the temperature coefficient to ensure that the model gradually adapts to the characteristics of the target domain, thereby improving the generalization ability of the model.

[0036] Multi-modal Dynamic Fusion Mechanism. The system designs a gated weight allocation module to achieve the fusion of cross-modal features. The specific mechanism is as follows: Joint Encoding Based on Question Semantic Vectors and Modal Features: Calculate the weight coefficients of each modality through a learnable parameter matrix to ensure the effective fusion of different modality information.

[0037] Knowledge Graph Dynamic Update Interface: Associate the fused multi-modal features with the corresponding entity nodes to form a multi-dimensional knowledge representation. This update mechanism can continuously optimize the knowledge base of the system to ensure that it always reflects the latest information.

[0038] Incremental Update and Fault Tolerance Mechanism. The system supports small-sample incremental learning and version rollback functions to ensure the reliability and flexibility of the system through the following methods: Knowledge injection: The newly added knowledge is classified into preset knowledge categories through hierarchical clustering algorithm, ensuring efficient knowledge injection with low resource consumption.

[0039] Version rollback: A deployment accuracy monitoring module is deployed. When it detects that the performance of historical tasks drops by more than a threshold, it automatically triggers a version rollback operation. This mechanism provides fault tolerance for the system and effectively reduces performance fluctuations.

[0040] Adaptive question - answering reasoning method. The system ensures accurate question - answering through three links: semantic parsing, dynamic reasoning, and conflict resolution. The specific method is as follows: Semantic understanding module. The semantic understanding module uses a sequence labeling model to identify key entities in the question and combines a lightweight language model for fine - grained intent classification. This module analyzes the implicit demand parameters in the question and generates a structured query statement containing domain constraints to ensure the accuracy of subsequent reasoning.

[0041] Dynamic reasoning engine. The system designs a reinforcement learning path selector and constructs a multi - objective reward function based on accuracy and latency metrics. The dynamic reasoning engine supports cross - modal knowledge retrieval and can parallelly execute tasks such as device parameter query, protocol rule matching, and historical case retrieval to ensure efficient multi - task reasoning ability.

[0042] Conflict resolution mechanism. The system establishes a multi - source knowledge verification rule library that can detect logical contradictions between different information sources. According to the preset priority, the system will execute a resolution strategy and output a structured answer with confidence evaluation to ensure that the answers provided by the system are still highly reliable in the face of conflicts.

[0043] System verification and deployment. To ensure the reliability and practicality of the system, the system undergoes three - stage verification: Performance testing: Compare the improvement effects of the system with traditional methods in core metrics such as accuracy and response latency to ensure a significant performance improvement of the system.

[0044] Update verification: Evaluate the impact of incremental learning on historical tasks and the adaptation efficiency of new knowledge to ensure that the system can flexibly handle the introduction of new knowledge.

[0045] Industrial deployment: Verify the large - scale application ability of the system in the operator network environment and monitor the improvement amplitude of fault location efficiency in the actual scenario to provide a basis for subsequent extensive deployment.

[0046] Test case 1 of the present invention: Knowledge fusion and reasoning ability test Objective: Verify the knowledge fusion and reasoning ability of the system.

[0047] Preparation: The Huawei D15 laptop has a running memory of 16GB and enables multi-modal data processing functions.

[0048] Prepare a knowledge base containing device parameters, 3 protocol texts, and 5 topology diagrams, such as device temperature parameters, 5G basic protocols, and base station connection topology diagrams.

[0049] List 5 comprehensive questions: Question 1: "According to the topology diagram, when the optical module temperature exceeds 45°C, what is the transmission power adjustment strategy stipulated by the protocol?" Associated knowledge sources: Device parameter table (temperature threshold), protocol document (power control clause), topology diagram (optical module location) Question 2: "The RRU unit of Base Station A is abnormally connected in the topology diagram. Explain the possible trigger protection mechanism in combination with Section 5.2 of the protocol" Associated knowledge sources: Topology diagram (RRU connection status), protocol document (fault protection clause) Question 3: "If the temperature of a certain device reaches 48°C and lasts for 3 minutes, list the operation steps that must be executed according to the protocol and the device parameter table" Question 4: "At the position marked 'Node X' in the topology diagram, what is the maximum signal delay allowed by the protocol?" Question 5: "When the physical distance between the baseband unit and the optical module exceeds 20 meters (the topology diagram shows an actual distance of 22 meters), how is the power compensation value in the device parameter table calculated?" Steps: Input the question and record the start time of processing.

[0050] Observe whether the system can integrate multi-modal information for reasoning and check if the answer combines the knowledge sources.

[0051] Test results: The cross-modal reasoning accuracy rate is 89%. For example, when dealing with device failures, it can accurately combine knowledge to provide solutions.

[0052] The lightweight student model parses the protocol 3.2 times faster. The original GPT-3.5 teacher model takes 4 seconds, now it takes 1.25 seconds; the GPU video memory is reduced from 6GB to 2GB, a reduction of 68%.

[0053] Expected results: The cross-modal reasoning accuracy rate exceeds 85%.

[0054] The protocol parsing speed is increased by 3 times and the GPU video memory is reduced by 65%.

[0055] It should be noted that the above content only illustrates the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements all fall within the protection scope of the claims of the present invention.

Claims

1. An adaptive question answering system based on knowledge distillation and multi-modal dynamic fusion, characterized in that Including: A knowledge distillation module, which is used to transfer the knowledge of a teacher model pre-trained on communication domain corpora to a lightweight student model, and realizes model compression by optimizing the distillation loss function. The loss function includes the KL divergence constraint between the soft labels output by the teacher model and the output of the student model; A multi-modal knowledge fusion module, including a feature extraction unit that uses a Transformer model to extract semantic vectors of text, device structure diagrams, and parameter tables respectively; an adaptive weighting unit that dynamically adjusts the fusion weights of text, image, and table modalities based on the question type; An attention fusion unit that realizes cross-modal semantic interaction through a multi-head attention mechanism; An adaptive inference engine that performs the following operations: a semantic parsing unit that uses a Bi-LSTM-CRF model to identify communication entities and intentions in the question; a dynamic path generation unit that matches knowledge base entries according to the question semantics and generates a weighted inference link; a conflict resolution unit that uses the student model to perform logical verification on the sub-task answers and sorts and outputs based on the evidence priority algorithm.

2. The adaptive question-answering system based on knowledge distillation and multi-modal dynamic fusion according to claim 1, wherein The knowledge distillation module includes: The pre-training stage of the teacher model, using communication device manuals, protocol RFC documents, and operation and maintenance work order data to construct training corpora; A progressive distillation strategy that first transfers the parameters of the semantic understanding layer and then transfers the parameters of the task inference layer; The distillation loss function is defined as: , where and are dynamically adjusted according to the communication task type. L_task represents the loss function of the student model on the original target task, and L_KL represents the loss function of teacher knowledge transfer. By dynamically adjusting, the accuracy of the student model's response to the task objective is ensured, and it forms a bi-objective optimization with L_KL to prevent it from simply mimicking the teacher model.

3. The adaptive question-answering system based on knowledge distillation and multi-modal dynamic fusion according to claim 1, wherein In the multi-modal knowledge fusion module: The adaptive weighting unit adopts a gating mechanism, and the gating value g is calculated by the cosine similarity between the question type encoding vector and the modal features; The attention fusion unit performs hierarchical fusion: first, perform intra-modal local attention, and then perform cross-modal global attention.

4. The adaptive question-answering system based on knowledge distillation and multi-modal dynamic fusion according to claim 1, wherein The dynamic path generation unit includes: A path selector based on reinforcement learning, and the reward function is defined as R = γ * answer accuracy + (1 - γ) * inference latency, where γ is an adjustable weight; A path weight update mechanism that automatically reduces the path weights of obsolete protocol rules when a device protocol version change is detected.

5. The adaptive question-answering system based on knowledge distillation and multi-modal dynamic fusion according to claim 1, wherein The conflict resolution unit performs: Contradiction detection based on the ontology of the communication domain, and the ontology defines the mutually exclusive relationships of device parameters, protocol versions, and physical layer specifications; Multi-evidence source credibility evaluation, defining that the weight of the device manual w1 > the weight of the protocol document w2 > the weight of the historical work order w3; If the conflict persists, trigger an artificial review mark and feedback it to the knowledge base update module.

6. An adaptive question answering method based on knowledge distillation and multi-modal dynamic fusion implemented by the system according to any one of claims 1 to 5, characterized in that, Including the following steps: The knowledge system construction stage includes: Multi-modal data preprocessing and feature extraction, performing text cleaning and structured parsing on the original communication domain corpora, and extracting key entities such as device parameters, protocol rules, and fault codes; using deep learning models to generate semantic vectors of text, images, and tables respectively; Knowledge distillation-driven lightweight model training, transferring the knowledge of the GPT-3.5 teacher model pre-trained in the communication domain to the TinyBERT-6L student model through the KL divergence loss function, and adopting a dynamic temperature coefficient adjustment strategy to optimize the protocol parsing and fault diagnosis capabilities; Multi-modal dynamic fusion and knowledge base construction, designing a gating weight allocation module to dynamically adjust the fusion weights of text, images, and tables, generating a unified multi-dimensional communication knowledge graph through a multi-head attention mechanism, and storing it in a graph database. Incremental knowledge update and version compatibility. When a new device protocol is added or the version is upgraded, small-sample distillation technology is used to inject new knowledge, combined with the historical rollback mechanism and the protocol dependency graph; The adaptive question-and-answer reasoning stage includes: Semantic parsing and intent recognition. The communication entities in the user's question are identified through the Bi-LSTM-CRF model, and the lightweight student model is used to parse the question intent; Dynamic inference link generation. Based on reinforcement learning, an inference path generator is constructed. The reward function is designed with a weighted accuracy and latency ratio to dynamically select the optimal knowledge base entries and decompose subtasks; Conflict resolution and answer generation. The ontology library in the communication field is used to verify the logical consistency, and the conflicts are resolved according to the priority of device manual > protocol RFC > historical work order, and a structured answer containing the inference path and confidence score is output.

7. The adaptive question answering method based on knowledge distillation and multi-modal dynamic fusion according to claim 6, characterized in that, Knowledge distillation training includes: Transfer the semantic understanding layer parameters of the teacher model; distill the inference layer parameters of the task, and retain the protocol parsing and fault diagnosis logic; dynamically adjust the temperature coefficient T to balance the soft label distribution and hard label supervision.

8. The adaptive question answering method based on knowledge distillation and multi-modal dynamic fusion according to claim 6, wherein The calculation method of the gated weight allocation module is: Input the question semantic vector and each modal feature vector; generate modal weights through the fully connected layer and the Sigmoid function; fuse multi-modal features by weighting according to the weights, and output a unified knowledge representation.

9. The adaptive question answering method based on knowledge distillation and multi-modal dynamic fusion according to claim 6, wherein The reinforcement learning reward function is designed as: Reward value R = accuracy weight × prediction accuracy + latency weight × (1 - latency ratio).

Images