A job scheduling optimization method based on LLM inference architecture

Abstract

The application belongs to the technical field of scheduling optimization system, and discloses a job scheduling optimization method based on LLM inference architecture. A generalization scheduling cognitive architecture search system based on LLM inference is constructed by using a labeled scheduling problem instance, and an optimal graph neural network architecture is outputted. Then, an unlabeled scheduling problem instance is inputted into the optimal graph neural network architecture, an optimal scheduling scheme is outputted, and a job is executed according to the optimal scheduling scheme. The labeled scheduling problem instance and the unlabeled scheduling problem instance are independently selected from one of JSSP, FJSP, DyFJSP and HFSP. The generalization scheduling cognitive architecture search system based on LLM inference comprises a cognitive blueprint construction module, an intelligent candidate generation module, a cognitive deliberation module and an architecture evaluation module. The job scheduling optimization method based on LLM inference architecture can break the limitations of traditional scheduling algorithms and meet the adaptive scheduling requirements of modern industry with high diversification and dynamicization.

Description

Technical Field

[0001] This invention belongs to the field of scheduling optimization system technology, and relates to a job scheduling optimization method based on LLM inference architecture. Background Technology

[0002] Scheduling optimization is a core requirement in manufacturing, supply chain collaboration, and other fields. It aims to optimize production efficiency, on-time delivery, and overall benefits through the rational allocation and prioritization of limited resources. In recent years, the development of artificial intelligence technology, especially the integration of deep learning and reinforcement learning, has provided new solutions to this classic problem.

[0003] For example, patent application CN118036979A discloses a scheduling optimization method and system based on the job shop scheduling problem in the supply chain. The core solution is to transform the supply chain scheduling problem into the job shop scheduling problem (JSSP), model it using disjunctive graphs, extract global features using pre-trained graph neural networks (GNNs), train the decision network using the DecayP3O reinforcement learning algorithm, and generate a scheduling scheme by iteratively updating the disjunctive arc direction.

[0004] For example, patent application CN116882669A discloses a profile production scheduling system based on graph neural networks and reinforcement learning. The core solution is to define JSSP scheduling as a sequential decision problem, use disjunctive graphs to represent the workshop state, learn node embedding features through GNN, and use PPO reinforcement learning algorithm to train the state representation module and decision strategy end-to-end to adapt to constraints such as mold changing and color changing in profile production.

[0005] Furthermore, in seeking better neural network model structures to improve scheduling performance, existing technologies typically use traditional Neural Architecture Search (NAS), evolutionary algorithms, or reinforcement learning to perform a "black box" search within a predefined search space, seeking better-performing network structures through extensive trial and error.

[0006] Despite the achievements of the aforementioned existing technologies, they still have the following significant drawbacks when facing complex and ever-changing industrial scenarios:

[0007] (1) Limitations of structural feature representation: Existing technologies (such as patent application CN118036979A) mostly use disjunctive graphs or ordinary heterogeneous graphs, while ordinary graphs are difficult to express complex high-order relationships in scheduling.

[0008] (2) Lack of cross-task versatility: The network structure of patent application CN116882669A is manually designed for specific problems (such as JSSP or specific profile production). Once the problem structure changes (for example, from JSSP to DyFJSP with AGV transportation introduced - dynamic flexible work shop scheduling problem, or to HFSP - hybrid flow shop scheduling problem), the original fixed network architecture often fails and cannot capture new constraint relationships.

[0009] (3) Black-box search is inefficient and lacks interpretability: Traditional NAS methods treat architecture search as a black-box optimization, blindly trying and failing, and cannot utilize prior knowledge.

[0010] In summary, existing technologies generally suffer from core problems such as weak generalization across scheduling variants, reliance on manual architecture design, low search efficiency, and lack of structured reasoning capabilities, which fail to meet the adaptive scheduling needs of diverse industrial scenarios.

[0011] Therefore, it is of great significance to study a job scheduling optimization method based on LLM inference architecture to solve the problems existing in the current technology. Summary of the Invention

[0012] The purpose of this invention is to address the problems existing in the prior art and provide a job scheduling optimization method based on LLM inference architecture.

[0013] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0014] A job scheduling optimization method based on LLM inference architecture is proposed. The method constructs a generalized scheduling cognitive architecture search system based on LLM inference using labeled scheduling problem instances and outputs the optimal graph neural network architecture. Then, unlabeled scheduling problem instances are input into the optimal graph neural network architecture to obtain the optimal scheduling scheme. The job is then executed according to the optimal scheduling scheme.

[0015] Labeled scheduling problem instances are scheduling problem instances where the scheduling scheme and its corresponding completion time are known. They are used to train the model to learn the mapping relationship from problem features to the optimal solution.

[0016] Unlabeled scheduling problem instances are scheduling problem instances where the scheduling scheme is unknown and the completion time is to be solved, requiring the use of a trained model to generate the scheduling scheme in real time;

[0017] The instances of labeled scheduling problems and the instances of unlabeled scheduling problems are each independently selected from one of JSSP, FJSP, DyFJSP, and HFSP;

[0018] JSSP: Each job consists of a series of steps in a fixed order, and each step can only be processed on a uniquely designated machine;

[0019] FJSP: Each process can be processed by selecting one of multiple machines;

[0020] DyFJSP: It not only possesses the flexibility of FJSP, but also needs to handle unexpected changes in the production process, such as new order arrivals and machine malfunctions. This means that the scheduling plan needs to be adjusted in real time, rather than being completed all at once;

[0021] HFSP: All jobs follow the same processing path (all are pipelines), but each processing stage has a set of parallel machines;

[0022] The generalized scheduling cognitive architecture search system based on LLM reasoning includes a cognitive blueprint construction module, an intelligent candidate generation module, a cognitive deliberation module, and an architecture evaluation module.

[0023] The cognitive blueprint building module is used to transform scheduling problems into cognitive blueprints of target scheduling tasks;

[0024] The formula used by the intelligent candidate generation module is:

[0025] ;

[0026] in:

[0027] This represents the final pool of elite candidate architectures, which contains the set of architectures that are ultimately selected to advance to the next stage.

[0028] This represents the selection function that selects the top K candidate architectures based on a given scoring criterion.

[0029] This represents the set of first candidate architectures recommended by Bayesian optimization (BO) based on the acquisition function.

[0030] This represents the set of second candidate architectures selected by the selector agent based on maximizing structural diversity.

[0031] This represents the average reward score of the candidate architecture;

[0032] The score represents the structural novelty of the candidate architecture;

[0033] This represents a coefficient that balances performance and exploration; it's a trade-off parameter. When searching for network architectures, the system faces a classic dilemma: should it choose the model with the highest current prediction score that seems the most secure (i.e., leveraging performance)? Or choose a model that, although unfamiliar, has a unique structure and may hold greater potential for breakthroughs (i.e., exploring the unknown, corresponding to...) ).coefficient It's used to adjust this. Turn it up. The system encourages exploring novel architectures, which effectively prevents the search process from getting trapped in local optima prematurely; adjusting the size... The system, on the other hand, places more emphasis on performance utilization;

[0034] Existing technologies (traditional BO or NAS) typically only use acquisition functions, which are purely based on numerical statistics (mean and variance), completely ignoring the "structural semantics" or "design novelty" of the architecture itself.

[0035] The intelligent candidate generation module of this invention uses a hybrid formula that forces the adoption of high-scoring mathematical predictions. With LLM's view of unique architecture To perform a union, we introduce... The item explicitly uses "novelty" as a core weight in the ranking, rather than just as a secondary factor.

[0036] Traditional business optimization (BO) is prone to getting stuck in local optima because it tends to recommend architectures similar to known good architectures.

[0037] This invention injects structural diversity: the introduction of LLM selection and novelty weighting is intended to force the exploration of architectures that are "not necessarily the highest in current prediction scores but have very unique structures" in the early stages of the search, which is crucial for discovering entirely new design patterns across tasks (such as migrating from JSSP to FJSP).

[0038] The cognitive deliberation module receives historical architecture performance data sets, a cognitive blueprint of the target scheduling task, and an elite candidate architecture pool as input. Through multi-agent collaboration and a multi-stage feedback deliberation mechanism, it outputs a structurally sound final graph neural network architecture to the architecture evaluation module. Traditional Neural Architecture Search (NAS) is a black-box optimization method, relying solely on random mutations for blind trial and error, lacking an understanding of "why to change." Existing large-model-based optimization methods (such as FunSearch and ReEvo) merely treat LLM as a one-way "code generator" of lightweight heuristic rules. Compared to this invention, they all lack the "cognitive reasoning" ability to diagnose problems, propose logical hypotheses, and conduct peer review to avoid ineffective modifications.

[0039] The architecture evaluation module takes the final graph neural network architecture and historical architecture performance dataset as input, and outputs performance evaluation results to update the system knowledge base through model training and performance feedback. When the search termination condition is met, it outputs the optimal graph neural network architecture.

[0040] As a preferred technical solution:

[0041] As described above, in a job scheduling optimization method based on an LLM inference architecture, the intelligent candidate generation module receives an initial candidate architecture pool, a set of historical architecture performance data, and a cognitive blueprint of the target scheduling task as input, and outputs an elite candidate architecture pool to the cognitive deliberation module.

[0042] The job scheduling optimization method based on LLM inference architecture described above includes the following steps in the execution process of the intelligent candidate generation module:

[0043] (1) Performance-oriented generation: using historical architecture performance data sets A Bayesian-optimized surrogate model is trained, taking candidate architectures as input and actual reward scores as theoretical output. The Bayesian-optimized surrogate model then predicts the reward scores of candidate architectures in the initial candidate architecture pool, obtaining the average reward score. Based on the principle of maximizing the collection function, the average reward score is extracted to form the first candidate architecture set. That is, by using a proxy model to calculate the acquisition function value of each architecture in the pool, and sorting them from highest to lowest score, the top-scoring architectures are directly selected. These architectures constitute the first candidate architecture set. ;

[0044] (2) Structural diversity generation: The selector agent based on the Large Language Model (LLM) is invoked. The selector agent uses the cognitive blueprint of the target scheduling task as the reasoning context to calculate the structural novelty score of each candidate architecture in the initial candidate architecture pool relative to the historical architecture performance data set. Based on the principle of maximizing structural diversity, candidate architectures are selected to form the second candidate architecture set. ;

[0045] (3) Hybrid evaluation and merging: The first candidate architecture set With the second candidate architecture set A union operation is performed, and the merged candidate architecture set is weighted and ranked using the average reward score and structural novelty score. The top K candidate architectures are selected to generate an elite candidate architecture pool. K is a preset hyperparameter used to control the search width. It determines the number of elite architectures that enter the review stage in each iteration. In practical applications, the specific value of K is dynamically set according to the actual available computing power resources and the size of the search space.

[0046] The job scheduling optimization method based on LLM inference architecture described above includes the following steps in the specific execution process of the cognitive deliberation module:

[0047] (a) Pattern recognition and diagnosis stage: The analyst agent is invoked. The analyst agent uses the historical architecture performance data set and the cognitive blueprint of the target scheduling task as the reasoning context (the analyst agent needs to identify patterns by combining the historical architecture performance data set and use the structural features of the cognitive blueprint for diagnosis), identifies the synergistic effects between architecture components and cross-task performance patterns, and outputs a structured analysis report.

[0048] (b) Hypothesis-driven proposal phase: The proposer agent is invoked. The proposer agent integrates the elite candidate architecture pool, the structured analysis report, and the current search strategy to generate a hypothesis-driven architecture modification proposal. The architecture modification proposal includes instructions for changing the network topology of the source architecture and its corresponding logical verification chain.

[0049] (c) Logic verification and closed-loop review stage: The evaluator agent is invoked to conduct a multi-dimensional review of the proposed architecture modification, verifying the rationality of its logic verification chain and its matching degree with the current search strategy.

[0050] The cognitive deliberation module of this invention is a multi-agent collaborative reasoning system, comprising an analyst agent, a proposer agent, and an evaluator agent. These three roles engage in peer-review style deliberation, offering the following advantages:

[0051] ① High sample efficiency: It transforms blind search into reasoned deduction. The state-of-the-art architecture can be found in just 50 iterations, while traditional methods require hundreds or thousands of iterations.

[0052] ② Explainability: Every modification has a reason, and engineers can understand why this network was designed.

[0053] ③ Ensuring validity: The evaluator mechanism significantly reduces the probability of generating invalid code.

[0054] In the job scheduling optimization method based on LLM inference architecture described above, in step (c), if the review is approved or the preset maximum number of review rounds is reached, the architecture modification proposal is instantiated into the final graph neural network architecture and output to the architecture evaluation module.

[0055] Conversely, the evaluator agent generates corrective feedback containing diagnostic information and sends it back to the proposer agent, triggering the next phase of the architecture refactoring proposal.

[0056] The job scheduling optimization method based on LLM inference architecture described above includes the following steps in the specific execution process of the architecture evaluation module:

[0057] (I) Model training and performance acquisition: The final graph neural network architecture is trained with a policy to obtain its actual completion time. The actual completion time is compared with the completion time of the baseline method to obtain the actual reward score.

[0058] (II) Knowledge base closed-loop update: The final graph neural network architecture and its corresponding actual reward score are used as new samples to update the historical architecture performance data set;

[0059] (III) Residual-based similarity update and transfer: The difference between the actual reward score corresponding to the final graph neural network architecture and the reward score predicted by the Bayesian optimized surrogate model is calculated to obtain the prediction residual. Then, the dynamic learnable similarity based on the prediction residual and the static structural similarity based on the cognitive blueprint are used to calculate the hybrid task similarity matrix. The hybrid task similarity matrix, by fusing the static structural similarity based on the cognitive blueprint and the dynamic learnable similarity based on the prediction residual, characterizes the correlation between the target scheduling task and the historical task, thereby guiding cross-task knowledge transfer in subsequent searches.

[0060] To achieve cross-task knowledge transfer (such as migrating from JSSP to FJSP—the flexible job shop scheduling problem), the system employs a hybrid kernel function to calculate the comprehensive similarity between tasks at the underlying Bayesian optimization layer. The hybrid kernel function is composed of a weighted fusion of static structural similarity based on cognitive blueprint feature extraction and dynamically updated learnable task similarity. Through this mechanism, even when faced with an unseen target scheduling task, the system can discover its structural commonalities with historical tasks through the cognitive blueprint, thereby directly reusing high-performance architecture fragments found in the past and achieving zero-sample or few-sample experience transfer.

[0061] (IV) Iteration and final output: The historical architecture performance data set updated in step (II) and the hybrid task similarity matrix in step (III) (these data will serve as prior knowledge for the next round of Bayesian optimization) are fed back to the intelligent candidate generation module for the next round of architecture search iteration. After the hybrid task similarity matrix is ​​fed back to the intelligent candidate generation module, the structure and parameters of the Bayesian optimized surrogate model will be adjusted. When the preset maximum number of iterations is reached, the graph neural network architecture with the best actual reward score and its parameters are selected from all the evaluated architectures.

[0062] As described above, the job scheduling optimization method based on LLM inference architecture uses reinforcement learning (RL) for model training in step (I). Specifically, it uses the near-end policy optimization algorithm as the backbone solver to perform end-to-end policy training on the generated GNN architecture.

[0063] Invention principle:

[0064] This invention presents a job scheduling optimization method based on the LLM inference architecture, which can break through the limitations of traditional scheduling algorithms and meet the highly diversified and dynamic adaptive scheduling needs of modern industry. Specifically, it establishes adaptive capability through the following three core improvements:

[0065] Firstly, cognitive blueprints enable unified dimensionality reduction and abstraction of non-standard scenarios.

[0066] Real-world scheduling scenarios are highly varied. For example, some workshops have fixed machines (JSSP), some workshops have multiple machines available for a single process (FJSP), and some even incorporate Automated Guided Vehicles (AGVs) for transportation (DyFJSP). Traditional methods require redesigning the graph structure and network for each new constraint.

[0067] This invention proposes a hypergraph-based "cognitive blueprint," which can uniformly abstract various entities or complex cross-constraints in an industrial scenario into "nodes" and "hyperedges." The system does not rely on hard-coded rules but instead "perceives" the scenario by extracting the topological statistical features of hyperedges (e.g., a higher cardinality of machine constraint hyperedges indicates greater processing flexibility in the workshop). This unified meta-representation enables large language models to understand the essential differences in the underlying structures of different factories, much like human experts, providing a universal foundation for adaptively generating custom architectures.

[0068] Secondly, the multi-agent cognitive engine endows the system with the ability to perform logical deductions in a context-specific manner.

[0069] Existing network architectures are static or based on random mutations using evolutionary algorithms. When faced with entirely new industrial scenarios, random mutations are highly likely to generate invalid architectures, resulting in extremely low search efficiency and an inability to adapt to rapid production line changes.

[0070] This invention introduces a multi-agent deliberation mechanism consisting of an analyst, proposer, and evaluator. When a new industrial scenario is input, the agents do not guess the network structure but instead perform deductive reasoning. For example, by reading the cognitive blueprint of the scenario, the analyst discovers that it contains a high cardinality of machine-constrained hyperedges (implying highly flexible choices). Based on this feature, the proposer will logically hypothesize: "This scenario requires the introduction of a deeper attention mechanism to handle complex machine-selective routing." The evaluator then verifies this logic. This mechanism ensures that the network architecture generated by the system for any new scenario is tailored and highly effective.

[0071] Third, the hybrid task similarity assessment breaks down knowledge barriers across scenarios, enabling zero-sample / few-sample transfer.

[0072] Traditional algorithms often require collecting a large amount of data to train from scratch when facing new factories and new business lines, which is extremely costly in actual industry.

[0073] This invention innovatively introduces "hybrid task similarity" into the Bayesian-optimized kernel function. When faced with a completely unfamiliar and complex scheduling scenario, the system can automatically calculate the structural similarity between the new scenario and historical scenarios. If the underlying logic is found to be similar, the system will directly migrate the high-performance architecture genes accumulated in the past for a hot start. Experiments show that this system can achieve a 52.20% reduction in completion time in new scenarios (compared to the expert baseline) without having to spend huge amounts of computing power to search again, truly achieving adaptive generalization of "learning by analogy".

[0074] Beneficial effects:

[0075] (1) The job scheduling optimization method based on LLM inference architecture of the present invention can realize intelligent search and cross-task generalization of the scheduling architecture through the collaborative design of "cognitive blueprint modeling + intelligent candidate generation module + multi-agent collaboration and multi-stage feedback review mechanism + Bayesian optimization".

[0076] (2) The job scheduling optimization method based on LLM reasoning architecture of the present invention can automatically design a suitable network architecture according to the characteristics of the problem through "cognitive blueprint" and LLM reasoning, and automatically discover a network topology pattern that is more suitable for the dynamic characteristics of scheduling, thereby having a stronger ability to express complex relationships; by adopting a hypergraph structure, the structural laws of different scheduling problems are explicitly abstracted, and knowledge transfer is realized.

[0077] (3) The job scheduling optimization method based on LLM inference architecture of the present invention uses a large language model as a cognitive engine and makes rational architectural modifications by analyzing performance bottlenecks (such as "insufficient attention layer leading to loss of long-distance dependencies") rather than random mutations.

[0078] (4) The job scheduling optimization method based on LLM inference architecture of the present invention innovatively introduces "hybrid task similarity" into the hybrid kernel function of Bayesian optimization. On tasks that have not participated in the search, the completion time can be reduced directly without re-searching, which completely breaks the knowledge barrier across tasks and realizes zero-sample / few-sample transfer.

[0079] (5) Traditional methods based on reinforcement learning or evolutionary algorithms require hundreds or thousands of iterations for evaluation, while the present invention only requires 50 iterations to discover a better architecture, which greatly reduces the search cost. Attached Figure Description

[0080] Figure 1 This is a schematic diagram illustrating the construction process of the generalized scheduling cognitive architecture search system based on LLM inference according to the present invention;

[0081] Figure 2This is a comparison chart showing the search efficiency of the generalized scheduling cognitive architecture search system (CogNAS Transfer) based on LLM inference of this invention and existing technology models (ABG and GNAS). Detailed Implementation

[0082] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0083] A job scheduling optimization method based on LLM inference architecture is proposed. The method constructs a generalized scheduling cognitive architecture search system based on LLM inference using labeled scheduling problem instances and outputs the optimal graph neural network architecture. Then, unlabeled scheduling problem instances are input into the optimal graph neural network architecture to obtain the optimal scheduling scheme. The job is then executed according to the optimal scheduling scheme.

[0084] The instances of labeled scheduling problems and the instances of unlabeled scheduling problems are each independently selected from one of JSSP, FJSP, DyFJSP, and HFSP.

[0085] like Figure 1 As shown, the generalized scheduling cognitive architecture search system based on LLM reasoning includes a cognitive blueprint construction module, an intelligent candidate generation module, a cognitive deliberation module, and an architecture evaluation module.

[0086] The cognitive blueprint building module is used to transform scheduling problems into cognitive blueprints of target scheduling tasks; the specific process is as follows:

[0087] The system determines the entity types and constraint relationship types involved in the target scheduling task based on its type and instance constraint characteristics. Entity types include processes and machines; in DyFJSP, AGVs and dynamic events are also included. Constraint relationship types include job priority constraints, machine selectability constraints, and resource contention constraints; in DyFJSP, transportation constraints and dynamic event constraints are also included. These constraints are not additional input files but are identified from the structural characteristics of the target scheduling task and organized into hyperedge types and their attributes in the cognitive blueprint. Subsequently, the system generates a cognitive blueprint of the target scheduling task based on the scheduling problem. The system extracts structural features such as the number of entities, whether AGVs are included, machine selectability, resource contention intensity, transportation constraints, and dynamic events, which are then used for LLM reasoning and task similarity calculation.

[0088] The cognitive blueprint building module serves as a "bridge" for LLM to understand mathematical problems, enabling agents to reason based on structural laws.

[0089] The intelligent candidate generation module receives the initial candidate architecture pool, the historical architecture performance data set, and the cognitive blueprint of the target scheduling task as input, and outputs the elite candidate architecture pool to the cognitive review module.

[0090] The specific execution process of the intelligent candidate generation module includes the following steps:

[0091] (1) Performance-oriented generation: A Bayesian optimization surrogate model is trained using a historical architecture performance dataset. During training, candidate architectures are used as inputs and actual reward scores are used as theoretical outputs. The Bayesian optimization (BO) surrogate model is then used to predict the reward scores of candidate architectures in the initial candidate architecture pool to obtain the mean reward score. Based on the principle of maximizing the collection function, the mean reward score is extracted to form the first candidate architecture set. ;

[0092] (2) Structural diversity generation: The selector agent based on a large language model is invoked. The selector agent uses the cognitive blueprint of the target scheduling task as the reasoning context to calculate the structural novelty score of each candidate architecture in the initial candidate architecture pool relative to the historical architecture performance data set. Based on the principle of maximizing structural diversity, candidate architectures are selected to form the second candidate architecture set. ;

[0093] (3) Hybrid evaluation and merging: The first candidate architecture set With the second candidate architecture set Perform a union operation, and use the mean of the reward score and the structural novelty score to perform a weighted evaluation and ranking of the merged candidate architecture set, and extract the top K candidate architectures to generate an elite candidate architecture pool.

[0094] The formula used by the intelligent candidate generation module is:

[0095] ;

[0096] in:

[0097] This represents the final pool of elite candidate architectures.

[0098] This represents the selection function that selects the top K candidate architectures based on a given scoring criterion.

[0099] This represents the set of first candidate architectures recommended by Bayesian optimization based on the acquisition function.

[0100] This represents the set of second candidate architectures selected by the selector agent based on maximizing structural diversity.

[0101] The mean reward score of the candidate architecture is represented by the Bayesian optimized surrogate model's prediction of the reward score of the candidate architecture, which is not the arithmetic mean of multiple runs.

[0102] The score represents the structural novelty of the candidate architecture;

[0103] This represents a coefficient that balances performance and exploration. In specific implementation cases, this coefficient is a preset hyperparameter, and its value is usually dynamically set according to the actual computing power and the exploration needs of the target scenario. It is used to adjust the system's weight preference between utilizing the current high-resolution architecture and exploring the novelty of unknown structures.

[0104] The cognitive deliberation module receives historical architecture performance data sets, cognitive blueprints of target scheduling tasks, and an elite candidate architecture pool as inputs. Through multi-agent collaboration (analyst agent, proposer agent, and evaluator agent) and a multi-stage feedback deliberation mechanism, it outputs a structurally sound final graph neural network architecture to the architecture evaluation module.

[0105] The specific implementation process of the cognitive deliberation module includes the following steps:

[0106] (a) Pattern recognition and diagnosis stage: The analyst agent is invoked. The analyst agent uses the historical architecture performance data set and the cognitive blueprint of the target scheduling task as the reasoning context to identify the synergistic effects between architecture components and cross-task performance patterns, and outputs a structured analysis report.

[0107] (b) Hypothesis-driven proposal phase: The proposer agent is invoked. The proposer agent integrates the elite candidate architecture pool, the structured analysis report, and the current search strategy to generate a hypothesis-driven architecture modification proposal. The architecture modification proposal includes instructions for changing the network topology of the source architecture and its corresponding logical verification chain.

[0108] (c) Logic verification and closed-loop review stage: The evaluator intelligent agent is invoked. The evaluator intelligent agent conducts a multi-dimensional review of the proposed architecture modification, verifying the rationality of its logical verification chain and its matching degree with the current search strategy.

[0109] If the review is approved or the preset maximum number of review rounds is reached, the architecture modification proposal will be instantiated into the final graph neural network architecture and output to the architecture evaluation module.

[0110] Conversely, the evaluator agent generates corrective feedback containing diagnostic information and sends it back to the proposer agent, triggering the next phase of the architecture refactoring proposal.

[0111] The architecture evaluation module takes the final graph neural network architecture and historical architecture performance dataset as input, outputs performance evaluation results to update the system knowledge base through model training and performance feedback, and outputs the optimal graph neural network architecture when the search termination condition is met.

[0112] The specific execution process of the architecture evaluation module includes the following steps:

[0113] (I) Model training and performance acquisition: The final graph neural network architecture is trained with a policy to obtain its actual completion time. The actual completion time is compared with the completion time of the baseline method to obtain the actual reward score.

[0114] The model is trained using reinforcement learning methods, specifically by using the proximal policy optimization algorithm as the backbone solver to perform end-to-end policy training on the generated GNN architecture.

[0115] (II) Knowledge base closed-loop update: The final graph neural network architecture and its corresponding actual reward score are used as new samples to update the historical architecture performance data set;

[0116] (III) Residual-based similarity update and transfer: The difference between the actual reward score corresponding to the final graph neural network architecture and the reward score predicted by the Bayesian optimized surrogate model is calculated to obtain the prediction residual. Then, the dynamic learnable similarity based on the prediction residual and the static structural similarity based on the cognitive blueprint are used to calculate the hybrid task similarity matrix. The hybrid task similarity matrix, by fusing the static structural similarity based on the cognitive blueprint and the dynamic learnable similarity based on the prediction residual, characterizes the correlation between the target scheduling task and the historical task, thereby guiding cross-task knowledge transfer in subsequent searches.

[0117] (IV) Iteration and final output: The updated data from steps (II) and (III) are fed back to the intelligent candidate generation module for the next round of architecture search iteration. After the hybrid task similarity matrix is ​​fed back to the intelligent candidate generation module, the structure and parameters of the Bayesian optimized surrogate model are adjusted. When the preset maximum number of iterations is reached, the graph neural network architecture with the best actual reward score and its parameters are selected from all the evaluated architectures.

[0118] Example 1

[0119] A job scheduling optimization method based on LLM inference architecture, the specific steps of which are as follows:

[0120] (1) Identify instances of label scheduling problems;

[0121] The labeled scheduling problem example is the JSSP case, specifically using all samples from the training set of the comprehensive scheduling benchmark dataset (Taillard Benchmarks) proposed by Éric Taillard in 1993, with each sample being a JSSP case;

[0122] (2) Construct a generalized scheduling cognitive architecture search system based on LLM inference using a labeled scheduling problem example, and output the optimal graph neural network architecture; the specific process is as follows:

[0123] (2.1) Execution of the cognitive blueprint construction module;

[0124] Transforming labeled scheduling problem instances into a cognitive blueprint for target scheduling tasks. ;

[0125] (2.2) The intelligent candidate generation module executes;

[0126] Input: Graph neural network architecture pool C (i.e., initial candidate architecture pool C), cognitive blueprint of the target scheduling task. and historical architecture performance data sets ;

[0127] The initial candidate architecture pool C is randomly sampled from the predefined architecture search space. The predefined architecture search space includes candidate GNN operators of each edge type, network layer number and entity embedding update order. In each search iteration, 100 candidate architectures are randomly sampled from the predefined architecture search space. The entire search process iterates for 50 rounds.

[0128] The first set of historical architecture performance data input to the intelligent candidate generation module The acquisition process is as follows: random sampling from the predefined architecture search space Twenty candidate architectures were selected, and their actual completion times were obtained through training and evaluation. These actual completion times were then compared with the completion times of the baseline method to obtain actual reward scores. These 20 candidate architectures and their actual reward scores constituted the historical architecture performance data set initially input into the intelligent candidate generation module. ;

[0129] In subsequent processes, historical architecture performance data sets This will be continuously updated; see below for details.

[0130] The completion time of the baseline method comes from the reference makespan obtained by expert-designed reference architecture on the same task under the same training and evaluation process. It can be understood as the reference completion time under the same evaluation criteria. The makespan represents the minimum completion time of the architecture discovered by the method on the test dataset up to a certain evaluation. The makespan is still used when explaining the scheduling effect because it is the most intuitive and the smaller the value, the better the scheduling effect.

[0131] Processing procedure:

[0132] (2.2.1) Performance-oriented generation: using historical architecture performance data sets A Bayesian-optimized surrogate model is trained, taking candidate architectures as input and actual reward scores as theoretical outputs. The Bayesian-optimized surrogate model then predicts the reward scores of candidate architectures in the initial candidate architecture pool C, yielding the average reward score. Based on the principle of maximizing the acquisition function, the first candidate architecture set is constructed. ;

[0133] (2.2.2) Generation of structural diversity: Invoking a selector agent based on a large language model, the selector agent uses the cognitive blueprint of the target scheduling task. As the inference context, the performance data set of each candidate architecture relative to the historical architecture is calculated in the initial candidate architecture pool C. Structural novelty score Based on the principle of maximizing structural diversity, candidate architectures were selected to form the second candidate architecture set. ;

[0134] (2.2.3) Mixed evaluation and merging: based on the formula The first candidate architecture set With the second candidate architecture set A union operation is performed, and the merged candidate architecture set is weighted and ranked using the mean of the reward score and the structural novelty score. The top K (10) candidate architectures are selected to generate an elite candidate architecture pool. ;

[0135] Output: Elite candidate architecture pool ;

[0136] (2.3) Execution of the cognitive deliberation module;

[0137] Input: Elite candidate architecture pool Historical architecture performance data set Cognitive blueprint for target scheduling tasks ;

[0138] Processing procedure:

[0139] (2.3.1) Pattern Recognition and Diagnosis Phase: Invoke the analyst agent, which uses historical architecture performance data sets. And a cognitive blueprint for target scheduling tasks As a reasoning context, it identifies the synergistic effects between architectural components and cross-task performance patterns, and outputs a structured analysis report. For example, the report points out that "in scenarios with dense resource conflicts, having more than 4 network layers will lead to performance degradation."

[0140] (2.3.2) Hypothesis-driven proposal phase: Invoking the proposer agent, the proposer agent synthesizes the elite candidate architecture pool Based on the structured analysis report and the current search strategy, hypothesis-driven architecture modification proposals are generated, such as "reduce the number of network layers from 4 to 3". The architecture modification proposals include instructions for changing the network topology of the source architecture and their corresponding logical verification chains.

[0141] (2.3.3) Logic verification and closed-loop review stage: The evaluator intelligent agent is invoked. The evaluator intelligent agent conducts a multi-dimensional review of the architecture modification proposal, verifying the rationality of its logical verification chain and its matching degree with the current search strategy.

[0142] If the review is approved or the preset maximum number of review rounds is reached, the architecture modification proposal will be instantiated into the final graph neural network architecture and output to the architecture evaluation module.

[0143] Conversely, the evaluator agent generates corrective feedback containing diagnostic information and sends the corrective feedback back to the proposer agent, triggering the next stage of architecture refactoring proposal.

[0144] Output: The final graph neural network architecture after logical verification and multiple rounds of correction, for example, a 3-layer heterogeneous graph neural network configuration with an attention mechanism;

[0145] (2.4) Execution of the architecture evaluation module;

[0146] Input: Final graph neural network architecture, historical architecture performance data set ;

[0147] Processing procedure: A near-end policy optimization algorithm is adopted, with the minimum completion time as the reward function. The final graph neural network architecture is trained to obtain its actual completion time. The actual completion time is compared with the completion time of the baseline method to obtain the actual reward score. The final graph neural network architecture and its corresponding actual reward score are used as new samples to update the historical architecture performance data set. The difference between the actual reward score and the reward score predicted by the surrogate model is calculated to obtain the prediction residual. Then, the dynamic learnable similarity based on the prediction residual and the static structural similarity based on the cognitive blueprint are used to calculate the hybrid task similarity matrix.

[0148] The formula for calculating the similarity matrix of hybrid tasks is as follows:

[0149] Smixed(T,T')=α·Sstatic(T,T')+(1-α)·Slearnable(T,T');

[0150] Sstatic(T,T')=cos(φ(B(T)),φ(B(T')));

[0151] Slearnable(T,T')=Corr({εT,i},{εT',j});

[0152] εT,i=yT,i-μt(aT,i);

[0153] εT',j=yT',j-μt(aT',j);

[0154] Where Smixed(T,T') represents the mixed task similarity matrix between task T and task T';

[0155] α represents the weight between static structural similarity and dynamic learnable similarity, α=0.4;

[0156] Sstatic(T,T') represents the static structural similarity based on the cognitive blueprint;

[0157] Slearnable(T,T') represents a dynamic learnable similarity based on the prediction residuals;

[0158] φ(B(T)) represents the cognitive blueprint feature vector of task T;

[0159] φ(B(T')) represents the cognitive blueprint feature vector of task T';

[0160] Corr(·) indicates correlation calculation;

[0161] εT,i represents the actual reward score of the final graph neural network architecture on task T. Reward score predicted by the proxy model The predicted residuals between;

[0162] εT',j represents the actual reward score of the final graph neural network architecture on task T'. Reward score predicted by the proxy model The predicted residuals between;

[0163] Output feedback: Output a hybrid task similarity matrix and compare it with the updated historical architecture performance data set. The results are fed back to the intelligent candidate generation module for the next round of architecture search iteration. After the hybrid task similarity matrix is ​​fed back to the intelligent candidate generation module, the structure and parameters of the Bayesian optimized surrogate model are adjusted. Each element of the hybrid task similarity matrix corresponds to the hybrid task similarity matrix between two tasks and is used as the input of the task kernel ktask(T,T') in Bayesian optimization for cross-task transfer. The search terminates after 50 iterations and outputs the optimal graph neural network architecture and weight parameters. The final output weight parameters refer to the learnable parameters obtained by training the optimal GNN / RL policy model, such as message passing layer weights, embedding update layer weights, policy network or value network parameters, etc.

[0164] (3) Identify instances of the unlabeled scheduling problem;

[0165] The example of the unlabeled scheduling problem is the JSSP case, specifically using a sample from the test set of the comprehensive scheduling benchmark dataset (Taillard Benchmarks) proposed by Éric Taillard in 1993, with a size of 100 (number of jobs) × 20 (number of machines), i.e., the Taillard 100×20 test group;

[0166] (4) After converting the unlabeled scheduling problem instance into a cognitive blueprint, input it into the optimal graph neural network architecture obtained in step (2), and obtain the optimal scheduling scheme from its output, and execute the job according to the optimal scheduling scheme;

[0167] The optimal scheduling scheme is as follows: the schedulable operations are selected step by step by the optimal graph neural network architecture to form a schedule table that includes the processing machine for each operation, the machine processing sequence, the start time and the end time; the schedule satisfies the order constraint within the operation and the machine capacity constraint.

[0168] The final completion time was 5363.4 time units (this result is the average of 5 runs), while the average time for the Taillard 100×20 test group was 5459.3 time units.

[0169] Comparative Example 1

[0170] A job scheduling optimization method uses a reinforcement learning scheduling model (RS) that is manually tuned by human experts for this scenario to generate the optimal scheduling scheme for the unlabeled scheduling problem instance in Example 1.

[0171] The final completion time was approximately 5442.2 time units (this result is the average of 5 runs), while the average time for the Taillard 100×20 test group was 5580.5 time units.

[0172] Comparative Example 2

[0173] A job scheduling optimization method is proposed, which uses the LLM-based rule optimization method ReEvo to generate the optimal scheduling scheme for the unlabeled scheduling problem instance in Example 1.

[0174] The final completion time was approximately 5519.8 time units (this result is the average of 5 runs), while the average time for the Taillard 100×20 test group was 5663.8 time units.

[0175] Comparative Example 3

[0176] A job scheduling optimization method employs ABG (ABG-NAS, Adaptive Bayesian Genetic Neural Architecture Search) to generate the optimal scheduling scheme for the unlabeled scheduling problem instance of Example 1.

[0177] The final completion time was approximately 5448.8 time units (this result is the average of 5 runs), while the average time for the Taillard 100×20 test group was 5492.3 time units.

[0178] Comparative Example 4

[0179] A job scheduling optimization method uses GNAS (Graph Neural Architecture Search) to generate the optimal scheduling scheme for the unlabeled scheduling problem instance of Example 1.

[0180] The final completion time was approximately 5589.6 time units (this result is the average of 5 runs), while the average time for the Taillard 100×20 test group was 5681.1 time units.

[0181] like Figure 2As shown in the figure, this invention's CogNAS Transfer, the prior art ABG, and GNAS compare their sample efficiency in the early stages of architecture search. CogNAS Transfer rapidly reduces the optimal makespan within its preset budget of 50 evaluations and stops searching after the 50th evaluation; the red dashed line after 50 evaluations in the figure represents the best result achieved by CogNAS Transfer when it stops searching, not the continuation of new architecture evaluations. In contrast, ABG and GNAS decrease at a slower rate within the same early evaluation interval, requiring further evaluation of more candidate architectures to approach the results of CogNAS Transfer. This figure illustrates that CogNAS Transfer can identify high-quality architectures faster with a smaller evaluation budget, thereby reducing the expensive candidate architecture training and validation costs in the NAS search phase.

[0182] Therefore, it can be concluded that the optimal graph neural network architecture searched by the system of the present invention has extremely high quality and perfectly matches the optimization terrain of complex scheduling problems. This not only enables the model to complete learning and training at a faster speed and with less sample cost, but also shows excellent training stability, thereby greatly improving the solution efficiency and reliability in practical industrial applications.

[0183] Example 2

[0184] A job scheduling optimization method based on LLM inference architecture, the specific steps of which are as follows:

[0185] (1) Identify instances of the unlabeled scheduling problem;

[0186] The unlabeled scheduling problem example is the FJSP case, specifically using a sample from the test set of the Large group of the Turkish Factory dataset, with a size of 20 (number of jobs) × 26 (number of machines).

[0187] (2) After converting the unlabeled scheduling problem instance into a cognitive blueprint, it is input into the optimal graph neural network architecture obtained in step (2) of Example 1. The optimal scheduling scheme is obtained from its output, and the job is executed according to the optimal scheduling scheme.

[0188] The optimal scheduling scheme is to output the candidate machine selection, machine processing sequence, start time and end time for each process. This scheme is suitable for the resource allocation characteristics of the Turkish factory, where the data is concentrated, machine compatibility is sparse, and the number of machines usually exceeds the number of jobs.

[0189] The final completion time is approximately 21 time units (this result is the average of 5 runs), while the average for the Large group of the Türkiye factory dataset is 43.98 time units.

[0190] Comparative Example 5

[0191] A job scheduling optimization method is proposed, which uses an expert-designed FJSP scheduling model to solve the unlabeled scheduling problem example in Example 2. This method is an expert baseline model designed manually for the FJSP scenario, which can directly handle machine selectability, process route constraints, and processing time information in the target FJSP instance to generate machine selection and process sequencing schemes.

[0192] The final completion time is approximately 22 time units (this result is the average of 5 runs), while the average for the Large group of the Türkiye factory dataset is 44.10 time units.

[0193] Example 3

[0194] A job scheduling optimization method based on LLM inference architecture, the specific steps of which are as follows:

[0195] (1) Identify instances of the unlabeled scheduling problem;

[0196] The unlabeled scheduling problem example is the HFSP case, specifically using a sample from the test set of the Carlier dataset, with a size of 15 (number of jobs) × 5 (number of machines); specifically j15c5d1; where j15 represents 15 jobs, c5 represents 5 stages / machine groups, and d1 represents the first instance of class d;

[0197] (2) After converting the unlabeled scheduling problem instance into a cognitive blueprint, it is input into the optimal graph neural network architecture obtained in step (2) of Example 1. The optimal scheduling scheme is obtained from its output, and the job is executed according to the optimal scheduling scheme.

[0198] The optimal scheduling scheme is to output the machine selection, processing order, start time and end time of each job in each stage; this scheme satisfies the constraint that all jobs are processed in a fixed stage sequence and that the capacity of the same machine is mutually exclusive.

[0199] The final completion time was approximately 179.0 time units (this result is the average of 5 runs), while the average for the J15C5 group was 145.90 time units.

[0200] Comparative Example 6

[0201] A job scheduling optimization method is proposed, which uses HGRLA (a hybrid flow shop scheduling method based on heterogeneous graph neural networks and deep reinforcement learning) to solve the unlabeled scheduling problem example in Example 3 (this method is designed for HFSP with manually designed heterogeneous graph neural networks and reinforcement learning strategies).

[0202] The final completion time was approximately 217.0 time units (this result is the average of 5 runs), while the average for the J15C5 group was 201.57 time units.

[0203] Example 4

[0204] A job scheduling optimization method based on LLM inference architecture, the specific steps of which are as follows:

[0205] (1) Identify instances of the unlabeled scheduling problem;

[0206] The unlabeled scheduling problem example is the DyFJSP case. Specifically, based on the literature (An improved deep reinforcement learning approach for the dynamic job shop scheduling problem with random job arrivals. [J]. Journal of Physics: Conference Series, 2021,1848(1):012029 (8pp)), a set of test samples from the test set of the randomly generated DyFJSP dataset is set, with a size of 20 (number of jobs) × 10 (number of machines) × 3 (number of AGVs); 100 test instances are randomly generated for each size and the average results are reported.

[0207] (2) After converting the unlabeled scheduling problem instance into a cognitive blueprint, it is input into the optimal graph neural network architecture obtained in step (2) of Example 1. The optimal scheduling scheme is obtained from its output, and the job is executed according to the optimal scheduling scheme.

[0208] The optimal scheduling scheme is as follows: output the machine selection, AGV transportation connection, machine processing sequence, start time and end time for each process; this scheme satisfies the constraints of job release time, process sequence, machine capacity and transportation time.

[0209] The final completion time was approximately 3743.2 time units, which is the average completion time of 100 dynamic test instances with a scale of 20×10×3. On the dynamic test instance group with a scale of 20×10×3, the average completion time of this invention was 3743.2 time units, which is lower than DFJSP-ITR's 4966.7 time units and CogNAS Direct's (which means performing architecture search from scratch on the target task) 4047.8 time units. This indicates that the method has better average scheduling effect and cross-task adaptability in scenarios with dynamic job arrival and AGV transportation constraints.

[0210] Comparative Example 7

[0211] A job scheduling optimization method is proposed, which uses DFJSP-ITR to solve the labelless scheduling problem example in Example 4. This method specifically models the dynamic job arrival and resource constraints in DyFJSP.

[0212] The final completion time was approximately 4966.7 time units, which is the average completion time of 100 dynamic test instances with a scale of 20×10×3. This is higher than the 4047.8 time units of CogNAS Direct and the 3743.2 time units of CogNASTransfer of this invention.

Claims

1. A job scheduling optimization method based on LLM inference architecture, characterized in that: After constructing a generalized scheduling cognitive architecture search system based on LLM inference using labeled scheduling problem instances and outputting the optimal graph neural network architecture, unlabeled scheduling problem instances are input into the optimal graph neural network architecture to obtain the optimal scheduling scheme, and jobs are executed according to the optimal scheduling scheme. The instances of labeled scheduling problems and the instances of unlabeled scheduling problems are each independently selected from one of JSSP, FJSP, DyFJSP, and HFSP; The generalized scheduling cognitive architecture search system based on LLM reasoning includes a cognitive blueprint construction module, an intelligent candidate generation module, a cognitive deliberation module, and an architecture evaluation module. The cognitive blueprint construction module transforms the scheduling problem into a cognitive blueprint of the target scheduling task, enabling the agent to reason based on structural laws. The specific transformation process involves: determining the entity types and constraint relationship types involved in the target scheduling task based on its type and instance constraint characteristics; entity types include processes and machines, and in DyFJSP, AGVs and dynamic events are also included; constraint relationship types include job priority constraints, machine selectability constraints, and resource contention constraints, and in DyFJSP, transportation constraints and dynamic event constraints are also included; subsequently, a cognitive blueprint of the target scheduling task is generated based on the scheduling problem. And extract structural features such as the number of entities, whether AGVs are included, machine selectability, resource competition intensity, transportation constraints, and dynamic events from them; The formula used by the intelligent candidate generation module is: ； in: This represents the final pool of elite candidate architectures. This represents the selection function that selects the top K candidate architectures based on a given scoring criterion. This represents the set of first candidate architectures recommended by Bayesian optimization based on the acquisition function. This represents the set of second candidate architectures selected by the selector agent based on maximizing structural diversity. This represents the average reward score of the candidate architecture; The score represents the structural novelty of the candidate architecture; A coefficient representing the balance between performance and exploration; The cognitive deliberation module receives historical architecture performance data sets, cognitive blueprints of target scheduling tasks, and an elite candidate architecture pool as inputs. Through multi-agent collaboration and a multi-stage feedback deliberation mechanism, it outputs the final graph neural network architecture to the architecture evaluation module. The architecture evaluation module takes the final graph neural network architecture and historical architecture performance dataset as input, and outputs performance evaluation results to update the system knowledge base through model training and performance feedback. When the search termination condition is met, it outputs the optimal graph neural network architecture.

2. The job scheduling optimization method based on LLM inference architecture according to claim 1, characterized in that, The intelligent candidate generation module receives the initial candidate architecture pool, the historical architecture performance data set, and the cognitive blueprint of the target scheduling task as input, and outputs the elite candidate architecture pool to the cognitive review module.

3. The job scheduling optimization method based on LLM inference architecture according to claim 2, characterized in that, The specific execution process of the intelligent candidate generation module includes the following steps: (1) Performance-oriented generation: A Bayesian-optimized surrogate model is trained using a historical architecture performance dataset. During training, candidate architectures are used as inputs and actual reward scores are used as theoretical outputs. The Bayesian-optimized surrogate model then predicts the reward scores of candidate architectures in the initial candidate architecture pool to obtain the mean reward score. Based on the principle of maximizing the collection function, the mean reward score is extracted to form the first candidate architecture set. ; (2) Structural diversity generation: The selector agent based on a large language model is invoked. The selector agent uses the cognitive blueprint of the target scheduling task as the reasoning context to calculate the structural novelty score of each candidate architecture in the initial candidate architecture pool relative to the historical architecture performance data set. Based on the principle of maximizing structural diversity, candidate architectures are selected to form the second candidate architecture set. ; (3) Hybrid evaluation and merging: The first candidate architecture set With the second candidate architecture set Perform a union operation, and use the mean of the reward score and the structural novelty score to perform a weighted evaluation and ranking of the merged candidate architecture set, and extract the top K candidate architectures to generate an elite candidate architecture pool.

4. The job scheduling optimization method based on LLM inference architecture according to claim 3, characterized in that, The specific implementation process of the cognitive deliberation module includes the following steps: (a) Pattern recognition and diagnosis stage: The analyst agent is invoked. The analyst agent uses the historical architecture performance data set and the cognitive blueprint of the target scheduling task as the reasoning context to identify the synergistic effects between architecture components and cross-task performance patterns, and outputs a structured analysis report. (b) Hypothesis-driven proposal phase: The proposer agent is invoked. The proposer agent integrates the elite candidate architecture pool, the structured analysis report, and the current search strategy to generate a hypothesis-driven architecture modification proposal. The architecture modification proposal includes instructions for changing the network topology of the source architecture and its corresponding logical verification chain. (c) Logic verification and closed-loop review stage: The evaluator agent is invoked to conduct a multi-dimensional review of the proposed architecture modification, verifying the rationality of its logic verification chain and its matching degree with the current search strategy.

5. The job scheduling optimization method based on LLM inference architecture according to claim 4, characterized in that, In step (c), if the review is approved or the preset maximum number of review rounds is reached, the architecture modification proposal is instantiated into the final graph neural network architecture and output to the architecture evaluation module. Conversely, the evaluator agent generates corrective feedback containing diagnostic information and sends it back to the proposer agent, triggering the next phase of the architecture refactoring proposal.

6. The job scheduling optimization method based on LLM inference architecture according to claim 5, characterized in that, The specific execution process of the architecture evaluation module includes the following steps: (I) Model training and performance acquisition: The final graph neural network architecture is trained with a policy to obtain its actual completion time. The actual completion time is compared with the completion time of the baseline method to obtain the actual reward score. (II) Knowledge base closed-loop update: The final graph neural network architecture and its corresponding actual reward score are used as new samples to update the historical architecture performance data set; (III) Similarity update and transfer based on residuals: calculate the difference between the actual reward score corresponding to the final graph neural network architecture and the reward score predicted by the Bayesian optimized surrogate model to obtain the prediction residuals, and then use the dynamic learnable similarity based on the prediction residuals and the static structural similarity based on the cognitive blueprint to calculate the hybrid task similarity matrix. (IV) Iteration and final output: The historical architecture performance data set updated in step (II) and the hybrid task similarity matrix in step (III) are fed back to the intelligent candidate generation module for the next round of architecture search iteration. After the hybrid task similarity matrix is ​​fed back to the intelligent candidate generation module, the structure and parameters of the Bayesian optimized surrogate model will be adjusted. When the preset maximum number of iterations is reached, the optimal graph neural network architecture and its parameters are selected from all evaluated architectures.

7. The job scheduling optimization method based on LLM inference architecture according to claim 6, characterized in that, In step (I), reinforcement learning is used to train the model. Specifically, the near-end policy optimization algorithm is used as the backbone solver to perform end-to-end policy training on the generated GNN architecture.

Images