A non-parametric Bayesian-based multi-agv scheduling optimization method and system

Abstract

The application belongs to the technical field of AGV control, and discloses a multi-AGV scheduling optimization method and system based on a non-parametric Bayesian method, which comprises the following steps: obtaining a historical data set, extracting model parameters from the posterior distribution of an infinite Gaussian process in combination with a sampling method, and constructing an infinite Gaussian process proxy model; constructing a scheduling parameter space, selecting a target scheduling parameter point in a candidate scheduling parameter point set, simulating the candidate scheduling scheme by taking the target scheduling parameter point as the candidate scheduling scheme, obtaining current historical data, and outputting a final scheduling scheme. The application describes the mapping relationship between the multi-AGV scheduling parameters and the system performance through the infinite Gaussian process proxy model, realizes adaptive optimization of the scheduling parameters, reduces the dependence on manual parameter adjustment, reduces invalid exploration in combination with the expert knowledge prior and the large language model recommendation mechanism, and improves the efficiency, stability and adaptability to complex working conditions of the multi-AGV scheduling optimization.

Description

Technical Field

[0001] This invention relates to the field of AGV control technology, and in particular to a multi-AGV scheduling optimization method and system based on nonparametric Bayes. Background Technology

[0002] In intelligent manufacturing and logistics systems, multi-AGV (Automated Guided Vehicle) systems are widely used in material distribution, warehousing and handling, and production line supply. A typical multi-AGV system operates within a shared transportation network containing multiple paths, intersections, and charging stations. Multiple AGVs perform tasks in parallel within the same space, and their operating states significantly influence each other. To ensure production cycle time and system stability, the scheduling system typically needs to comprehensively determine multiple scheduling parameters, including task assignment order, path selection or path cost weighting, intersection traffic control rules, and charging triggering and charging station selection strategies. System-level performance indicators (such as average workstation waiting time, throughput per unit time, energy consumption, or their weighted combinations) are used as optimization objectives.

[0003] Existing multi-AGV scheduling methods mainly include rule-based methods and model-based methods. Rule-based methods (such as first-come-first-served, fixed priority, shortest path, fixed charging threshold, etc.) are simple to implement, but they are usually based on only local information and lack a characterization of the overall operating state of the multi-AGV system. They are prone to congestion, queuing, and increased energy consumption in peak task periods or resource conflict scenarios. Model-driven methods (such as integer programming, spatiotemporal network flow, or reinforcement learning methods) can describe scheduling constraints to a certain extent, but they often rely on accurate system modeling. In multi-AGV systems with random task arrivals, traffic disturbances, and frequent changes in operating state, model errors accumulate significantly, and the solution complexity is high, making it difficult to support rolling replanning and online optimization.

[0004] To reduce reliance on precise mechanistic models, engineering practice has increasingly treated the multi-AGV scheduling parameter optimization problem as a black-box optimization problem: that is, instead of explicitly modeling the analytical relationship between scheduling parameters and system performance, it evaluates a given set of scheduling parameters through simulation or actual workshop operation, obtains corresponding system-level performance feedback, and iteratively updates the parameters accordingly. In this approach, Bayesian optimization (BO) has been introduced as a common framework for scheduling parameter search and update. Its basic approach is to take the joint scheduling parameter vector of the multi-AGV system as input, and the system-level performance indicators obtained from simulation or actual vehicle operation as output. A surrogate model is used to learn the parameter-performance mapping relationship, and the scheduling strategy is gradually improved under a limited evaluation budget.

[0005] Existing Bayesian optimization-based multi-AGV scheduling methods typically employ a Gaussian Process (GP) as a surrogate model, combined with acquisition strategies such as Confidence Upper Bound (UCB), Expectation Improvement (EI), or Thompson Sampling (TS) to select the next set of scheduling parameters to be evaluated in each iteration. The advantage of the GP model lies in its analytically computable posterior distribution, which can simultaneously provide the predicted mean and variance, thus quantifying the performance uncertainty of scheduling parameters in situations where single-evaluation costs are high in multi-AGV systems. However, existing techniques often implicitly assume that the system-level performance feedback function is highly smooth and that the evaluation noise follows a sub-Gaussian distribution, which is often difficult to hold true in the actual operation of multi-AGV systems. For example, during peak task periods, congestion on key road sections, concentrated traffic at intersections, or queuing at charging stations, the performance feedback of multi-AGV systems often exhibits significant non-stationarity, heteroscedasticity, or even heavy-tailed distribution characteristics, leading to a significant decrease in the prediction accuracy and uncertainty assessment capability of traditional GP surrogate models.

[0006] To address the aforementioned issues, researchers have proposed several improved Bayesian optimization methods and attempted to apply them to engineering optimization problems, including multi-AGV scheduling. For example, they have enhanced the model's adaptability to changes in operating conditions by using input distortion or non-stationary kernel functions; handled heavy-tailed performance feedback caused by sudden congestion or abnormal waiting by using truncated or robust GP-UCB algorithms; and replaced GP with neural networks to improve the ability to express the relationship between complex scheduling parameters and system performance. However, in the typical complex operating scenario of multi-AGV systems, these methods often still rely on strong prior assumptions or specific noise structures, achieving improvement only under certain operating conditions and failing to maintain stable and consistent scheduling optimization performance across a wide range of operating states.

[0007] Another related research direction is nonparametric Bayesian modeling, such as hybrid GP models that rely on Dirichlet processes. In multi-AGV scheduling scenarios, such methods can be understood as characterizing system performance feedback through multiple potential operating modes, thereby reflecting system behavior to some extent under different traffic conditions or resource conflict modes. However, these methods still have limitations within the sequential sampling BO framework: their inference process typically assumes "batch simultaneous sampling," which does not fully match the online optimization requirements of successive evaluation and rolling updates in multi-AGV scheduling; at the same time, existing Dirichlet process and spatial Dirichlet process models are mostly used for static density estimation or cluster analysis, lacking deep integration with the sequential decision-making mechanism and real-time parameter update process required for multi-AGV scheduling.

[0008] From a theoretical perspective, existing Bayesian optimization-based scheduling methods largely rely on the RKHS smoothness assumption and information-theoretic inequalities for regret analysis. However, in multi-AGV systems, when there is a significant difference between the true distribution of system performance feedback and the posterior distribution of the general GP (Gross Parameters), such as due to sporadic congestion propagation, uncertain intersection deadlock resolution times, or multi-peaked or heavy-tailed feedback distributions caused by charging conflicts, the relevant regret bounds may no longer hold. Even within improved EI or TS frameworks, additional decoupling or approximation techniques are typically required to barely guarantee sublinear regret in multi-AGV scheduling problems, thus limiting the applicability of the methods in complex operating environments.

[0009] Finally, regarding computational complexity, the standard GP proxy model requires kernel matrix operations whose size increases with the number of evaluations, resulting in a computational complexity of O(n^2). 3 This approach is difficult to directly apply in scenarios requiring frequent rolling optimization or online replanning in multi-AGV scheduling. Although existing research has attempted to reduce computational overhead through sparse generalization (GP), random features, or efficient posterior sampling methods, effectively reducing computational complexity while maintaining scheduling optimization effectiveness remains a prominent issue in multi-AGV systems, which require simultaneous consideration of model expressiveness, operational stability, and real-time performance. In summary, existing multi-AGV scheduling methods, including rule-based methods, some model-driven methods, and GP-based Bayesian optimization scheduling methods, face challenges such as insufficient collaborative optimization, limited adaptability to complex performance feedback distributions, and insufficient computational scalability in complex operating environments.

[0010] Therefore, how to provide a multi-AGV scheduling optimization method and system based on nonparametric Bayes is an urgent problem to be solved. Summary of the Invention

[0011] This invention provides a multi-AGV scheduling optimization method and system based on nonparametric Bayesian methods to solve the problems mentioned above in the prior art.

[0012] According to a first aspect of the present invention, a multi-AGV scheduling optimization method based on nonparametric Bayes is provided.

[0013] In one embodiment, the multi-AGV scheduling optimization method based on nonparametric Bayesianism includes: Obtain the historical dataset of the AGV and set the number of iterations and the exploration probability. The historical dataset includes scheduling parameter points and corresponding performance feedback. Based on the historical dataset of AGVs, model parameters are extracted from the posterior distribution of the infinite Gaussian process using a sampling method, and the model parameters are used to construct a proxy model of the infinite Gaussian process. Construct a scheduling parameter space, and uniformly mark the scheduling parameter space and / or use a large oracle model to obtain a set of candidate scheduling parameter points; Based on the infinite Gaussian process surrogate model and exploration probability, target scheduling parameter points are selected from the set of candidate scheduling parameter points, and these target scheduling parameter points are used as candidate scheduling schemes. The candidate scheduling scheme is simulated to obtain the current historical data, and the current historical data is added to the historical dataset to obtain a new historical dataset. The above steps are repeated based on the new historical dataset until the preset number of iterations is reached, and then the final scheduling scheme is output.

[0014] In one embodiment, based on the historical dataset of the AGV, model parameters are extracted from the posterior distribution of the infinite Gaussian process using a sampling method, and a proxy model of the infinite Gaussian process is constructed using the model parameters, including: Calculate the posterior probability of each latent function surface of the scheduling parameter points in the historical dataset, and sample to determine the label parameters of the latent function surface of the scheduling parameter points; Based on the potential surface label parameters and combined with the preset prior knowledge, a posterior distribution is constructed and sampled to obtain the trend term parameters that characterize the global changes in performance feedback. Calculate the residual values ​​between performance feedback and trend term parameters in the historical dataset, construct the posterior distribution based on the residual values ​​and sample them to obtain the process perturbation parameters that characterize local fluctuations; An infinite Gaussian process surrogate model is constructed based on potential surface label parameters, trend term parameters, process perturbation parameters, and preset hyperparameters.

[0015] In one embodiment, the posterior probability values ​​of each latent function surface of the scheduling parameter points in the historical dataset are calculated, and random sampling is performed to determine the latent function surface label parameters of the scheduling parameter points, including: A set of independent auxiliary random variables following a Beta distribution is introduced. The weight values ​​of each latent function surface are generated by the auxiliary random variables through a stepwise decomposition method. The weight values ​​of each latent function surface are non-negative and the sum of the weight values ​​is 1. Calculate the conditional probability of the scheduling parameter point under each potential surface, and combine the weights of each potential function surface to calculate the relative posterior probability value of each potential function surface; The relative posterior probability values ​​are normalized so that the sum of the relative posterior probability values ​​of each latent function surface is 1. Based on the normalized posterior probability values, the latent function surface label parameters of the scheduling parameter points are determined using the Gibbs sampling method.

[0016] In one embodiment, based on potential surface label parameters and combined with preset prior knowledge, a posterior distribution is constructed and sampled to obtain trend term parameters characterizing the global change in performance feedback, including: Based on the latent surface label parameters, the historical dataset is classified and filtered to assign the labels to the scheduling parameter points of the same latent function surface and their corresponding performance feedback values. By utilizing pre-defined prior knowledge and combining the relationship between scheduling parameter points and their corresponding performance feedback values, the prior mean of the Gaussian process on the latent function surface is determined. The posterior distribution is constructed from the prior mean using the performance feedback value, and then sampled using the Gibbs sampling method to obtain the trend term parameters.

[0017] In one embodiment, the residual value between the performance feedback and the trend term parameter in the historical dataset is calculated, and a posterior distribution is constructed based on the residual value and sampled to obtain the process perturbation parameter characterizing local fluctuations. Based on the trend term parameter, the trend baseline value of the scheduling parameter point is calculated, and the residual value is used to calculate the performance feedback to obtain the residual value. Using the residuals as observation samples, a conditional posterior distribution is constructed and sampled using the Gibbs sampling method to obtain the process perturbation parameters.

[0018] In one embodiment, constructing a scheduling parameter space, uniformly marking the scheduling parameter space, and / or using a large oracle model to obtain a set of candidate scheduling parameter points includes: Construct a scheduling parameter space based on the dimensions of AGV operation scheduling parameters; The scheduling parameter space is standardized using pre-configured operational constraints and operational experience, and the standardized scheduling parameter space is evenly marked with points at preset intervals to obtain basic scheduling parameter points. Based on the scheduling parameter points, performance feedback, and AGV running status information in historical data, prompt words are constructed and input into the preset large language model to obtain supplementary scheduling parameter points; The basic scheduling parameter points and the supplementary scheduling parameter points are merged and removed to obtain a set of candidate scheduling parameter points.

[0019] In one embodiment, constructing the scheduling parameter space based on the scheduling parameter dimensions of AGV operation includes: The dimensions include the task assignment matrix, path cost weight, right-of-way control parameters, and charging strategy parameters.

[0020] In one embodiment, selecting a target scheduling parameter point from the set of candidate scheduling parameter points based on an infinite Gaussian process surrogate model and exploration probability, and using the target scheduling parameter point as a candidate scheduling scheme includes: The candidate scheduling parameter point set is input into the infinite Gaussian process surrogate model. The Thompson sampling method is used to randomly sample a latent function surface from the posterior distribution of the infinite Gaussian process. Using this surface, the prediction performance value of each candidate scheduling parameter point under each latent function surface is calculated. Candidate scheduling parameter points are selected using a preset exploration probability method, and these scheduling parameter points are used as target scheduling parameter points.

[0021] In one embodiment, the expression for the posterior probability value is: ; In the formula, z The latent function surface label represents the scheduling parameter point. Indicates the first k The function values ​​of a potential function surface. Represents the weight of the potential surface. y Indicates performance feedback, x Indicates the scheduling parameter point. This represents the conditional probability of the performance feedback of the scheduling parameter point under the latent function surface. p Represents probability symbols.

[0022] According to a second aspect of the present invention, a multi-AGV scheduling optimization system based on nonparametric Bayesian methods is provided.

[0023] In one embodiment, the multi-AGV scheduling optimization system based on nonparametric Bayesian methods includes: The data acquisition module is used to acquire the historical dataset of the AGV and set the number of iterations and the exploration probability. The historical dataset includes scheduling parameter points and corresponding performance feedback. The model building module is used to extract model parameters from the posterior distribution of the infinite Gaussian process based on the historical dataset of AGV, combined with the sampling method, and to build an surrogate model of the infinite Gaussian process using the model parameters. The candidate set module is used to construct the scheduling parameter space, uniformly mark the scheduling parameter space and / or use the big oracle model to obtain the candidate scheduling parameter point set; The candidate scheme module is used to select the target scheduling parameter point from the set of candidate scheduling parameter points based on the infinite Gaussian process surrogate model and exploration probability, and use the target scheduling parameter point as a candidate scheduling scheme. The feedback iteration module is used to simulate the candidate scheduling schemes, obtain the current historical data, and add the current historical data to the historical dataset to obtain a new historical dataset. Based on the new historical dataset, the module repeatedly executes the construction of an infinite Gaussian process proxy model, obtains the set of candidate scheduling parameter points, and selects the target scheduling parameter point until the preset number of iterations is reached, and then outputs the final scheduling scheme.

[0024] According to a third aspect of the present invention, a computer device is provided.

[0025] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0026] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0027] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0028] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention introduces an infinite Gaussian process surrogate model to model the mapping relationship between multi-AGV scheduling parameters and system performance, enabling adaptive optimization of scheduling parameters, reducing reliance on manual parameter tuning experience, improving the expressive power of surrogate modeling and the accuracy of posterior uncertainty characterization, and reducing ineffective exploration by introducing a collaborative optimization mechanism that combines expert knowledge prior guidance with candidate point recommendation from a large language model, thereby improving the efficiency and stability of multi-AGV optimization, enhancing scheduling performance, reducing trial and error costs, and strengthening adaptability to complex working conditions.

[0029] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0030] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0031] Figure 1 This is a flowchart illustrating a multi-AGV scheduling optimization method based on nonparametric Bayes, according to an exemplary embodiment. Figure 2 This is a block diagram illustrating the principle of a multi-AGV scheduling optimization system based on nonparametric Bayes, according to an exemplary embodiment. Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment; Figure 4 This is a schematic diagram illustrating a comparison between a nonparametric Bayesian-based multi-AGV scheduling optimization method and other methods, according to an exemplary embodiment. Detailed Implementation

[0032] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0033] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0034] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0035] Figure 1 An embodiment of a multi-AGV scheduling optimization method based on nonparametric Bayesian methods according to the present invention is shown.

[0036] In this optional embodiment, the multi-AGV scheduling optimization method based on nonparametric Bayes includes: S101. Obtain the historical dataset of the AGV and set the number of iterations and the exploration probability. The historical dataset includes scheduling parameter points and corresponding performance feedback. S102. Based on the historical dataset of AGV, model parameters are extracted from the posterior distribution of the infinite Gaussian process using the sampling method, and the model parameters are used to construct an surrogate model of the infinite Gaussian process. In this optional embodiment, based on the historical dataset of the AGV, model parameters are extracted from the posterior distribution of the infinite Gaussian process using a sampling method, and an infinite Gaussian process surrogate model is constructed using the model parameters. This includes: calculating the posterior probability of each latent function surface of the scheduling parameter point in the historical dataset and sampling it to determine the latent function surface label parameters of the scheduling parameter point; constructing a posterior distribution based on the latent surface label parameters and combined with preset prior knowledge and sampling it to obtain the trend term parameters characterizing the global change in performance feedback; calculating the residual value between the performance feedback and the trend term parameters in the historical dataset, constructing a posterior distribution based on the residual value and sampling it to obtain the process disturbance parameters characterizing local fluctuations; and constructing an infinite Gaussian process surrogate model based on the latent surface label parameters, trend term parameters, process disturbance parameters, and preset hyperparameters.

[0037] It should be noted that the trend term parameter is used to characterize the overall trend of the objective function in the decision space, such as a linear function, a polynomial function, or a regression term formed by expanding basis functions. The process disturbance parameter is used to characterize the random fluctuation characteristics of the objective function in a local region after removing the trend term. It is used to describe the local deviations caused by uncertain factors such as random task arrivals, traffic congestion, and power fluctuations. By superimposing the trend term parameter and the process disturbance parameter, the infinite Gaussian process surrogate model (hereinafter referred to as the surrogate model) can simultaneously capture the global structure and local randomness of multiple AGVs.

[0038] Furthermore, the objective function is used to characterize the overall operational efficiency of multiple AGVs under conditions of random task arrival and resource constraints, and its expression is: ; In the formula, W represents minimizing the average workstation waiting time, E represents minimizing energy consumption, and T represents maximizing throughput per unit time.

[0039] In addition, the new hyperparameters include mean coefficient, variance parameter, and kernel function. The mean coefficient is used to determine the weight of the trend term parameter in the surrogate model, reflecting the overall change level of the objective function in the decision space, and is updated using normal distribution sampling. The variance parameter is used to control the magnitude of the process disturbance parameter, characterizing the degree of uncertainty of the objective function in the local region, and is updated using inverse Gamma distribution sampling. The kernel function is used to characterize the correlation range of the objective function in the decision space, controlling the degree of mutual influence between different scheduling decision points, and is updated using Gamma distribution sampling.

[0040] In this optional embodiment, determining the latent function surface label parameters of the scheduling parameter points by calculating the posterior probability values ​​of each latent function surface in the historical dataset and performing random sampling includes: introducing a set of independent auxiliary random variables that follow a Beta distribution; generating weight values ​​for each latent function surface using the auxiliary random variables through a stepwise decomposition method, wherein the weight values ​​of each latent function surface are non-negative and the sum of the weight values ​​is 1; calculating the conditional probability of the scheduling parameter point under each latent surface, and calculating the relative posterior probability value of each latent function surface in combination with the weight of each latent function surface; normalizing the relative posterior probability values ​​so that the sum of the relative posterior probability values ​​of each latent function surface is 1; and determining the latent function surface label parameters of the scheduling parameter points using the Gibbs sampling method based on the normalized posterior probability values.

[0041] Specifically, a set of auxiliary random variables is introduced. Each of them Independently follows a Beta distribution, used to generate the th element through a stepwise decomposition. The weights of each latent function surface are generated and decay naturally with the latent function surface labels. This allows the surrogate model to adaptively determine the actual number of latent function surfaces involved in modeling without pre-setting the number of latent surfaces, while maintaining the stability and computability of the posterior update process.

[0042] Next, weights for the latent function surface are generated through a stepwise decomposition process. These weights measure the relative contribution of the latent surface to the overall objective function, and their calculation expression is as follows: ; ; In the formula, This indicates that the weights of the first latent function surface are directly determined by the corresponding auxiliary random variables. k Represents the latent function surface, Indicates the first k The weights of the surface of a latent function. Represents an auxiliary random variable. ∏ represents the auxiliary random variable corresponding to the past potential surface, and ∏ represents the product calculation.

[0043] when At that time, the first k The weights of each latent function surface are determined by the current auxiliary random variable. Compared to the past The product of the residual weights of the auxiliary random variables determines that the surface weights of each latent function generated in the above manner are all non-negative, and the sum of all latent surface weights is 1, thus satisfying the basic constraint condition of probability weights.

[0044] Specifically, first, the scheduling parameter point is set to x, and the performance feedback obtained after execution is y. Then, weights are applied to the surface of the current latent function. and the corresponding function values ​​on each latent function surface Below, the scheduling decision point belongs to the first... The posterior probability values ​​of a potential function surface satisfy the following relationship: ; In the formula, z The latent function surface label represents the scheduling parameter point. Indicates the first k The function values ​​of a potential function surface. Represents the weight of the potential surface. y Indicates performance feedback, x Indicates the scheduling parameter point. This represents the conditional probability of the performance feedback of the scheduling parameter point under the latent function surface. p Represents probability symbols.

[0045] The conditional probability is used to measure the degree of matching between the scheduling decision point and the latent function surface. It follows a Gaussian distribution, and the mean of the Gaussian distribution is composed of the trend term and the function value of the corresponding latent function surface at the scheduling decision point.

[0046] In this optional embodiment, based on the latent surface label parameters and combined with preset prior knowledge, a posterior distribution is constructed and sampled to obtain trend term parameters characterizing the global change in performance feedback. This includes: classifying and filtering historical datasets based on the latent surface label parameters, assigning labels to scheduling parameter points belonging to the same latent function surface and their corresponding performance feedback values; using preset prior knowledge and combining the relationship between scheduling parameter points and their corresponding performance feedback values, determining the prior mean of the Gaussian process on the latent function surface; constructing a posterior distribution from the prior mean using the performance feedback values ​​and then sampling using the Gibbs sampling method to obtain the trend term parameters.

[0047] It should be noted that prior knowledge, based on manual scheduling experience, operational procedures, and historical operational summaries, includes at least the following forms: First, the path cost weight range that has been proven to cause severe congestion at critical road sections or intersections during long-term AGV operation is assigned a lower prior performance mean. Second, charging triggering threshold ranges that are known to easily lead to concentrated charging of AGVs, frequent task interruptions, or power risks are assigned lower prior performance mean values. Third, scheduling parameter combinations that repeatedly resulted in decreased system throughput or significantly increased workstation waiting time during historical workshop operations were given a punitive prior mean. Fourth, parameter regions that violate implicit engineering constraints or operational logic under specific operating conditions but are difficult to formally model are marked as high-risk regions.

[0048] In this optional embodiment, the residual values ​​between performance feedback and trend term parameters in the historical dataset are calculated. A posterior distribution is constructed based on the residual values ​​and sampled to obtain process disturbance parameters characterizing local fluctuations. Based on the trend term parameters, the trend benchmark values ​​of scheduling parameter points are calculated, and residual calculations are performed on the performance feedback using the trend benchmark values ​​to obtain residual values. The residual values ​​are used as observation samples to construct a conditional posterior distribution and sampled using the Gibbs sampling method to obtain process disturbance parameters.

[0049] S103. Construct a scheduling parameter space, and uniformly mark the scheduling parameter space and / or use the big oracle model to obtain a set of candidate scheduling parameter points; In this optional embodiment, constructing a scheduling parameter space and uniformly marking the scheduling parameter space and / or obtaining a set of candidate scheduling parameter points using a large oracle model includes: constructing a scheduling parameter space based on the scheduling parameter dimensions of AGV operation; standardizing the scheduling parameter space using pre-configured operation constraints and operation experience, and uniformly marking the standardized scheduling parameter space at preset intervals to obtain basic scheduling parameter points; constructing prompt words based on scheduling parameter points and performance feedback from historical data and AGV operation status information, and inputting them into a preset large language model to obtain supplementary scheduling parameter points; merging and removing the basic scheduling parameter points and supplementary scheduling parameter points to obtain a set of candidate scheduling parameter points.

[0050] Specifically, the scheduling parameters, including the task assignment matrix, path cost weight matrix, right-of-way control parameters, and charging strategy parameters, are used as scheduling dimensions to construct a scheduling parameter space.

[0051] The task assignment matrix, a 30×80 binary matrix, determines the task queue order for each AGV. , indicating the first i Taiwan AGV was assigned to perform the first j One task; During execution, a task queue for each AGV is generated based on the task assignment matrix, and the corresponding path planning and execution operations are triggered sequentially. The path cost weight assigns weighting coefficients to the distance, congestion, and energy consumption of different road segments, with values ​​ranging from [0.5, 2.0]. Specifically, when evaluating any road segment, its path cost is calculated by the following weighting factors: Road segment length, historical or real-time congestion levels, and energy consumption per unit distance.

[0052] The candidate paths are compared based on the above weighted factors, and the path with the lowest overall cost is selected as the AGV's travel route.

[0053] Right-of-way control parameters are used to adjust the passage rules of AGVs at intersections, including the allocation of intersection signal time slots (0.5s~5s) and priority rules (priority for critical materials and / or first-come-first-served). During operation, the passage order of AGVs is scheduled according to the right-of-way control parameters, thereby affecting the waiting time at intersections and the overall traffic efficiency.

[0054] The electric strategy parameters are used to trigger the charging threshold (SOC=20%-40%) and the charging station selection rules (shortest waiting time priority or nearest distance priority). The specific charging location is selected from multiple available charging stations, and the AGV is judged to enter the charging state according to the electric strategy parameters, and its task execution and path planning strategies are adjusted accordingly.

[0055] The scheduling parameter space is standardized and marked with points according to the preset marking interval to generate basic scheduling parameter points.

[0056] In addition, a large language model is introduced to recommend candidate scheduling parameters. Specifically, historical scheduling parameter points from the optimization process, corresponding performance feedback data, prediction results from the surrogate model (such as function values ​​on the latent function surface), and structured information on the operating status of multiple AGVs are combined to construct prompt words, which are then input into the large language model. Based on its understanding of AGV scheduling rules, engineering constraints, and historical operating characteristics, the large oracle model outputs a set of recommended supplementary scheduling parameter points, as shown in Table 1. Table 1: List of Prompt Words ; The basic scheduling parameter points and the supplementary scheduling parameter points are merged and removed to obtain a set of candidate scheduling parameter points.

[0057] It should be noted that the large language model adopts an open-source pre-trained large language model, and the candidate scheduling parameter points generated by the large language model can also directly constitute a set of candidate scheduling parameter points, improving the engineering rationality and overall quality of the candidate set.

[0058] In this optional embodiment, constructing the scheduling parameter space based on the scheduling parameter dimensions of AGV operation includes: The dimensions include the task assignment matrix, path cost weight, right-of-way control parameters, and charging strategy parameters.

[0059] S104. Based on the infinite Gaussian process surrogate model and exploration probability, select the target scheduling parameter point from the candidate scheduling parameter point set, and use the target scheduling parameter point as the candidate scheduling scheme.

[0060] In this optional embodiment, selecting a target scheduling parameter point from the candidate scheduling parameter point set based on the infinite Gaussian process surrogate model and exploration probability, and using the target scheduling parameter point as a candidate scheduling scheme includes: inputting the candidate scheduling parameter point set into the infinite Gaussian process surrogate model, randomly sampling a latent function surface from the posterior distribution of the infinite Gaussian process using the Thompson sampling method, and using the surface to calculate the prediction performance value of each candidate scheduling parameter point under each latent function surface; selecting a candidate scheduling parameter point using a preset exploration probability method, and using the scheduling parameter point as the target scheduling parameter point.

[0061] Specifically, the candidate scheduling scheme is simulated to obtain the actual performance feedback corresponding to the scheme. The actual performance feedback corresponding to the candidate scheduling scheme is then added to the historical dataset as a new data pair to obtain the historical dataset of the new AGV.

[0062] For example, the task completion time, path travel time, energy consumption, waiting time at intersections, and task interruption or delay caused by charging are statistically analyzed for each AGV. The above statistical results are then combined and / or weighted summed to obtain the actual performance feedback value of the corresponding scheduling scheme.

[0063] In addition, the mutual influence between multiple AGVs is reflected in the following ways: when multiple AGVs choose the same or adjacent paths at the same time, it will cause local congestion, thereby increasing travel time and energy consumption; when multiple AGVs pass through the same intersection in a concentrated manner, queuing will occur, extending the task completion time; when multiple AGVs enter the same charging station at the same time for charging, it will cause charging waiting or task suspension.

[0064] It should be noted that, in order to achieve a balance between exploration and utilization, candidate scheduling parameter points are determined by exploration probability. Make a selection, through The method for selecting the candidate scheduling parameter point with the highest predicted reward is ( Or directly use the exploration probability. Randomly select a candidate scheduling parameter point.

[0065] S105. Simulate the candidate scheduling scheme to obtain the current historical data, and add the current historical data to the historical dataset to obtain a new historical dataset. Based on the new historical dataset, repeatedly execute the construction of an infinite Gaussian process proxy model, obtain the set of candidate scheduling parameter points and select the target scheduling parameter point until the preset number of iterations is reached, and then output the final scheduling scheme.

[0066] For example, for 30 AGVs, 12 intersections, and 4 charging stations, a scheduling scheme is selected based on the scheduling parameter space. This scheme is then run in a multi-AGV system or simulation system to obtain corresponding real performance feedback, reflecting phenomena such as waiting, congestion, energy consumption, and task interruptions experienced by multiple AGVs under the current scheduling scheme. Next, the agent model is updated using historical scheduling parameters and performance feedback as new data, and a new scheduling scheme is selected and determined using the updated agent model.

[0067] Specifically, using a 60-minute cycle, when historical operating data shows that multiple AGVs frequently converge on the same charging station during certain time periods, leading to a significant increase in charging waiting time, the scheduling parameter combination corresponding to this operating state will be reflected in poor performance feedback in the model. Subsequently, in subsequent iterations, the proxy model will reduce the priority of parameters such as charging trigger thresholds or charging station selection strategies, making the system more inclined to choose strategy parameters that disperse charging demand or delay charging timing, thereby alleviating charging station resource conflicts, as shown in Table 2: Table 2: Comparison Table Before and After Optimization ; Based on the performance indicators in Table 2, the operating performance of AGVs is evaluated using an objective function, thereby characterizing the overall operating efficiency of the multi-AGV system under conditions of random task arrival and resource constraints.

[0068] like Figure 4 As shown, several existing Bayesian optimization-based scheduling methods were selected as benchmarks for comparison, including: scheduling methods based on the Confidence Upper Bound (UCB) policy, scheduling methods based on the Expected Improvement (EI) criterion, scheduling methods based on the Posterior Sampling (TS) criterion, and scheduling methods based on the Knowledge Gradient (KG) criterion. The present invention is ∞-GP-TS (Infinite Gaussian Thompson Sampling). In the figure, the horizontal axis represents the number of optimization iterations, and the vertical axis represents the cumulative reward obtained after the execution of the scheduling scheme selected by the method.

[0069] The cumulative reward is calculated based on workstation waiting time, throughput, and energy consumption indicators according to preset weights. The calculation process is as follows: During the execution of each scheduling plan, various basic operational data are recorded in real time, including: 1) Material arrival time and waiting time at each workstation; 2) The number of delivery tasks completed by each AGV and the completion time; 3) The travel distance, travel time, and energy consumption of each AGV during the task execution process; 4) Waiting and stopping status of AGVs at intersections or charging stations.

[0070] Next, based on the above operational data, calculate the key performance indicators (objective function expression above) within a single evaluation cycle (60 minutes).

[0071] For the workstation waiting time indicator, the waiting time of each workstation due to waiting for AGV delivery is counted within the evaluation period, and the average waiting time of all workstations is taken to obtain the average workstation waiting time within the period. For the throughput metric, the total number of delivery tasks completed by the multi-AGV system within the evaluation period is counted and divided by the evaluation period length to obtain the system throughput per unit time. For energy consumption indicators, the total energy consumption of all AGVs during the evaluation period is calculated, including driving energy consumption, start-stop energy consumption, and additional energy consumption caused by congestion or queuing, and normalized as needed.

[0072] Subsequently, the aforementioned performance indicators are converted into evaluation quantities with a unified direction. Since the workstation waiting time and system energy consumption are better the smaller they are, while the system throughput is better the larger they are, different indicators are processed in a unified direction when calculating the cumulative reward, so that they can be weighted and combined in the same objective function. Then, according to the preset weight coefficients, the workstation waiting time, throughput and energy consumption indicators are weighted and summed to obtain the reward value of the corresponding scheduling scheme within the evaluation period.

[0073] The weighting coefficients are determined by the production management requirements in the application scenario and are used to reflect the relative importance of different performance indicators in actual production. In production lines with high cycle time requirements, the weight of throughput indicators can be increased; in energy cost-sensitive scenarios, the weight of energy consumption indicators can be increased accordingly.

[0074] Finally, the reward values ​​obtained over multiple consecutive evaluation periods are summed to form a cumulative reward, which is used to measure the long-term operational effectiveness of the scheduling scheme throughout the optimization process.

[0075] For example, using the same optimization budget (100 times) in multiple scenarios (A, B, C), the optimization effect of the present invention is shown in Table 3: Table 3: Comparison of Optimization Across Multiple Scenarios ; Scenario A: 30 AGVs, 80 workstations, 12 intersections, and 4 charging stations; Scenario B: 60 AGVs, 160 workstations, 24 intersections, and 8 charging stations; Scenario C: 100 AGVs, 240 workstations, 36 intersections, and 12 charging stations.

[0076] Figure 2 An embodiment of a multi-AGV scheduling optimization system based on nonparametric Bayesian methods according to the present invention is shown.

[0077] In this optional embodiment, a multi-AGV scheduling optimization system based on nonparametric Bayes includes: The data acquisition module 201 is used to acquire the historical dataset of the AGV and set the number of iterations and the exploration probability. The historical dataset includes scheduling parameter points and corresponding performance feedback. The model building module 202 is used to extract model parameters from the posterior distribution of the infinite Gaussian process based on the historical dataset of AGV, combined with the sampling method, and to build an infinite Gaussian process proxy model using the model parameters. The candidate set module 203 is used to construct the scheduling parameter space, uniformly mark the scheduling parameter space and / or use the big oracle model to obtain the candidate scheduling parameter point set; The candidate scheme module 204 is used to select the target scheduling parameter point from the candidate scheduling parameter point set based on the infinite Gaussian process proxy model and exploration probability, and use the target scheduling parameter point as a candidate scheduling scheme. The feedback iteration module 205 is used to simulate the candidate scheduling scheme, obtain the current historical data, and add the current historical data to the historical dataset to obtain a new historical dataset. Based on the new historical dataset, the construction of the infinite Gaussian process proxy model, the acquisition of the candidate scheduling parameter point set, and the selection of the target scheduling parameter point are repeatedly executed until the preset number of iterations is reached, and then the final scheduling scheme is output.

[0078] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0079] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0080] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0081] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0083] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A multi-AGV scheduling optimization method based on nonparametric Bayesian methods, characterized in that, include: Obtain the historical dataset of the AGV and set the number of iterations and the exploration probability. The historical dataset includes scheduling parameter points and corresponding performance feedback. Based on the historical dataset of AGVs, model parameters are extracted from the posterior distribution of the infinite Gaussian process using a sampling method, and the model parameters are used to construct a proxy model of the infinite Gaussian process. Construct a scheduling parameter space, and uniformly mark the scheduling parameter space and / or use a large oracle model to obtain a set of candidate scheduling parameter points; Based on the infinite Gaussian process surrogate model and exploration probability, target scheduling parameter points are selected from the set of candidate scheduling parameter points, and these target scheduling parameter points are used as candidate scheduling schemes. The candidate scheduling schemes are simulated to obtain the current historical data, which is then added to the historical dataset to obtain a new historical dataset. Based on the new historical dataset, the process of constructing an infinite Gaussian process proxy model, obtaining a set of candidate scheduling parameter points, and selecting target scheduling parameter points is repeated until the preset number of iterations is reached. Finally, the final scheduling scheme is output.

2. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 1, characterized in that, The historical dataset based on AGVs, combined with a sampling method to extract model parameters from the posterior distribution of the infinite Gaussian process, and the construction of an infinite Gaussian process surrogate model using these model parameters, includes: Calculate the posterior probability of each latent function surface of the scheduling parameter points in the historical dataset, and sample to determine the label parameters of the latent function surface of the scheduling parameter points; Based on the potential surface label parameters and combined with the preset prior knowledge, a posterior distribution is constructed and sampled to obtain the trend term parameters that characterize the global changes in performance feedback. Calculate the residual values ​​between performance feedback and trend term parameters in the historical dataset, construct the posterior distribution based on the residual values ​​and sample them to obtain the process perturbation parameters that characterize local fluctuations; An infinite Gaussian process surrogate model is constructed based on potential surface label parameters, trend term parameters, process perturbation parameters, and preset hyperparameters.

3. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 2, characterized in that, The calculation of the posterior probability values ​​of each latent function surface of the scheduling parameter points in the historical dataset, and the random sampling to determine the latent function surface label parameters of the scheduling parameter points, includes: A set of independent auxiliary random variables following a Beta distribution is introduced. The weight values ​​of each latent function surface are generated by the auxiliary random variables through a stepwise decomposition method. The weight values ​​of each latent function surface are non-negative and the sum of the weight values ​​is 1. Calculate the conditional probability of the scheduling parameter point under each potential surface, and combine the weights of each potential function surface to calculate the relative posterior probability value of each potential function surface; The relative posterior probability values ​​are normalized so that the sum of the relative posterior probability values ​​of each latent function surface is 1. Based on the normalized posterior probability values, the latent function surface label parameters of the scheduling parameter points are determined using the Gibbs sampling method.

4. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 2, characterized in that, The process of constructing a posterior distribution based on potential surface label parameters and combining it with pre-defined prior knowledge, and then sampling to obtain trend term parameters characterizing the global changes in performance feedback, includes: Based on the latent surface label parameters, the historical dataset is classified and filtered to assign the labels to the scheduling parameter points of the same latent function surface and their corresponding performance feedback values. By utilizing pre-defined prior knowledge and combining the relationship between scheduling parameter points and their corresponding performance feedback values, the prior mean of the Gaussian process on the latent function surface is determined. The posterior distribution is constructed from the prior mean using the performance feedback value, and then sampled using the Gibbs sampling method to obtain the trend term parameters.

5. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 2, characterized in that, The residual values ​​between performance feedback and trend term parameters in the historical dataset are calculated. Based on the residual values, a posterior distribution is constructed and sampled to obtain process perturbation parameters characterizing local fluctuations. Based on the trend term parameter, the trend baseline value of the scheduling parameter point is calculated, and the residual value is used to calculate the performance feedback to obtain the residual value. Using the residuals as observation samples, a conditional posterior distribution is constructed and sampled using the Gibbs sampling method to obtain the process perturbation parameters.

6. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 1, characterized in that, The construction of the scheduling parameter space, including uniformly marking the scheduling parameter space and / or using a large oracle model to obtain a set of candidate scheduling parameter points, includes: Construct a scheduling parameter space based on the dimensions of AGV operation scheduling parameters; The scheduling parameter space is standardized using pre-configured operational constraints and operational experience, and the standardized scheduling parameter space is evenly marked with points at preset intervals to obtain basic scheduling parameter points. Based on the scheduling parameter points, performance feedback, and AGV running status information in historical data, prompt words are constructed and input into the preset large language model to obtain supplementary scheduling parameter points; The basic scheduling parameter points and the supplementary scheduling parameter points are merged and removed to obtain a set of candidate scheduling parameter points.

7. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 1, characterized in that, The scheduling parameter space constructed based on the scheduling parameter dimensions of AGV operation includes: The dimensions include the task assignment matrix, path cost weight, right-of-way control parameters, and charging strategy parameters.

8. The multi-AGV scheduling optimization method based on nonparametric Bayes as described in claim 1, characterized in that, The step of selecting the target scheduling parameter point from the candidate scheduling parameter point set based on the infinite Gaussian process surrogate model and exploration probability, and using the target scheduling parameter point as a candidate scheduling scheme, includes: The candidate scheduling parameter point set is input into the infinite Gaussian process surrogate model. The Thompson sampling method is used to randomly sample a latent function surface from the posterior distribution of the infinite Gaussian process. Using this surface, the prediction performance value of each candidate scheduling parameter point under each latent function surface is calculated. Candidate scheduling parameter points are selected using a preset exploration probability method, and these scheduling parameter points are used as target scheduling parameter points.

9. A multi-AGV scheduling optimization method based on nonparametric Bayesian methods according to claim 3, characterized in that, The expression for the posterior probability value is: ； In the formula, z The latent function surface label represents the scheduling parameter point. Indicates the first k The function values ​​of a potential function surface. Represents the weight of the potential surface. y Indicates performance feedback, x Indicates the scheduling parameter point. This represents the conditional probability of the performance feedback of the scheduling parameter point under the latent function surface. p Represents probability symbols.

10. A multi-AGV scheduling optimization system based on nonparametric Bayesian methods, characterized in that, include: The data acquisition module is used to acquire the historical dataset of the AGV and set the number of iterations and the exploration probability. The historical dataset includes scheduling parameter points and corresponding performance feedback. The model building module is used to extract model parameters from the posterior distribution of the infinite Gaussian process based on the historical dataset of AGV, combined with the sampling method, and to build an surrogate model of the infinite Gaussian process using the model parameters. The candidate set module is used to construct the scheduling parameter space, uniformly mark the scheduling parameter space and / or use the big oracle model to obtain the candidate scheduling parameter point set; The candidate scheme module is used to select the target scheduling parameter point from the set of candidate scheduling parameter points based on the infinite Gaussian process surrogate model and exploration probability, and use the target scheduling parameter point as a candidate scheduling scheme. The feedback iteration module is used to simulate the candidate scheduling schemes, obtain the current historical data, and add the current historical data to the historical dataset to obtain a new historical dataset. Based on the new historical dataset, the steps are repeated until the preset number of iterations is reached, and then the final scheduling scheme is output.

Images