Method, device and equipment for determining traffic control candidate scheme and storage medium

Abstract

The application discloses a kind of methods for determining traffic control candidate scheme, device, equipment and storage medium, it is related to intelligent transportation and microcosmic traffic simulation technical field, comprising: inputting road network file, traffic file and additional control file into traffic simulator to simulate initial scheme, obtain simulation result;Extract multi-dimensional evaluation index in simulation result, with the index of reference scheme as benchmark to normalize multi-dimensional evaluation index, obtain target index;According to preset optimization mode and target index, the initial scheme is optimized to obtain intermediate scheme;Extract the pareto front of intermediate scheme;When the number of intermediate scheme corresponding to the pareto front exceeds threshold value, the intermediate scheme is screened according to the congestion degree of target index, to obtain traffic control candidate scheme.The problem that traffic control scheme is formulated to rely on engineering experience and a small amount of enumeration trial and error, cannot converge to pareto front, and the traffic control scheme formulated has multi-objective conflict problem.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and micro-traffic simulation technology, specifically to methods, apparatus, equipment, and storage media for determining candidate traffic control schemes. Background Technology

[0002] The development of traffic control schemes has long relied on engineering experience and limited trial-and-error enumeration. However, the scheme space grows exponentially with the combination of multiple variables such as speed limits, accident locations, lane occupancy, start and end times, and traffic density levels, making manual enumeration or linear scanning impractical. Developed traffic control schemes suffer from conflicting objectives; average travel time, queue length, capacity utilization, and fairness are mutually constrained, and any single-objective optimal solution inevitably compromises other aspects in engineering practice. Using a pure enumeration platform to develop traffic control schemes only performs batch parameter scanning, with parameters still manually selected, making convergence to the Pareto front impossible.

[0003] Therefore, existing technologies suffer from problems such as relying on engineering experience and a small amount of enumeration and trial and error to formulate traffic control schemes, failing to converge to the Pareto front, and having multiple conflicting objectives in the formulated traffic control schemes. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for determining candidate traffic control schemes, in order to solve the problems that the formulation of traffic control schemes relies on engineering experience and a small amount of enumeration and trial and error, cannot converge to the Pareto front, and has multiple conflicting objectives.

[0005] In a first aspect, the present invention provides a method for determining candidate traffic control schemes, the method comprising: Input the road network file, traffic flow file, and additional control file rendered from the initial plan into the traffic simulator, and use the traffic simulator to simulate the initial plan to obtain the simulation results; Multi-dimensional evaluation indicators are extracted from the simulation results. Based on the indicators of the reference scheme, the multi-dimensional evaluation indicators are normalized to obtain the target indicators. The target indicators are used to compare and merge the initial scheme across indicators. The initial solution is optimized based on the preset optimization mode and target indicators to obtain the intermediate solution; Perform non-dominated sorting on the intermediate schemes and extract the Pareto front corresponding to the intermediate schemes based on the sorting results; When the number of intermediate solutions corresponding to the Pareto front exceeds the first preset threshold, the intermediate solutions are screened according to the congestion level corresponding to the target indicator to obtain traffic control candidate solutions.

[0006] The method for determining candidate traffic control schemes provided in this embodiment uses a traffic simulator as the truth-driven kernel, combining a multi-objective optimization algorithm and a probabilistic surrogate model to automatically search, evaluate performance, and extract Pareto fronts for the scheme space of various traffic control actions, and then injects the optimal candidate set back into the strategy generation module. This method uses simulation results as the truth-driven approach, combined with preset optimization modes, to automatically optimize the scheme space for actions such as speed limits, accidents, and lane control. It approximates the multi-objective Pareto front with a small simulation budget, outputting a structured and interpretable set of candidate schemes and their actual performance indicators under simulation, significantly reducing the cost of manual trial and error. It solves the problems of traffic control scheme formulation relying on engineering experience and limited enumeration trial and error, failing to converge to the Pareto front, and exhibiting multi-objective conflicts.

[0007] In one alternative implementation, after obtaining candidate traffic control solutions, the method further includes: Based on the preset information of traffic control candidate solutions, a structured summary of the traffic control candidate solutions is generated, and the structured summary is written into the context anchor of the policy generator as a priority source for downstream recommendations. Display constraints in the policy generator and obtain the confidence level output by the policy generator based on the confidence level labels of the traffic control candidate solutions; Use structured logs to record the default behavior of the policy generator.

[0008] In this implementation, the structured summary anchoring strategy generator forces priority fine-tuning of Pareto candidates and links confidence propagation, which can effectively suppress random divergence of large language models and improve the feasibility of recommendations and the interpretability of solutions. Log recording supports offline evaluation, realizes closed-loop feedback, and enhances the credibility of system engineering and iterative optimization capabilities.

[0009] In an alternative implementation, the method further includes, before inputting the road network file, traffic flow file, and additional control file rendered from the initial scheme into the traffic simulator: Determine the control plan and represent it as a structured object, where the structured object is used to represent the initial plan; Define validity constraints for each decision variable in the initial scheme; The continuous variables in the initial scheme are discretized into candidate levels, and sampling is performed based on the discrete variables to obtain the initial scheme.

[0010] In this implementation, illegal schemes are eliminated in advance through structured representation and hard constraint pre-filtering, avoiding unnecessary simulation overhead; continuous variables are discretized according to physical meaning, effectively compressing the combination space, improving search efficiency while maintaining coverage, and providing a feasible and compact initial scheme space for subsequent multi-objective optimization.

[0011] In one optional implementation, a traffic simulator is used to simulate the initial plan, and the simulation results are obtained, including: The initial scheme was repeatedly run using a traffic simulator under multiple different random seeds to obtain the mean and variance of the index. The simulation results are obtained based on the index mean and index variance; If a crash or illegal exit code occurs during the simulation, determine the cause of the failure and mark the initial solution as infeasible.

[0012] In this implementation, the mean and variance of the index are obtained by repeated simulation with multiple random seeds, which effectively suppresses the ranking reversal caused by random fluctuations and improves the robustness of the evaluation. Collapsed or illegal exit schemes are marked as infeasible and excluded to ensure that subsequent optimization is based only on reliable data and to avoid abnormal results interfering with the Pareto search.

[0013] In one optional implementation, the initial solution is optimized according to a preset optimization mode and target indicators to obtain an intermediate solution, including: With the preset optimization mode being Bayesian optimization mode, the evaluation point is determined according to the Gaussian process surrogate model corresponding to the target index under the Pareto hypervolume improvement meaning. The expected future hypervolume improvement is predicted based on the evaluation point, the mean of the target index, and the variance of the target index, and the prediction result is obtained. The prediction result is used to optimize the initial scheme to obtain the intermediate scheme. Determine the deviation between the prediction result and the simulation result. If the deviation is greater than the second preset threshold, trigger agent retraining and mark the initial scheme.

[0014] In this implementation, the deviation between the prediction results and the simulation results is monitored. When the deviation exceeds the threshold, retraining is triggered and abnormal samples are marked, which effectively prevents abnormal results from misleading the search and improves the efficiency and reliability of multi-objective optimization.

[0015] In one optional implementation, the initial solution is optimized according to a preset optimization mode and target indicators to obtain an intermediate solution, including: When the preset optimization mode is the evolutionary multi-objective mode, an evolutionary algorithm based on non-dominated sorting and crowding is used to perform a global Pareto search on the initial scheme, and the initial scheme is optimized according to the search results and soft constraint repair operators to obtain an intermediate scheme. When the preset optimization mode is space-filling sampling mode, determine whether the initial solution violates the engineering hard constraints or the engineering soft constraints. If the initial solution violates the engineering hard constraints, mark the initial solution. If the initial solution violates the soft constraints, the initial solution is optimized by the soft constraint repair operator according to the preset arrangement order to obtain the intermediate solution.

[0016] In this implementation, a global Pareto search is achieved, and the soft constraint repair operator adjusts the offspring according to the risk order to ensure the feasibility of the scheme. The space-filling sampling mode generates a cold-start seed through hard constraint filtering and soft constraint repair, providing a high-quality initial population for Bayesian or evolutionary modes. Multiple modes share a unified output interface, enhancing the framework's adaptability and robustness.

[0017] In one alternative implementation, after obtaining candidate traffic control solutions, the method further includes: Obtain the action list, target indicator value, percentage improvement relative to the reference solution, and confidence label corresponding to the traffic control candidate solution, and generate summary information of the traffic control candidate solution; Based on the action list, target indicator values, improvement percentages, confidence labels, and summary information, generate structured descriptions of traffic control candidate solutions.

[0018] Secondly, the present invention provides an apparatus for determining candidate traffic control schemes, the apparatus comprising: The simulation module is used to input road network files, traffic flow files, and additional control files rendered from the initial plan into the traffic simulator, and use the traffic simulator to simulate the initial plan to obtain simulation results. The index processing module is used to extract multi-dimensional evaluation indicators from the simulation results. Using the indicators of the reference scheme as a benchmark, the multi-dimensional evaluation indicators are normalized to obtain the target indicators. The target indicators are used to compare and merge the initial schemes across indicators. The scheme optimization module is used to optimize the initial scheme according to the preset optimization mode and target indicators to obtain the intermediate scheme; The extraction module is used to perform non-dominated sorting on intermediate solutions and extract the Pareto front corresponding to the intermediate solutions based on the sorting results. The filtering module is used to filter intermediate solutions based on the congestion level corresponding to the target indicator when the number of intermediate solutions corresponding to the Pareto front exceeds a first preset threshold, thereby obtaining traffic control candidate solutions.

[0019] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method for determining traffic control candidate schemes described in the first aspect or any corresponding embodiment thereof.

[0020] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method for determining traffic control candidate schemes described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a method for determining candidate traffic control schemes according to an embodiment of the present invention. Figure 2 This is a structural block diagram of a device for determining candidate traffic control schemes according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0025] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0026] The formulation of traffic control schemes has long relied on engineering experience and limited enumeration and trial-and-error, resulting in problems such as scheme space explosion, multi-objective conflict, high simulation costs and noisy results, and lack of transferability across scenarios. Specifically, scheme space explosion means that the scheme space grows exponentially after combining multiple variables such as speed limits, accident locations, occupied lanes, start and end times, and traffic density levels, making manual enumeration or linear scanning infeasible. Multi-objective conflict means that indicators such as average travel time, queue length, capacity utilization, and fairness are mutually constrained, and any single-objective optimal solution will inevitably compromise other aspects in engineering. High simulation costs and noisy results mean that a single simulation requires a minute-level central processing unit, and the results fluctuate significantly with different random seeds, requiring averaging of multiple seeds for comparison, making full-space evaluation impractical in engineering. Lack of transferability across scenarios means that the same scheme performs drastically under different traffic density levels and mileage segments, and the optimal solutions for accidents, speed limits, and lane closures are different, lacking a unified and reusable optimization framework.

[0027] Furthermore, when formulating traffic control plans, if a pure enumeration platform is used, the resulting plans cannot converge to the Pareto front because the platform only performs batch parameter scanning and still relies on manual selection. If a deep learning-based surrogate model is used, although it can predict indicators, it lacks interpretability of the mapping between control and indicators and is highly dependent on the sample size of the target domain, making it less practical in small-sample scenarios. If reinforcement learning methods are used, they require a differentiable simulation environment and long-term reward signals, resulting in high engineering costs and difficulties in implementation compared to existing discrete event simulators. If a one-time suggestion based on a large language model is used, although it can efficiently generate traffic control plans, this model tends to provide empirical parameters and lacks a mechanism for global optimization within the plan space, making it prone to getting trapped in local solutions.

[0028] Based on the above, this invention provides a method for determining candidate traffic control schemes. Using a traffic simulator as the truth-driven kernel, and combining a multi-objective optimization algorithm and a probabilistic surrogate model, it automatically searches, evaluates performance, and extracts the Pareto front for the scheme space of various traffic control actions. The optimal candidate set is then back-injected into the strategy generation module. This method uses simulation results as the truth-driven kernel and combines a multi-objective optimization algorithm (Bayesian optimization and evolutionary algorithms are optional) to automatically optimize the scheme space for actions such as speed limits, accidents, and lane control. It approximates the multi-objective Pareto front with a small simulation budget, outputting a structured and interpretable candidate scheme set and its true performance under simulation. The candidate set is then back-injected into the upstream strategy generator to improve the quality of downstream decision-making and significantly reduce the cost of manual trial and error.

[0029] According to an embodiment of the present invention, a method embodiment for determining candidate traffic control schemes is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0030] This embodiment provides a method for determining candidate traffic control schemes, which can be used in the aforementioned computer system. Figure 1 This is a flowchart of a method for determining candidate traffic control schemes according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Input the road network file, traffic flow file, and additional control file rendered from the initial scheme into the traffic simulator, and use the traffic simulator to simulate the initial scheme to obtain simulation results.

[0031] Specifically, a traffic simulator can be connected, such as micro-traffic flow simulation software like SUMO or VISSIM. The traffic simulator is started via an external process call, and the road network file, the traffic flow file generated by the traffic flow density level, and the additional control file rendered by the initial traffic control plan (hereinafter referred to as the initial plan) are input into the traffic simulator.

[0032] The initial scheme is simulated using a traffic simulator to obtain simulation results. For example, each initial scheme is repeatedly run under multiple different random seeds, and the mean and variance of the indicators are taken as the final evaluation value, i.e., the simulation result, to reduce the ranking reversal caused by the randomness of the simulation. Simulation results, such as detailed vehicle trajectory data, will be extracted into macroscopic multi-dimensional evaluation indicators.

[0033] Step S102: Extract multi-dimensional evaluation indicators from the simulation results. Using the indicators of the reference scheme as a benchmark, normalize the multi-dimensional evaluation indicators to obtain the target indicators. The target indicators are used to compare and merge the initial scheme across indicators.

[0034] Specifically, multi-dimensional evaluation indicators are extracted from the simulation results, such as: average travel time in key sections, queue length quantiles, capacity utilization, imbalance index between upstream and downstream road segments, and safe alternative indicators.

[0035] The reference scheme refers to the specific indicators in the simulation of an uncontrolled situation in the accident occurrence environment. Using the indicators of the reference scheme as a benchmark, the multi-dimensional evaluation indicators are normalized to obtain a dimensionless relative improvement ratio, which serves as the target indicator. The target indicator facilitates cross-indicator comparison and merging of the initial schemes.

[0036] Furthermore, this embodiment also supports extracting regional sub-indicators by mileage segment and lane granularity, making candidate solutions interpretable at the regional level. These regional sub-indicators are not directly used at the optimization level but can be used as a visualization interface to enhance interpretability.

[0037] Step S103: Optimize the initial solution according to the preset optimization mode and target indicators to obtain the intermediate solution.

[0038] Specifically, this embodiment includes a multi-objective optimization engine module, which can provide various preset optimization modes, such as Bayesian optimization mode, evolutionary multi-objective mode, and space-filling sampling mode. These three modes can be dynamically switched in engineering practice based on the sample budget and the number of objectives.

[0039] The target index is input into the multi-objective optimization engine module, which determines a suitable preset optimization mode, optimizes the initial solution, and obtains an intermediate solution.

[0040] Step S104: Perform non-dominated sorting on the intermediate schemes and extract the Pareto front corresponding to the intermediate schemes based on the sorting results.

[0041] Specifically, the intermediate schemes are the schemes that complete the simulation. All intermediate schemes are sorted in a multi-objective non-dominated order. The Pareto front corresponding to the intermediate schemes is extracted based on the sorting results.

[0042] Perform a multi-objective non-dominated sort on all intermediate solutions. For example, the sorting can use NSGA-II fast non-dominated sort, which involves three steps: 1. Calculate the dominance relationship: For all intermediate solutions P... success Compare the strengths and weaknesses of each element pairwise across the five objective dimensions. X is said to dominate y if and only if x is not inferior to y in all five objective dimensions, and is strictly superior to y in at least one objective. 2. Stratification: P... success The data is categorized into layers F0, F1, F2, etc., based on the number of solutions it can be dominated by. F0 represents the set of solutions that cannot be dominated by any other solution, i.e., the Pareto front. 3. Incremental Update: Each time a new solution is simulated, only an incremental comparison is performed on F0, at a cost of O(m·|F0|+|new|), where m represents the number of target dimensions and "|new|" represents the number of newly added solutions, completed in milliseconds. Extraction means that members in F0 that are no longer dominated by any solution are filtered out again based on "information redundancy".

[0043] Step S105: When the number of intermediate solutions corresponding to the Pareto front exceeds the first preset threshold, the intermediate solutions are screened according to the congestion level corresponding to the target indicator to obtain traffic control candidate solutions.

[0044] Specifically, the first preset threshold is, for example, 4, 5, or other values. When the number of intermediate solutions corresponding to the Pareto front exceeds the first preset threshold, it is determined that the number of Pareto candidates exceeds the first preset threshold. In this case, the intermediate solutions need to be simplified. The intermediate solutions are screened according to the congestion in different directions of each target indicator. Redundant solutions are eliminated and representative candidate solutions are retained as traffic control candidate solutions.

[0045] The method for determining candidate traffic control schemes provided in this embodiment inputs road network files, traffic flow files, and additional control files rendered from the initial scheme into a traffic simulator. The traffic simulator then simulates the initial scheme to obtain simulation results. Multi-dimensional evaluation indicators are extracted from the simulation results. Using the indicators of a reference scheme as a benchmark, the multi-dimensional evaluation indicators are normalized to obtain target indicators. These target indicators are used for cross-indicator comparison and merging of the initial scheme. The initial scheme is optimized according to a preset optimization mode and the target indicators to obtain intermediate schemes. The intermediate schemes are then non-dominatedly ranked, and the Pareto fronts corresponding to the intermediate schemes are extracted based on the ranking results. When the number of intermediate schemes corresponding to Pareto fronts exceeds a first preset threshold, the intermediate schemes are filtered based on the congestion level corresponding to the target indicators to obtain candidate traffic control schemes. This method uses the traffic simulator as the truth-driven kernel, combining multi-objective optimization algorithms and probabilistic surrogate models to automatically search, evaluate performance, and extract Pareto fronts for the scheme space of various traffic control actions. The optimal candidate set is then reinjected into the strategy generation module. This method uses simulation results as the truth-driven approach, combined with preset optimization modes, to automatically optimize the solution space for actions such as speed limits, accidents, and lane control. It approximates the Pareto front for multiple objectives with a small simulation budget, outputting a structured and interpretable set of candidate solutions and their actual performance metrics under simulation, significantly reducing the cost of manual trial and error. It solves the problems of traffic control schemes relying on engineering experience and limited enumeration trial and error, failing to converge to the Pareto front, and exhibiting multi-objective conflicts.

[0046] As an optional embodiment, after obtaining candidate traffic control solutions, the method further includes: Based on the preset information of traffic control candidate solutions, a structured summary of the traffic control candidate solutions is generated, and the structured summary is written into the context anchor of the policy generator as a priority source for downstream recommendations. Display constraints in the policy generator and obtain the confidence level output by the policy generator based on the confidence level labels of the traffic control candidate solutions; Use structured logs to record the default behavior of the policy generator.

[0047] Specifically, the preset information for traffic control candidate solutions is determined, such as key information like speed limits, speed-limited lanes, and speed limit types. Based on this preset information, a structured summary of the traffic control candidate solutions is generated. This structured summary is then written into the context anchor of the policy generator, i.e., the structured summary of the Pareto candidate set is written into the context anchor of the policy generator, serving as a priority source for downstream recommendations.

[0048] Create a guidance mechanism that displays constraints in the strategy generator. For example, it requires prioritizing the selection of the solution closest to the user's goal from traffic control candidate solutions and fine-tuning it, rather than designing parameters out of thin air. In this case, the constraints are the prompt word layer constraints of the large language model integrated into the strategy generator.

[0049] The process of selecting a solution from traffic control candidate solutions is as follows: Step 1, the user inputs natural language requirements, which are parsed by the LLM (Large Language Model) into a 5-dimensional target vector g. target g target = (ΔT1, ΔT2, ΔT3, ΔT4, ΔT5), where each component is the expected percentage improvement of the objective relative to the reference scheme (negative for travel / queueing, positive for utilization / imbalance / safety). Step 2, in the representative candidate set S repr The above calculates each candidate x∈S repr target distance As shown in formula (1).

[0050] (1) in, , These are the importance weights of each objective, inferred by the LLM from the user's language (default). =1).

[0051] Step 3, press Sort in ascending order and select the top N. anchor =5 traffic control candidate solutions have been entered into the fine-tuning pool (M pool Step 4: Put M pool Each traffic control candidate solution is injected into the [candidate solution anchor point] field of the LLM context, requiring the LLM to be in the M... pool Adjustments are made within a defined scope, rather than freely combining new actions.

[0052] The confidence level of the strategy generator is obtained by using the confidence level labels of the traffic control candidate solutions, thereby enabling the linkage between the confidence level labels of the traffic control candidate solutions and the confidence level output by the strategy generator, and realizing the cascading transmission of uncertainty.

[0053] Table 1 explains the linkage between the confidence labels of traffic control candidate schemes and the confidence output of the policy generator: 1. When injecting traffic control candidate schemes, `conf` is injected simultaneously. `conf` represents the confidence label, not the confidence level. In the "candidate scheme anchor point" field prompted by the LLM, each traffic control candidate scheme carries a confidence label. The LLM knows the weight based on `conf`, and records this weight as `LLM_conf`; 2. Interception via schema verification during generation: When the wrapper layer performs schema verification on the LLM output, if `LLM_conf > rank_conf(conf(x) - 1)`, it is directly rejected; schema verification represents pattern verification. 3. Degradation chain: When S... repr All traffic control candidate solutions are conf=low or S repr If empty, LLM does not output based on traffic control candidate solutions, but instead falls back to the rule template. LLM_conf is forced to be low and "This suggestion is generated by the rule template" is displayed on the front end.

[0054] Table 1. Confidence labels of traffic control candidate solutions and confidence scores of policy generator output.

[0055] Preset behaviors include, for example, the traffic control candidate schemes actually adopted by the policy generator and whether they have undergone manual fine-tuning. These preset behaviors of the policy generator are recorded using structured logs for offline evaluation of the improvement in policy generation quality caused by candidate set injection.

[0056] In this implementation, the structured summary anchoring strategy generator forces priority fine-tuning of Pareto candidates and links confidence propagation, which can effectively suppress random divergence of large language models and improve the feasibility of recommendations and the interpretability of solutions. Log recording supports offline evaluation, realizes closed-loop feedback, and enhances the credibility of system engineering and iterative optimization capabilities.

[0057] As an optional embodiment, the method further includes, before inputting the road network file, traffic flow file, and additional control file rendered from the initial scheme into the traffic simulator: Determine the control plan and represent it as a structured object, where the structured object is used to represent the initial plan; Define validity constraints for each decision variable in the initial scheme; The continuous variables in the initial scheme are discretized into candidate levels, and sampling is performed based on the discrete variables to obtain the initial scheme.

[0058] Specifically, this embodiment sets up a scheme space and decision variable modeling module, which is used to determine the control scheme. A single control scheme is represented as a structured object, such as: speed limit action arrays, accident action arrays, lane closure action arrays, and traffic flow density levels and mileage ranges. The structured object is used to represent the initial scheme.

[0059] The scheme space and decision variable modeling module performs hard constraint pre-filtering, defining legality constraints for each decision variable in the initial scheme, such as: the time interval is non-negative and strictly increasing, the spatial interval is non-empty, the speed limit value is in a legal interval, the lane index is in a legal range, and the direction is legal. If the initial scheme does not meet the above legality constraints, the initial scheme is directly rejected and does not enter the simulation stage.

[0060] The scheme space and decision variable modeling module includes a scheme encoder, which is used to discretize continuous variables into candidate levels according to their physical meaning. For example, a continuous road network will be discretized into a set of edges and junctions in the micro simulation. When the simulation results are output, they will be based on the key section information and the statistical information of the edges to discretize the road network.

[0061] Discrete variables are sampled as sets, compressing the original combination space into a finite but highly comprehensive initial scheme. For example, speed limits are discretized as a percentage of the design speed, instead of directly dividing km / h into equal intervals, the road segment design speed is used as a benchmark, and discretized into 5 candidate levels by percentage; start and end times are discretized according to the minimum control duration, and continuous variables in the time dimension are discretized according to "minimum control duration Δt". min "Slice the data, Δt" min The candidate duration is further constrained to Δt, taking 300 seconds (5 minutes). min Integer multiples of.

[0062] In this implementation, illegal schemes are eliminated in advance through structured representation and hard constraint pre-filtering, avoiding unnecessary simulation overhead; continuous variables are discretized according to physical meaning, effectively compressing the combination space, improving search efficiency while maintaining coverage, and providing a feasible and compact initial scheme space for subsequent multi-objective optimization.

[0063] As an optional embodiment, a traffic simulator is used to simulate the initial plan, and the simulation results are obtained, including: The initial scheme was repeatedly run using a traffic simulator under multiple different random seeds to obtain the mean and variance of the index. The simulation results are obtained based on the index mean and index variance; If a crash or illegal exit code occurs during the simulation, determine the cause of the failure and mark the initial solution as infeasible.

[0064] Specifically, the road network file, the traffic flow file generated from the traffic flow density level, and the additional control file rendered by the current scheme are used as inputs to the traffic simulator.

[0065] The initial scheme is repeatedly run under multiple different random seeds using a traffic simulator. For example, each initial scheme is repeatedly run under multiple different random seeds, and the mean and variance of the index are taken as the final evaluation value to reduce the order reversal caused by the randomness of the simulation. After the simulation, the mean and variance of the index of each initial scheme are obtained.

[0066] The simulation results are obtained based on the index mean and index variance, and these simulation results serve as the optimization objective for Bayesian optimization.

[0067] If a crash or illegal exit code occurs during the simulation, identify and record the reason for the failure, mark the initial solution as infeasible, and do not participate in the subsequent sorting.

[0068] In this implementation, the mean and variance of the index are obtained by repeated simulation with multiple random seeds, which effectively suppresses the ranking reversal caused by random fluctuations and improves the robustness of the evaluation. Collapsed or illegal exit schemes are marked as infeasible and excluded to ensure that subsequent optimization is based only on reliable data and to avoid abnormal results interfering with the Pareto search.

[0069] As an optional embodiment, the initial solution is optimized according to a preset optimization mode and target indicators to obtain an intermediate solution, including: With the preset optimization mode being Bayesian optimization mode, the evaluation point is determined according to the Gaussian process surrogate model corresponding to the target index under the Pareto hypervolume improvement meaning. The expected future hypervolume improvement is predicted based on the evaluation point, the mean of the target index, and the variance of the target index, and the prediction result is obtained. The prediction result is used to optimize the initial scheme to obtain the intermediate scheme. Determine the deviation between the prediction result and the simulation result. If the deviation is greater than the second preset threshold, trigger agent retraining and mark the initial scheme.

[0070] Specifically, in this embodiment, a Gaussian Process (GP) surrogate model is independently trained for each target indicator (such as travel time, queue length, capacity utilization, imbalance index, and safety alternative indicator) to select the next evaluation point under the Pareto hypervolume improvement context. For example, the Gaussian Process surrogate model uses a Matern5 / 2+ARD (autocorrelation determination for anisotropy) kernel function, and the training data consists of simulated triplets (scheme x, indicator mean μ, variance σ²).

[0071] When the preset optimization mode is Bayesian optimization, the evaluation point is determined according to the Gaussian process surrogate model corresponding to the target index, under the Pareto hypervolume improvement meaning. For example, under the Pareto hypervolume improvement meaning, the q-Expected Hypervolume Improvement (qEHVI) function is used to select the next evaluation point x_{t+1}, q=1, specifically determined by GP. i Their respective means and variances predict the expected future hypervolume improvement; in the surrogate-real dual loop, the surrogate refers to GP. i In the prediction and reality phases, the SUMO subprocess provides the mean μ and variance σ², which are coupled through the Bayesian optimization update rule (Expected Improvement). The execution process of the surrogate-real dual-loop is as follows: The surrogate loop predicts the mean and variance of candidate evaluation points based on existing samples using the Gaussian process model corresponding to each target indicator, and determines the next evaluation point based on the expected improvement (EI) or the qEHVI acquisition function in the multi-objective case. The reality loop sends this evaluation point to the SUMO subprocess for simulation to obtain the observed values ​​of the target indicators, or obtains the sample mean and sample variance through multiple repeated simulations. The simulation results are then added to the training data and the posterior distribution of the Gaussian process is updated, forming a closed-loop coupling that includes surrogate prediction, acquisition function point selection, SUMO simulation verification, and surrogate model update.

[0072] With the preset optimization mode set to Bayesian optimization mode, the expected improvement of the future hypervolume is predicted based on the mean and variance of the evaluation points and target indicators. A surrogate-real dual loop is used to concentrate the simulation cost in the most promising region to obtain the prediction results. The prediction results are used to optimize the initial scheme to obtain the intermediate scheme.

[0073] This embodiment sets up an anti-spoofing mechanism for the surrogate model. With the preset optimization mode set to Bayesian optimization, it monitors the deviation between the surrogate model's predictions and simulation results. If the deviation exceeds a second preset threshold, surrogate retraining is triggered, and the initial scheme is marked as high-value to prevent abnormal simulation results from misleading subsequent sampling. The value of the second preset threshold is set according to actual needs.

[0074] In this implementation, the expected improvement of the hypervolume is predicted by using a Gaussian process surrogate model and the qEHVI acquisition function. Under the surrogate-real dual loop, the limited simulation budget is concentrated in the high-potential region. At the same time, the deviation between the surrogate prediction and the real simulation is monitored. When the deviation exceeds the threshold, retraining is triggered and abnormal samples are marked, which effectively prevents abnormal results from misleading the search and improves the efficiency and reliability of multi-objective optimization.

[0075] As an optional embodiment, the initial solution is optimized according to a preset optimization mode and target indicators to obtain an intermediate solution, including: When the preset optimization mode is the evolutionary multi-objective mode, an evolutionary algorithm based on non-dominated sorting and crowding is used to perform a global Pareto search on the initial scheme, and the initial scheme is optimized according to the search results and soft constraint repair operators to obtain an intermediate scheme. When the preset optimization mode is space-filling sampling mode, determine whether the initial solution violates the engineering hard constraints or the engineering soft constraints. If the initial solution violates the engineering hard constraints, mark the initial solution. If the initial solution violates the soft constraints, the initial solution is optimized by the soft constraint repair operator according to the preset arrangement order to obtain the intermediate solution.

[0076] Specifically, under the preset optimization mode of evolutionary multi-objective mode, an evolutionary algorithm based on non-dominated sorting and crowding is adopted to perform a global Pareto search on the initial scheme under a large budget, and optimize the initial scheme according to the search results and soft constraint repair operators to ensure the feasibility of the offspring and obtain the intermediate scheme.

[0077] Evolutionary multi-objective mode is a method that consumes relatively high computational resources and has a relatively slow running efficiency. It is more suitable for use when reviewing major scenarios, including steps A1 to A4.

[0078] Step A1, Algorithm and Encoding. The NSGA-III (Nondominated Sorting Genetic Algorithm III) algorithm is used to perform nondominated sorting and niche selection on five objective dimensions (such as critical journey time, queue length, capacity utilization, imbalance index, and safe alternative index). Variables are encoded using a hybrid method: speed limit / location / duration / range are encoded as real numbers, action type / density / accident level are encoded as integers, and lane set is encoded as a bit vector. The total dimension D = 8~12. Reference directions are generated by the Das-Dennis rule, with a quantity H = 12.

[0079] Step A2, Initial Population and Evolutionary Operator: The initial population consists of 50-100 P samples from LHS. LHS P loaded from historical decision logs history P from upstream reinjection seed It consists of three parts and shares the same initial population pool as the Bayesian model; the continuous component adopts SBX(η) c =20, p c =0.9)+PM(η m =20, p m=1 / D) generates offspring; discrete / integer components use uniform crossover and bit flipping, while bit vector components use single-point flipping; SBX is an abbreviation for Simulated Binary Crossover; PM is an abbreviation for Polynomial Mutation. Where η c =20 represents the preset cross-distribution index, η m η represents the preset variation distribution index. c and η m Used to control the distribution range and variation of offspring relative to their parents; p c =0.9 indicates that each parent pair undergoes SBX crossover with a 90% probability; D represents the number of continuous decision variables, p m =1 / D indicates that each continuous variable undergoes polynomial mutation with a probability of 1 / D, resulting in an average of approximately one continuous variable mutation in each offspring. The process of generating offspring is as follows: first, two parents are selected from the population; then, their continuous variables are recombinated using SBX to generate temporary offspring; finally, the continuous variables of the temporary offspring are processed according to p... m Perform a polynomial perturbation to obtain the final offspring.

[0080] Step A3: The soft constraint repair operator repairs the offspring sequentially according to risk order, avoiding the merging / bifurcation point ±50m buffer zone at the accident location; speed limits and serious or above accidents on the same road segment cannot coexist; if the duration exceeds 80% of the total simulation time, the simulation is truncated; for multiple actions on the same road segment at the same time, the dominant action is retained according to the information content priority of "speed limit > lane closure > accident"; serious or above accidents cannot occupy all lanes; offspring that fail to repair are eliminated and do not enter the simulation.

[0081] Step A4, Environment Selection and Termination. Merge the parent and offspring into a temporary population of 2 × population size. Fill the next generation population starting from F0 according to the non-dominant sort and reference direction. Remove any excess population based on niche score. Termination conditions include: simulation budget of 300-500 runs, no new schemes entering F0 for 20 consecutive generations, F0 crowding all less than 0.05, and a wall clock interval of 30 minutes. After termination, call the representative candidate interface of Module 5 to output S. repr (|S) repr (|=30), sharing the same output interface with the Bayesian pattern.

[0082] The space-filling sampling pattern serves as a cold-start seed scheme. It employs methods such as Latin hypercube to perform space-filling sampling within the scheme space, providing an initial population for Bayesian or evolutionary schemes.

[0083] When the preset optimization mode is space-filling sampling mode, it is determined whether the initial scheme violates the engineering hard constraints or the engineering soft constraints. If the initial scheme violates the engineering hard constraints, the initial scheme that violates the engineering hard constraints (physical illegality) is directly rejected and marked to ensure that it does not enter the simulation budget. The preset order of arrangement is as follows: the specific repair order is arranged according to the level of engineering risk (accident location > speed limit value > duration), that is, high-risk attributes are repaired first.

[0084] If the initial solution violates soft constraints (engineering preferences, such as a speed limit of 120 km / h coupled with a large-scale accident), the initial solution is adjusted to minimize the disturbance by using a soft constraint repair operator according to a preset arrangement order, resulting in an intermediate solution.

[0085] This embodiment does not use the degenerate mode (the evolutionary multi-objective mode is strictly preserved, and all candidates satisfy the engineering significance of Pareto ordering), and only switches to the space-filling sampling mode to supplement seeds when the simulation budget is extremely tight or the Bayesian proxy bias is too large.

[0086] In this implementation, the evolutionary multi-objective mode uses NSGA-III and non-dominated sorting to achieve global Pareto search, and the soft constraint repair operator adjusts the offspring according to the risk order to ensure the feasibility of the scheme; the space-filling sampling mode generates cold-start seeds through hard constraint filtering and soft constraint repair, providing a high-quality initial population for Bayesian or evolutionary modes, and multiple modes share a unified output interface to enhance the framework's adaptability and robustness.

[0087] As an optional embodiment, after obtaining candidate traffic control solutions, the method further includes: Obtain the action list, target indicator value, percentage improvement relative to the reference solution, and confidence label corresponding to the traffic control candidate solution, and generate summary information of the traffic control candidate solution; Based on the action list, target indicator values, improvement percentages, confidence labels, and summary information, generate structured descriptions of traffic control candidate solutions.

[0088] Specifically, this embodiment is used to generate a structured report. It obtains the action list, target indicator values, percentage improvement relative to the reference plan, and confidence level labels corresponding to the traffic control candidate solutions, and generates a summary of the traffic control candidate solutions.

[0089] Summary information includes, for example, a human-readable one-sentence summary. Each traffic control candidate solution generates a structured description, including a list of actions that can be directly invoked by the downstream policy generation module, the values ​​of each target indicator and the percentage improvement relative to the reference solution, confidence labels, and a human-readable one-sentence summary.

[0090] As an optional embodiment, the method also includes a simulation execution and index feedback module, and a closed-loop iteration and human-machine collaboration module.

[0091] Specifically, the functions of the simulation execution and indicator feedback module include execution deployment, indicator calculation and normalization, and result writing back and interpretation.

[0092] The execution functions include: applying the action sequence proposed by the large language model and verified to the draft of the current session, rendering it to the attachment file and traffic flow file, starting the simulation subprocess to run for a specified duration, and generating simulation output; the index calculation and normalization functions include: extracting key indicators according to the established index rules, and extracting sub-indicators by region, supporting fine-grained evaluation at the mileage segment and lane level; the result writing and interpretation functions include: after the simulation is completed, the result interpretation action allows the large language model to obtain the indicators and regions, and produces natural language interpretation under the constraints of prioritizing facts and prohibiting the fabrication of numbers not appearing in the indicators; the results are also verified by a dedicated schema and can only reference whitelisted region fields.

[0093] The functions of the closed-loop iteration and human-machine collaboration module include target and indicator inversion, candidate library injection, multi-round reflection threshold, and human-machine collaboration breakpoint.

[0094] The target and indicator inversion function includes: after the user provides a target to further reduce the queue by a certain percentage, the system uses the difference between the current indicator and the target as a new context to call the large language model a second time. The large language model tends to adjust the combination of speed limits, accidents, and lane actions, forming a small closed loop including the target, actions, simulation, and indicators. The candidate library injection function includes: using the Pareto front generated by the simulation-optimized pre-module as a reference candidate for the large language model, appearing in the prompt as a summary. The large language model can make local fine-tuning based on this instead of designing out of thin air, greatly improving the feasibility of the initial proposal. The multi-round reflection threshold function includes: to avoid the large language model getting stuck in a self-loop, setting a maximum number of reflection rounds and stopping when the indicator improvement is below a threshold, ensuring that a complete closed loop is completed within minutes in engineering. The human-machine collaborative breakpoint function includes: all drafts, indicators, and large language model summaries of any iteration are persisted, allowing traffic engineers to take over at any round (manually modifying a speed limit or accident and then returning it to the system), supporting a hybrid workflow that includes AI drafting, manual fine-tuning, and AI review.

[0095] This embodiment also provides an apparatus for determining candidate traffic control schemes. This apparatus is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0096] This embodiment provides a device for determining candidate traffic control schemes, such as... Figure 2 As shown, it includes: The simulation module 201 is used to input the road network file, traffic flow file and additional control file rendered by the initial scheme into the traffic simulator, and use the traffic simulator to simulate the initial scheme to obtain simulation results. The index processing module 202 is used to extract multi-dimensional evaluation indicators from the simulation results, and normalize the multi-dimensional evaluation indicators based on the indicators of the reference scheme to obtain the target indicators. The target indicators are used to compare and merge the initial scheme across indicators. The scheme optimization module 203 is used to optimize the initial scheme according to the preset optimization mode and target indicators to obtain the intermediate scheme; Extraction module 204 is used to perform non-dominated sorting on intermediate schemes and extract the Pareto front corresponding to the intermediate schemes based on the sorting results; The filtering module 205 is used to filter intermediate solutions based on the congestion level corresponding to the target indicator when the number of intermediate solutions corresponding to the Pareto front exceeds a first preset threshold, thereby obtaining traffic control candidate solutions.

[0097] In some alternative embodiments, the device further includes: The first generation module is used to generate a structured summary of the traffic control candidate scheme based on the preset information of the traffic control candidate scheme, and write the structured summary into the context anchor of the policy generator as a priority source for downstream recommendation. The module is used to display constraints in the policy generator and obtain the confidence level output by the policy generator based on the confidence level labels of the traffic control candidate solutions. The logging module is used to record the default behaviors of the strategy generator using structured logs.

[0098] In some alternative embodiments, the device further includes: The determination module is used to determine the control plan, which is represented as a structured object. The structured object is used to represent the initial plan. The definition module is used to define legality constraints for each decision variable in the initial scheme; The sampling module is used to discretize the continuous variables in the initial scheme into candidate levels, and to sample based on the discrete variables to obtain the initial scheme.

[0099] In some alternative implementations, simulation module 201 includes: The simulation unit is used to repeatedly run the initial scheme under multiple different random seeds using a traffic simulator to obtain the mean and variance of the index. The obtained unit is used to obtain simulation results based on the index mean and index variance; The first judgment unit is used to determine the cause of failure and mark the initial solution as infeasible if a crash or illegal exit code occurs during the simulation.

[0100] In some optional implementations, the scheme optimization module 203 includes: The prediction unit is used to determine the evaluation point under the Pareto hypervolume improvement meaning based on the Gaussian process surrogate model corresponding to the target index when the preset optimization mode is Bayesian optimization mode. It predicts the expected future hypervolume improvement based on the evaluation point, the mean of the target index and the variance of the target index, and obtains the prediction result. The prediction result is used to optimize the initial scheme to obtain the intermediate scheme. The determination unit is used to determine the deviation between the prediction result and the simulation result. If the deviation is greater than the second preset threshold, the agent retraining is triggered and the initial scheme is marked.

[0101] In some optional implementations, the scheme optimization module 203 includes: The optimization unit is used to perform a global Pareto search on the initial scheme using an evolutionary algorithm based on non-dominated sorting and crowding when the preset optimization mode is an evolutionary multi-objective mode. The initial scheme is then optimized based on the search results and soft constraint repair operators to obtain an intermediate scheme. The second judgment unit is used to determine whether the initial scheme violates the hard or soft constraints of the engineering when the preset optimization mode is space filling sampling mode. If the initial scheme violates the hard constraints of the engineering, the initial scheme is marked. The third judgment unit is used to optimize the initial scheme by using the soft constraint repair operator according to the preset arrangement order if the initial scheme violates the soft constraint, so as to obtain the intermediate scheme.

[0102] In some alternative embodiments, the device further includes: The acquisition module is used to acquire the action list, target indicator value, improvement percentage relative to the reference plan, and confidence label corresponding to the traffic control candidate plan, and generate summary information of the traffic control candidate plan. The second generation module is used to generate structured descriptions of traffic control candidate solutions based on the action list, target indicator values, improvement percentages, confidence labels, and summary information.

[0103] The apparatus for determining candidate traffic control schemes provided in this embodiment of the invention can execute the method for determining candidate traffic control schemes provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0104] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0105] The following is a detailed reference. Figure 3 The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 301, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 302 or a program loaded from memory 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device. The processor 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0106] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0107] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 309, or installed from a memory 308, or installed from a ROM 302. When the computer program is executed by the processor 301, it performs the functions defined in the method for determining traffic control candidate schemes according to embodiments of the present invention.

[0108] Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0109] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the method for determining traffic control candidate schemes shown in the above embodiments is implemented.

[0110] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0111] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for determining candidate traffic control schemes, characterized in that, The method includes: The road network file, traffic flow file, and additional control file rendered from the initial scheme are input into the traffic simulator, and the traffic simulator is used to simulate the initial scheme to obtain simulation results. Multi-dimensional evaluation indicators are extracted from the simulation results. Using the indicators of the reference scheme as a benchmark, the multi-dimensional evaluation indicators are normalized to obtain target indicators. The target indicators are used to compare and merge the initial scheme across indicators. The initial solution is optimized according to the preset optimization mode and the target index to obtain an intermediate solution; The intermediate schemes are sorted non-dominated, and the Pareto front corresponding to the intermediate schemes is extracted based on the sorting results. When the number of intermediate solutions corresponding to the Pareto front exceeds a first preset threshold, the intermediate solutions are filtered according to the congestion level corresponding to the target indicator to obtain traffic control candidate solutions.

2. The method according to claim 1, characterized in that, After obtaining the traffic control candidate solutions, the method further includes: Based on the preset information of the traffic control candidate schemes, a structured summary of the traffic control candidate schemes is generated, and the structured summary is written into the context anchor of the policy generator as a priority source for downstream recommendations. Constraints are displayed in the policy generator, and the confidence level output by the policy generator is obtained based on the confidence level label of the traffic control candidate scheme. The preset behaviors of the policy generator are recorded using structured logs.

3. The method according to claim 1, characterized in that, Before inputting the road network file, traffic flow file, and additional control file rendered from the initial scheme into the traffic simulator, the method further includes: A control scheme is determined, and the control scheme is represented as a structured object, wherein the structured object is used to represent the initial scheme; Define a validity constraint for each decision variable in the initial scheme; The continuous variables in the initial scheme are discretized into candidate levels, and sampling is performed based on the discrete variables to obtain the initial scheme.

4. The method according to claim 1, characterized in that, The simulation of the initial plan using the traffic simulator to obtain simulation results includes: The initial scheme was repeatedly run using the traffic simulator under multiple different random seeds to obtain the index mean and index variance. The simulation results are obtained based on the mean and variance of the indicators. If a crash or illegal exit code occurs during the simulation, determine the cause of the failure and mark the initial solution as infeasible.

5. The method according to claim 1, characterized in that, The step of optimizing the initial solution according to the preset optimization mode and the target index to obtain an intermediate solution includes: When the preset optimization mode is Bayesian optimization mode, the evaluation point is determined according to the Gaussian process surrogate model corresponding to the target index under the Pareto hypervolume improvement meaning. The expected future hypervolume improvement is predicted according to the evaluation point, the mean of the target index and the variance of the target index, and the prediction result is obtained. The prediction result is used to optimize the initial scheme to obtain the intermediate scheme. The deviation between the prediction result and the simulation result is determined. If the deviation is greater than a second preset threshold, agent retraining is triggered and the initial scheme is marked.

6. The method according to claim 5, characterized in that, The step of optimizing the initial solution according to the preset optimization mode and the target index to obtain an intermediate solution includes: When the preset optimization mode is the evolutionary multi-objective mode, the initial scheme is subjected to a global Pareto search using an evolutionary algorithm based on non-dominated sorting and crowding, and the initial scheme is optimized according to the search results and soft constraint repair operators to obtain the intermediate scheme. When the preset optimization mode is space-filling sampling mode, it is determined whether the initial scheme violates the engineering hard constraints or the engineering soft constraints. If the initial scheme violates the engineering hard constraints, the initial scheme is marked. If the initial solution violates the soft constraints, the initial solution is optimized by the soft constraint repair operator according to the preset arrangement order to obtain an intermediate solution.

7. The method according to claim 1, characterized in that, After obtaining the traffic control candidate solutions, the method further includes: Obtain the action list, target index value, improvement percentage relative to the reference scheme, and confidence label corresponding to the traffic control candidate scheme, and generate summary information of the traffic control candidate scheme; Based on the action list, the target indicator value, the improvement percentage, the confidence level label, and the summary information, a structured description of the traffic control candidate solution is generated.

8. An apparatus for determining candidate traffic control schemes, characterized in that, The device includes: The simulation module is used to input road network files, traffic flow files, and additional control files rendered from the initial scheme into the traffic simulator, and use the traffic simulator to simulate the initial scheme to obtain simulation results. The index processing module is used to extract multi-dimensional evaluation indicators from the simulation results, and normalize the multi-dimensional evaluation indicators based on the indicators of the reference scheme to obtain target indicators. The target indicators are used to perform cross-index comparison and merging of the initial scheme. The scheme optimization module is used to optimize the initial scheme according to the preset optimization mode and the target index to obtain an intermediate scheme; The extraction module is used to perform non-dominated sorting on the intermediate schemes and extract the Pareto front corresponding to the intermediate schemes based on the sorting results. The filtering module is used to filter the intermediate solutions according to the congestion level corresponding to the target index when the number of intermediate solutions corresponding to the Pareto front exceeds a first preset threshold, so as to obtain traffic control candidate solutions.

9. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method for determining candidate traffic control schemes as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method for determining a traffic control candidate scheme as described in any one of claims 1 to 7.

Images