A Multi-Objective Signal Control Optimization Method Applicable to Multi-Mode Traffic

Abstract

This invention belongs to the technical field of intelligent traffic signal control, specifically a multi-objective signal control optimization method applicable to multi-modal traffic, including model preparation, constrained multi-objective full-probability Bayesian optimization, and termination iteration. The model preparation part constructs a multi-modal traffic integrated model and obtains a prior training set; the constrained multi-objective full-probability Bayesian optimization part optimizes the next sampling point (a set of signal timing schemes) based on the prior training set from the model preparation part, and inputs the results into the multi-modal traffic integrated model to obtain all objective function values; then, the termination iteration part determines whether to terminate the optimization process. This invention can be used to solve the multi-objective optimization problem of multi-modal traffic signal control. When the conditions set in the termination iteration are met, a satisfactory signal timing scheme can be obtained. This invention is also applicable to model-free scenarios (i.e., completely independent of microscopic traffic simulation platforms and traffic safety assessment software).

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent traffic signal control, and relates to multi-objective Bayesian optimization algorithms, constrained Bayesian optimization algorithms, active learning, full Bayesian Gaussian processes, and multi-modal traffic signal optimization methods. Specifically, it is a multi-objective signal control optimization method applicable to multi-modal traffic. Background Technology

[0002] Transportation systems are diverse and complex (e.g., private cars, buses, pedestrians, bicycles, trucks, emergency vehicles, etc.). To better address traffic control problems, realistic modeling and control research of multimodal transportation systems based on modal interactions is essential. Many previous studies have employed strategies that add buses, bicycles, or pedestrians as alternative modes of transportation to multimodal traffic signal control research (e.g., bus signal priority and pedestrian crossing), primarily focusing on optimizing traffic efficiency. For example, Tang et al., in "Multi-Modal Traffic Signal Control in Shared Space Street," established a periodic-based multimodal (private car, bus, and light rail) traffic signal control optimization model for the multimodal signal coordination problem. The model aimed to minimize the total travel cost and traffic delay of the three modes of transportation, using a particle swarm optimization algorithm to select the optimal signal scheme. Li et al., in "Regional Coordinated Bus Priority Signal Control Considering Pedestrian and Vehicle Delays at Urban Intersections," proposed a regionally coordinated bus priority signal control method. This is a network-level bus priority control method that considers pedestrian and passenger delays, and uses a genetic algorithm to optimize and obtain an approximately optimal network signal scheme. In their paper "Optimal Presignal Control for Two-Mode Traffic at Isolated Signalized Intersections," Khwais and Haddad proposed an optimal presignal control strategy that uses the Pontryagin maximization principle to determine the signaling scheme based on minimizing the total travel time for both private cars and buses.

[0003] Existing multimodal traffic signal control optimization methods suffer from the following problems: First, they primarily address multimodal traffic consisting mainly of private cars and buses, while real-world traffic environments are far more complex, making it difficult to obtain signal timing schemes suitable for real-world traffic conditions. Second, they often prioritize traffic efficiency as the optimization objective, failing to optimize traffic flows composed of multiple traffic modes from a safer, fairer, and more efficient perspective. Third, existing heuristic optimization algorithms often require numerous iterations to obtain an approximate optimal solution, while multimodal traffic models, due to the interaction and influence of multiple traffic modes, are complex and computationally expensive. This strategy of trading high computational costs for better signal timing schemes hinders their application to real-world traffic problems. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a multi-objective signal control optimization method applicable to multi-modal traffic. Based on traffic flow data, turning ratio data, original signal timing scheme data, and network road baseline data for all road segments, this invention constructs a multi-modal traffic integration model using microscopic traffic simulation software and traffic safety analysis software. An optimizer is designed based on the microscopic traffic simulation model, the traffic safety analysis model, and a constrained multi-objective fully probabilistic Bayesian optimization algorithm. This optimizer is used to solve the aforementioned multi-modal multi-objective traffic signal optimization model until the optimal signal timing scheme for the intersection is obtained. The decision variables for the signal control optimization problem to be solved by this invention are signal control parameters (phase display order, cycle duration, green light time, and phase difference).

[0005] The technical solution of this invention:

[0006] A multi-objective signal control optimization method applicable to multimodal traffic includes model preparation, constrained multi-objective full-probability Bayesian optimization, and iteration termination. The model preparation part constructs a multimodal traffic integrated model and obtains a prior training set. The constrained multi-objective full-probability Bayesian optimization part optimizes the next sampling point (a set of signal timing schemes) based on the prior training set from the model preparation part, and inputs the results into the multimodal traffic integrated model to obtain all objective function values. Finally, the iteration termination part determines whether to terminate the optimization process.

[0007] The specific steps are as follows:

[0008] Step 1. Model Preparation

[0009] (1.1) Building a multimodal transportation integration model

[0010] The multimodal traffic integration model is built using micro-traffic simulation software and traffic safety assessment software. First, reliable road network data, vehicle attributes and proportions in the network traffic flow, original signal timing data, traffic flow data for each road segment, and steering ratio data are input into the micro-traffic simulation software (such as VISSIM or SUMO) to construct a micro-traffic simulation model. Second, running the micro-traffic simulation model allows for the acquisition of operational status and trajectory data for all vehicles. Finally, the vehicle trajectory data is input into the traffic safety assessment software (such as SSAM) to calculate the total number of traffic conflicts.

[0011] Based on the multimodal traffic integration model, users can obtain values ​​for various objective functions. These objective functions can be divided into three categories: traffic efficiency assessment functions, traffic equity assessment functions, and traffic safety assessment functions. The values ​​of traffic efficiency assessment functions (such as total waiting time, total delay time, and average travel time) and traffic equity assessment functions (such as the Gini coefficient) are calculated using operational status data from all vehicles; the values ​​of traffic safety assessment functions (such as conflict rate and estimated number of traffic accidents) are calculated using the total number of traffic conflicts.

[0012] In summary, the input of the multimodal traffic integration model is a set of signal timing schemes, and the output is the corresponding multi-class objective function values.

[0013] (1.2) Obtain the prior training set required for constrained multi-objective full probability Bayesian optimization and set the relevant parameters.

[0014] Selected by Latin hypercube sampling method Group signal timing scheme , ,in Indicates the first Group signal timing schemes. Input these schemes one by one into the multimodal traffic integration model constructed in step (1.1) to obtain the corresponding multi-class objective function values. The initial objective function set is... , ,in Indicates the first Group 1 The objective function value, This represents the number of objective functions. Finally, a prior training set is constructed based on the signal timing scheme and the objective function values. .

[0015] Let the number of iterations be denoted as . ,make Preset the maximum number of iterations Pre-set the maximum values ​​of multiple objective functions. and minimum value ,based on The initial Pareto front hypervolume value is calculated and denoted as . ,counter And preset the maximum value of the counter. .

[0016] Step 2. Constrained Multi-Objective Full Probabilistic Bayesian Optimization

[0017] (2.1) Using the prior training set obtained in step (1.2), fit the latent function based on the full Bayesian Gaussian process regression model. With decision variables The relationship between (i.e., signal control parameters). It is the latent function and Gaussian noise The observed value, i.e. , , .

[0018] Let the dataset used for Gaussian process regression training be denoted as . ,make Then the multivariate Gaussian distribution can be expressed as:

[0019]

[0020] in , ; , It is the covariance kernel function. The kernel function is represented by hyperparameters. For parameterization, the kernel function can be a radial basis function, a Marton kernel function, or a quadratic rational function, etc.

[0021] (2.2) Use full Bayesian estimation to estimate the hyperparameters of the model in step (2.1).

[0022] Placing prior information on hyperparameters And approximates the complete posterior distribution of the model, that is:

[0023]

[0024] Then the Markov Chain Monte Carlo (MCMC) sampling method is used to select... The optimal hyperparameters for each sample are determined by maximizing the posterior distribution. ,Right now:

[0025]

[0026] in , .

[0027] (2.3) Determine the next sampling point based on the sampling results of step (2.2). The mean and variance of.

[0028] Each MCMC sampling can obtain one , Thus, the corresponding mean and variance can be obtained, that is:

[0029]

[0030]

[0031] in yes The covariance matrix between the training set input and the training set input. yes and The covariance matrix between them.

[0032] Each prediction based on MCMC sampling can be viewed as a Gaussian mixture model, so the hierarchical prediction posterior is actually... A mixture of Gaussian processes. The mean of the posterior distribution at the next sampling point. and variance Divided into:

[0033]

[0034]

[0035] (2.4) Based on the mean of step (2.3) and variance Construct an acquisition function and maximize the acquisition function to determine the next sampling point.

[0036] Maximize the function Determine the next sampling point ,Right now:

[0037]

[0038] in It is the decision variable space that satisfies the period duration constraint. , It is an improved probability function based on hypervolume. These are weight parameters. It is a posteriori The derived mean function The variance is given, that is .

[0039] The calculation process is as follows:

[0040]

[0041]

[0042]

[0043] in This represents the increase in the Pareto front hypervolume. , It is the first The mean function of the objectives; It is the current set of Pareto front points; It is a hypervolume indicator function; Indicates the probability of improvement; Represents the non-dominated region of the objective function space; It is the first The probability density function of the objective function; It is the first The latent function of an objective.

[0044] (2.5) Evaluate the next sampling point

[0045] The result obtained in step (2.4) Input the data into the multimodal traffic integration model in step (1.1), and output the target value corresponding to the next sampling point. .

[0046] (2.6) Calculate the hypervolume value of the Pareto front

[0047] Based on the results of step (2.5), the current Pareto front point set is updated and determined, and then based on... and Calculate the area (hypervolume) of the non-dominated region of the current Pareto front. .if Let the counter ;if ,but .

[0048] Step 3. Terminate the iteration

[0049] Determine if the current iteration meets the termination condition. There are two termination conditions: the first is that the number of iterations exceeds the maximum number of iterations. The second condition is that the counter exceeds its maximum value. .if or If the signal timing scheme is correct, the signal timing scheme corresponding to the Pareto front convergence point is returned directly; otherwise, the training set is updated. , Return to step 2 and continue iterative optimization.

[0050] Users can utilize this invention to solve multi-objective optimization problems in multi-modal traffic signal control. A satisfactory signal timing scheme can be obtained when the conditions set in the termination iteration are met. This invention is also applicable to model-free scenarios (i.e., completely independent of microscopic traffic simulation platforms and traffic safety assessment software). Where practical conditions permit, users can select an experimental area in the real world, set the signal timing scheme for that area, and obtain traffic evaluation data over a certain period using detection equipment installed on roads and intersections, thereby calculating the multi-objective function values.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] (1) This invention can solve the multi-objective optimization problem of multi-modal traffic signal control. Previous solutions to traffic signal control optimization problems have mostly focused on the traffic efficiency of motor vehicles. However, real-world traffic is often composed of multiple modes of transportation, and maximizing system traffic efficiency can no longer be the sole criterion for evaluating the quality of the traffic environment. Previous solutions lacked consideration of the mutual influence of multiple modes of traffic at intersections. The multi-objective signal optimization method of this invention can not only realistically reflect the traffic environment, but also enable all types of travelers to pass through signal-controlled intersections more safely, fairly, and efficiently.

[0053] (2) This invention uses a constrained multi-objective full-probability Bayesian optimization algorithm, which can quickly find the approximate optimal solution in the constrained space (i.e., significantly reduce the computational cost) and ensure the superiority of the Pareto front set (i.e., high-quality solution under multi-objective trade-off optimization). In the past, genetic algorithms were often used, which had a slow convergence speed and were prone to obtaining local optima. However, this invention incorporates an active learning method into the method used, which can quickly find the optimal solution under multi-objective trade-off optimization in the constrained space. Attached Figure Description

[0054] Figure 1 This is a flowchart of the constrained multi-objective full-probability Bayesian optimization in this invention;

[0055] Figure 2 This is a schematic diagram of the multimodal traffic integration model in the embodiment;

[0056] Figure 3 This is a three-dimensional Pareto front plot in the embodiment;

[0057] Figure 4 This is a two-dimensional planar plot of the average waiting time versus the Gini coefficient in the embodiment;

[0058] Figure 5 This is a two-dimensional planar graph showing the average waiting time versus the conflict rate in the embodiment.

[0059] Figure 6 This is a two-dimensional planar diagram of the Gini coefficient and conflict rate in the embodiment;

[0060] Figure 7 This is a graph showing how the hypervolume value changes with the number of iterations in the example. Detailed Implementation

[0061] The following describes in detail the specific embodiments of the present invention with reference to the accompanying drawings and technical solutions, and simulates the implementation effects of the invention.

[0062] This embodiment uses a typical signalized intersection in Copenhagen, Denmark as a case study for verification and optimization of the signal timing scheme. The final result is the optimal signal timing scheme considering multiple objectives, which simultaneously balances traffic safety, traffic equity, and traffic efficiency. Details are as follows:

[0063] 1. Construction of a multimodal transportation integration model

[0064] A typical signalized intersection in a central urban area of ​​a city was selected as the research object. The traffic composition of this intersection mainly consists of private cars, buses, and bicycles. There is no dedicated bus lane or corresponding dedicated phase for buses, while there is a dedicated bicycle lane but no dedicated phase for bicycles.

[0065] Based on field research and by combining data from HCM2010 and the Danish Transport Yearbook 2020, the following traffic data for the evening rush hour from 17:00 to 18:00 was obtained:

[0066] Table 1. Traffic data at the intersection.

[0067]

[0068] There are a total of 8 bus stops and 12 bus routes in the four directions of the intersection, with buses departing every 15 minutes.

[0069] The initial signal timing scheme has four phases with a cycle length of 124 seconds, and the yellow light is set to 4 seconds. The specific phases and green light durations are shown in Table 2.

[0070] Table 2 Original signal timing scheme for intersections

[0071]

[0072] Using the above data, a multimodal traffic integration model is constructed using urban micro-traffic simulation software (SUMO in this embodiment) and traffic safety assessment software (SSAM in this embodiment). Figure 1As shown in the figure. In this embodiment, the simulation run time is set to 4500 seconds, of which the first 900 seconds are the unstable period of the simulation model, and the last 3600 seconds are the stable period of the model (i.e., the effective simulation time). By running the simulation model and recording and statistically analyzing the results during the effective simulation time, the waiting time and the number of traffic conflicts for all vehicles at the intersection are obtained.

[0073] 2. Description of solving the multi-mode, multi-objective signal control optimization problem

[0074] To simplify the problem, this embodiment considers the traffic signal optimization problem under timed signal control. Specifically, the phase structure and phase display order are fixed, the cycle duration is fixed, and only the green light ratio of the four phases is optimized. Based on the principle of minimum green light time to ensure safe pedestrian crossing, the minimum green light time for the vehicle phase is set to 5 seconds. The above multi-mode multi-objective signal optimization problem can be expressed by the following formula:

[0075]

[0076] st

[0077]

[0078]

[0079] in Represents the objective function vector. This is the first phase of the green light time. This is the second phase of the green light time. This is the third phase of the green light time; It is the average waiting time (traffic efficiency) at the intersection. It is the Gini coefficient (traffic equity). It is the traffic conflict rate (traffic safety).

[0080] 3. Solve the above problem

[0081] In this embodiment, the relevant parameters are set as follows: , , , , Based on the cycle time and minimum green light time constraints, the value range of all decision variables in this embodiment is set to... .

[0082] The method of this invention optimizes the signal timing scheme of signal-controlled intersections from three dimensions: traffic safety, fairness, and efficiency. The resulting Pareto front is as follows: Figure 2 As shown, at the same time Figure 4 , Figure 5 and Figure 6 A two-dimensional planar diagram showing the pairwise relationships between the three targets is provided. Figure 7 The graph shows the change in the Pareto front hypervolume value with the number of iterations. It can be seen from the graph that the hypervolume value changes very little starting from the 250th iteration, indicating that the optimization iteration has converged. The initial Pareto front value based on the prior dataset is 25455. After 300 iterations, the final Pareto front value is 30972, an improvement of 21.67% compared to the initial state. This demonstrates a significant improvement in the Pareto front, indicating a very good improvement effect in this implementation case. Table 3 shows the signal timing schemes corresponding to the convergence points on the Pareto front. Table 4 compares the results before and after optimization using one of the optimization results, showing that the method of this invention can minimize the average waiting time of the system while ensuring fairer right-of-way and improving traffic safety. Furthermore, users can choose an appropriate signal timing scheme based on actual traffic conditions. If users prioritize traffic safety, they can choose a scheme with the lowest possible traffic conflict rate; if users prioritize traffic fairness, they can choose a scheme with a lower Gini coefficient.

[0083] Table 3. Meeting points on the Pareto front and corresponding signal timing schemes

[0084]

[0085] Table 4 Comparison of results before and after optimization

[0086]

Claims

1. A multi-objective signal control optimization method applicable to multi-modal traffic, characterized in that, The process includes model preparation, constrained multi-objective full-probability Bayesian optimization, and iteration termination. The model preparation part constructs a multi-modal traffic integration model and obtains a prior training set. The constrained multi-objective full-probability Bayesian optimization part optimizes the next sampling point (i.e., a set of signal timing schemes) based on the prior training set from the model preparation part, and inputs the results into the multi-modal traffic integration model to obtain all objective function values. Finally, the iteration termination part determines whether to terminate the optimization process. The specific steps are as follows: Step 1. Model Preparation (1.1) Building a multimodal transportation integration model The multimodal traffic integration model is built using microscopic traffic simulation software and traffic safety assessment software. First, reliable road network data, vehicle attributes and proportions in the network traffic flow, original signal timing scheme data, traffic flow data for each road segment, and steering ratio data are input into the microscopic traffic simulation software to construct a microscopic traffic simulation model. Second, by running the microscopic traffic simulation model, the operating status and trajectory data of all vehicles can be obtained. Finally, the vehicle trajectory data is input into the traffic safety assessment software to calculate the total number of traffic conflicts. Based on a multimodal traffic integration model, multiple objective function values ​​are obtained. These objective functions are divided into three categories: traffic efficiency evaluation function, traffic equity evaluation function, and traffic safety evaluation function. The values ​​of the traffic efficiency evaluation function and the traffic equity evaluation function are calculated using the operational status data of all vehicles. The value of the traffic safety evaluation function is calculated using the total number of traffic conflicts. Therefore, the input of the multimodal traffic integration model is a set of signal timing schemes, and the output is the corresponding multi-class objective function values; (1.2) Obtain the prior training set required for constrained multi-objective full probability Bayesian optimization and set the relevant parameters. Selected by Latin hypercube sampling method Group signal timing scheme , ,in Indicates the first Group signal timing schemes; input these schemes one by one into the multimodal traffic integration model constructed in step (1.1) to obtain the corresponding multi-class objective function values, the initial objective function set is as follows. , ,in Indicates the first Group 1 The objective function value, The number of objective functions is given; finally, a prior training set is constructed based on the signal timing scheme and the objective function values. ; Let the number of iterations be denoted as... ,make Preset the maximum number of iterations Pre-set the maximum value of multiple objective functions and minimum value ,based on The initial Pareto front hypervolume value is calculated and denoted as . ,counter And preset the maximum value of the counter. ; Step 2. Constrained Multi-Objective Full Probabilistic Bayesian Optimization (2.1) Using the prior training set obtained in step (1.2), fit the latent function based on the full Bayesian Gaussian process regression model. With decision variables That is, the relationship between signal control parameters; and It is the latent function and Gaussian noise The observed value, i.e. , , ; Let the dataset used for Gaussian process regression training be denoted as . ,make Then the multivariate Gaussian distribution is expressed as: ； in , ; , It is the covariance kernel function. The kernel function is represented by hyperparameters. Perform parameterization; (2.2) Use full Bayesian estimation to estimate the hyperparameters of the model in step (2.1); Placing prior information on hyperparameters And approximates the complete posterior distribution of the model, that is: ； Then the Markov chain Monte Carlo (MCMC) sampling method is used to select... The optimal hyperparameters for each sample are determined by maximizing the posterior distribution. ,Right now: ； in , ; (2.3) Determine the next sampling point based on the sampling results of step (2.2). The mean and variance; Each MCMC sampling yields one , Thus, the corresponding mean and variance are obtained, i.e.: ； ； in yes The covariance matrix between the training set input and the training set input. yes and The covariance matrix between them; Each prediction based on MCMC sampling can be viewed as a Gaussian mixture model, so the hierarchical prediction posterior is actually... A mixture of Gaussian processes; the mean of the posterior distribution at the next sampling point. and variance Divided into: ； ； (2.4) Based on the mean of step (2.3) and variance Construct an acquisition function and maximize the acquisition function to determine the next sampling point; Maximize the function Determine the next sampling point ,Right now: ； in It is the decision variable space that satisfies the period duration constraint. , It is an improved probability function based on hypervolume. These are weight parameters. It is a posteriori The derived mean function The variance is given, that is ; The calculation process is as follows: ； ； ； in This represents the increase in the Pareto front hypervolume. , It is the first The mean function of the objectives; It is the current set of Pareto front points; It is a hypervolume indicator function; Indicates the probability of improvement; Represents the non-dominated region of the objective function space; It is the first The probability density function of the objective function; It is the first The latent function of an objective; (2.5) Evaluate the next sampling point The result obtained in step (2.4) Input the data into the multimodal traffic integration model in step (1.1), and output the target value corresponding to the next sampling point. ; (2.6) Calculate the hypervolume value of the Pareto front Based on the results of step (2.5), the current Pareto front point set is updated and determined, and then based on... and Calculate the area of ​​the non-dominated region of the current Pareto front. ;if Let the counter ;if ,but ; Step 3. Terminate the iteration Determine if the current iteration meets the termination condition; there are two termination conditions: the first condition is that the number of iterations exceeds the maximum number of iterations. The second condition is that the counter exceeds its maximum value. ;if or If the signal timing scheme is correct, the signal timing scheme corresponding to the Pareto front convergence point is returned directly; otherwise, the training set is updated. , Return to step 2 and continue iterative optimization.

Images