Intelligent parameter calibration method for freeway weaving area microscopic traffic simulation model

Abstract

The application discloses a parameter intelligent calibration method for a microcosmic traffic simulation model of an expressway interlaced area, and belongs to the technical field of microcosmic traffic simulation. The method comprises the following steps: acquiring road geometry data, road flow data and vehicle trajectory data of a target road section; based on the acquired data, establishing a microcosmic traffic simulation basic model comprising at least one car-following behavior model and one lane-changing behavior model; constructing a reinforcement learning intelligent agent for calibrating the parameters of the microcosmic traffic simulation basic model, wherein the reinforcement learning intelligent agent adopts a double-layer decision framework; training the reinforcement learning intelligent agent by using a proximal policy optimization algorithm; and outputting an optimal parameter combination corresponding to the trained reinforcement learning intelligent agent as the parameters of the microcosmic traffic simulation model. The application realizes the automation and intelligentization of simulation parameter calibration, can effectively avoid local optimization, and significantly improves the authenticity and reliability of traffic simulation in the interlaced area.

Description

Technical Field

[0001] This invention relates to the field of microscopic traffic simulation, and in particular to a method for intelligent parameter calibration of a microscopic traffic simulation model for highway weaving areas. Background Technology

[0002] Microscopic traffic simulation technology can accurately reproduce the interactive behaviors of individual vehicles, such as following and changing lanes. It is an indispensable core tool for analyzing traffic conditions in weaving areas. However, the effectiveness and reliability of the simulation model depend heavily on the accuracy of its inherent behavioral parameter settings.

[0003] As a critical node in the road system, weaving zones are prone to traffic disorder and congestion due to lane-changing and weaving phenomena. Unlike other regular road sections, traffic flow in weaving zones is influenced by the interaction of three traffic flows: mainline traffic, ramp inflow, and mainline outflow. This results in highly dynamic and nonlinear traffic flow characteristics, thus demanding higher accuracy from simulation models. Using default parameters directly in simulations may distort key indicators such as traffic speed, density, and delays, severely impacting the reliability and application value of the simulation results.

[0004] Current simulation parameter calibration mainly relies on manual parameter adjustment, which is inefficient and heavily dependent on personal experience, making it difficult to guarantee the accuracy and optimality of the calibration results. Therefore, there is an urgent need for an efficient, automated parameter calibration method with collaborative optimization capabilities to improve the accuracy and practicality of microscopic traffic simulation models in highway weaving areas. Summary of the Invention

[0005] Purpose of the invention: To address the above problems, the purpose of this invention is to provide an intelligent parameter calibration method for a microscopic traffic simulation model of a highway weaving area.

[0006] Technical solution: The intelligent parameter calibration method for the microscopic traffic simulation model of highway weaving areas of the present invention includes the following steps:

[0007] Step 1: Obtain road geometry data, traffic flow data, and vehicle trajectory data for the target road segment;

[0008] Step 2: Based on the acquired data, establish a basic micro-traffic simulation model that includes at least one car-following behavior model and one lane-changing behavior model.

[0009] Step 3: Construct a reinforcement learning agent for parameter calibration of the microscopic traffic simulation basic model. The reinforcement learning agent adopts a two-layer decision framework, including: an upper-layer decision module, which dynamically selects the type of behavior model to be optimized based on the comparison between the simulation output and the real data; and a lower-layer execution module, which outputs the adjustment amount of one or more specific parameters of the selected type of behavior model based on the selection of the upper-layer decision module.

[0010] Step 4: Train the reinforcement learning agent using the proximal policy optimization algorithm;

[0011] Step 5: Output the optimal parameter combination corresponding to the reinforcement learning agent after training, as the parameters of the micro-traffic simulation model.

[0012] Furthermore, the steps for constructing a reinforcement learning agent for parameter calibration of a microscopic traffic simulation model include:

[0013] Define the state space as ,in, For state space, For traffic volume, The average speed of the vehicle. The average acceleration of the vehicle. The number of times a vehicle changes lanes. This refers to the lateral speed during lane changes;

[0014] Define the action space as ,in, This is the maximum acceleration adjustment value. The desired vehicle speed adjustment value. For safe headway adjustment value, For comfort deceleration adjustment value, Adjustment value for lane-changing intention. This is the minimum safety clearance adjustment value;

[0015] Define the reward function as follows: ,in, As a reward value, This is the upper-level reward value. This is the reward value for the lower level. To collaboratively optimize reward values, The penalty value. This is the upper-level reward value coefficient. This is the coefficient for the lower-level reward value. To collaboratively optimize the reward value coefficient, This is the penalty value coefficient.

[0016] Furthermore, step 4 includes:

[0017] Step 41: Construct a two-layer actor network, consisting of an upper-layer actor network and a lower-layer actor network. The input to the upper-layer actor network is the state space. The output is the action probability distribution. , where parameters 0 represents optimized car-following parameters, and 1 represents optimized lane-changing parameters;

[0018] For the lower-level actor network, its input is the state space. and parameters of the upper layer actions The output is a continuous action vector. mean and standard deviation ,Right now ,in, It is the identity matrix;

[0019] Construct a network of critics, with the state space as its input. The output is a value estimate for that state. , used to evaluate the long-term expected return of a state;

[0020] Step 42: In each round of training, update the parameters by interacting with the environment. ,in, For the first The state space at the next iteration. For the first Action space during the next iteration For the first The reward function value at the next iteration;

[0021] The advantage function is calculated using generalized advantage estimation. The calculation formula is:

[0022] ,

[0023] ,

[0024] In the formula, As a discount factor, For parameters to estimate the generalized dominance;

[0025] Step 43, construct the actor loss function, expressed as:

[0026] ,

[0027] In the formula, Indicates time step Expectations This represents all trainable parameters in the two-layer actor network; This represents a clipping function that restricts input values ​​to a range. Inside, This represents the probability ratio between the old and new strategies. These are the trimming parameters;

[0028] Construct the critic loss function, expressed as:

[0029] ,

[0030] In the formula, This indicates that the critics' network is in response to the state. The value estimate, For target value;

[0031] The total loss function is calculated using the following formula:

[0032] ,

[0033] In the formula, For the value function loss weights, For entropy reward weight, The entropy of the policy distribution;

[0034] Step 44: Determine if the current iteration count has reached the maximum iteration count. ,like If the iteration stops, the strategy with the highest reward value in the training history is selected as the final calibration result.

[0035] like Calculate the nearest The moving average of the reward function values ​​over the training epochs If the moving average is in continuous In this assessment, the absolute value of its change is less than If the training is considered to have converged, the iteration stops. The policy obtained at this point is the optimal parameter calibration policy. Otherwise, the iteration count is updated. Then return to step 41 for the next iteration.

[0036] Furthermore, the process of obtaining the action probability distribution in the upper-level actor network includes:

[0037] state space After two layers of nonlinear transformation, a two-dimensional probability distribution is output through the Softmax function, as follows:

[0038] ,

[0039] ,

[0040] ,

[0041] ,

[0042] In the formula, This represents the first hidden layer after the first layer is activated; These are the weighting parameters for the first layer of nonlinear transformation; These are the bias parameters for the first layer of nonlinear transformation; This represents the second hidden layer after the second activated layer; These are the weighting parameters for the second-level nonlinear transformation; These are the bias parameters for the second-level nonlinear transformation; The raw scores for the two actions; These are the weight parameters for the output layer; These are the bias parameters for the output layer; This is the activation function.

[0043] Furthermore, the basic model of micro-traffic simulation is established based on micro-traffic simulation software, the car-following behavior model is the intelligent driver model, and the lane-changing behavior model is the SL2015 model.

[0044] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0045] 1. This invention automates and automates the parameter calibration process based on a two-layer reinforcement learning model, which can significantly shorten the calibration cycle, reduce reliance on human experience, and improve the systematicness and repeatability of calibration work.

[0046] 2. By dynamically deciding whether to optimize the following parameters or lane-changing parameters in the upper-level module and executing specific parameter adjustments in the lower-level module, the collaborative optimization of the two types of key behavioral parameters is achieved. This mechanism can effectively avoid the local optimum problem caused by adjusting a single parameter, making the simulation model more adaptable to traffic flow in complex weaving areas.

[0047] 3. This invention constructs a state space and reward function using multiple real traffic indicators, and continuously approximates the real traffic flow characteristics through reinforcement learning, thereby making the simulation output more statistically close to the actual observation data, significantly improving the credibility and practicality of the simulation results. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the interlacing area;

[0049] Figure 2 This is the initial simulation process;

[0050] Figure 3 Optimize the process for parameters;

[0051] Figure 4 A comparison chart of reward curves;

[0052] Figure 5This is a comparison chart of reward differences. Detailed Implementation

[0053] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0054] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0055] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0056] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0057] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0058] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0059] The intelligent parameter calibration method for the microscopic traffic simulation model of the highway weaving zone described in this embodiment includes the following steps:

[0060] Step 1: Obtain road geometry data, traffic flow data, and vehicle trajectory data for the target road segment.

[0061] The target road sections include highway weaving areas, ramp merging areas, or diverging areas. This example mainly focuses on parameter calibration for highway weaving area scenarios. Figure 1 This is a schematic diagram of the weaving area. This method is also applicable to the merging or diverging areas of ramps.

[0062] In one example, the collected road geometry data includes road length, number of lanes, lane width, and road signs and markings; road traffic flow data includes the number of vehicles in each lane, vehicle travel time, and headway; vehicle trajectory data includes: vehicle speed, acceleration, horizontal coordinate, vertical coordinate, vehicle length, vehicle width, and the lane ID of the vehicle.

[0063] Step 2: Based on the acquired data, establish a basic micro-traffic simulation model that includes at least one car-following behavior model and one lane-changing behavior model.

[0064] Furthermore, the basic model for microscopic traffic simulation is built upon microscopic traffic simulation software, such as SUMO software. The car-following behavior model is an intelligent driver model, and the lane-changing behavior model is the SL2015 model.

[0065] Evaluation metrics were selected, and several key simulation results were chosen as the basis for calibrating the parameters of the vehicle following behavior model and lane-changing behavior model. The simulation results included traffic volume, average vehicle speed, vehicle acceleration, number of lane changes, and lane-changing duration. Simulation parameters such as maximum acceleration, desired vehicle speed, safe headway, comfortable deceleration, lane-changing intention, and minimum safe clearance were set to default values. Microscopic traffic simulation was then started, and the corresponding simulation results were output. The flowchart of the initial simulation is shown below. Figure 2 As shown.

[0066] Step 3: Construct a reinforcement learning agent for parameter calibration of the microscopic traffic simulation basic model. The reinforcement learning agent adopts a two-layer decision framework, including: an upper-layer decision module, which dynamically selects the type of behavior model to be optimized based on the comparison between the simulation output and the real data; and a lower-layer execution module, which outputs the adjustment amount of one or more specific parameters of the selected type of behavior model based on the selection of the upper-layer decision module.

[0067] Furthermore, the steps for constructing a reinforcement learning agent for parameter calibration of a microscopic traffic simulation model include:

[0068] Define the state space as ,in, For state space, For traffic volume, The average speed of the vehicle. The average acceleration of the vehicle. The number of times a vehicle changes lanes. This refers to the lateral speed during lane changes;

[0069] Define the action space as ,in, This is the maximum acceleration adjustment value. The desired vehicle speed adjustment value. For safe headway adjustment value, For comfort deceleration adjustment value, Adjustment value for lane-changing intention. This is the minimum safety clearance adjustment value;

[0070] Define the reward function as follows: ,in, As a reward value, This is the upper-level reward value coefficient. This is the coefficient for the lower-level reward value. To collaboratively optimize the reward value coefficient, This is the penalty value coefficient; This is the upper-level reward value. This is the reward value for the lower level. To collaboratively optimize reward values, The penalty value is expressed as follows:

[0071] ,

[0072] ,

[0073] In the formula, The average speed weight of the upper-level reward value. The upper-level reward value is the vehicle acceleration weight. The weight of the number of lane changes is used to determine the upper-level reward value. The horizontal speed weight is the upper-level reward value. This represents actual traffic volume. To simulate traffic volume, For the true average speed, To simulate the average speed, To simulate the acceleration of real vehicles, To simulate vehicle acceleration, To reflect the actual number of lane changes, To simulate the number of lane changes, For the actual lane change duration, To simulate the duration of lane changes, This represents the variance of the corresponding parameter.

[0074] ,

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] In the formula, For the improved value of the car-following parameter, These are improved values ​​for lane-changing parameters. For the new round of simulation car-following parameter results, The results of the previous simulation car-following parameters, For the new round of simulation lane-changing parameter results, These are the lane-changing parameter results from the previous simulation.

[0080] ,

[0081] ,

[0082] In the formula, To exceed the maximum acceleration number of times, The total number of vehicles in operation. This represents the number of times the distance between vehicles is less than the minimum safe distance. As the weight of the first penalty item, As the weight of the second penalty term, This is the weight of the third penalty item.

[0083] Step 4: Train the reinforcement learning agent using the proximal policy optimization algorithm.

[0084] Combination Figure 3 The parameter optimization process shown further includes, in step 4:

[0085] Step 41: Construct a two-layer actor network, consisting of an upper-layer actor network and a lower-layer actor network. The input to the upper-layer actor network is the state space. The output is the action probability distribution. , where parameters 0 represents optimized car-following parameters, and 1 represents optimized lane-changing parameters;

[0086] If the upper-level judgment result is to optimize the following parameters, the lower-level mainly modifies the maximum acceleration, desired speed, safe headway, and comfortable deceleration; if the upper-level judgment result is to optimize the lane-changing parameters, the lower-level mainly modifies the lane-changing intention and minimum safe clearance.

[0087] For the lower-level actor network, its input is the state space. and parameters of the upper layer actions The output is a continuous action vector. mean and standard deviation ,Right now ,in, It is the identity matrix;

[0088] Construct a network of critics, with the state space as its input. The output is a value estimate for that state. , used to evaluate the long-term expected return of a state;

[0089] Step 42: In each round of training, update the parameters by interacting with the environment. ,in, For the first The state space at the next iteration. For the first Action space during the next iteration For the first The reward function value at the next iteration;

[0090] The advantage function is calculated using generalized advantage estimation. The calculation formula is:

[0091] ,

[0092] ,

[0093] In the formula, As a discount factor, For parameters to estimate the generalized dominance;

[0094] Step 43, construct the actor loss function, expressed as:

[0095] ,

[0096] In the formula, Indicates time step Expectations This represents all trainable parameters in the two-layer actor network; This represents a clipping function that restricts input values ​​to a range. Inside, This represents the probability ratio between the old and new strategies. The pruning parameter is typically set to a value between 0.1 and 0.2. A value of 0.2 is used to prevent the strategy update step from being too large.

[0097] Construct the critic loss function, expressed as:

[0098] ,

[0099] In the formula, This indicates that the critics' network is in response to the state. The value estimate, For target value;

[0100] The total loss function is calculated using the following formula:

[0101] ,

[0102] In the formula, The loss weight for the value function (usually taken as 0.5). This is the entropy reward weight (usually 0.01). The entropy of the policy distribution is used to encourage exploration;

[0103] Step 44: Determine if the current iteration count has reached the maximum iteration count. ,like Take 500, if If the iteration stops, the strategy with the highest reward value in the training history is selected as the final calibration result.

[0104] like Calculate the nearest The moving average of the reward function values ​​over the training epochs If the moving average is in continuous In this assessment, the absolute value of its change is less than If the training is considered convergent and the reward function no longer shows significant improvement, the iteration stops. The policy obtained at this point is the optimal parameter calibration policy. Otherwise, the iteration count is updated. Then return to step 41 for the next iteration.

[0105] Furthermore, the process of obtaining the action probability distribution in the upper-level actor network includes:

[0106] state space After two layers of nonlinear transformation, a two-dimensional probability distribution is output through the Softmax function, as follows:

[0107] ,

[0108] ,

[0109] ,

[0110] ,

[0111] In the formula, This represents the first hidden layer after the first layer is activated; These are the weighting parameters for the first layer of nonlinear transformation; These are the bias parameters for the first layer of nonlinear transformation; This represents the second hidden layer after the second activated layer; These are the weighting parameters for the second-level nonlinear transformation; These are the bias parameters for the second-level nonlinear transformation; The raw scores for the two actions; These are the weight parameters of the output layer. , , The initial weights are randomly sampled from a uniform distribution with a mean close to 0, and the initial biases are all 0. These are the bias parameters for the output layer; The activation function is expressed as:

[0112] .

[0113] For the commentator network, the core processing part in the middle includes a fully connected layer of 256 neurons, a first ReLU activation layer, a fully connected layer of 128 neurons, and a second ReLU activation layer. The output layer is a 1-neuron layer connected to a linear activation layer.

[0114] Step 5: Output the optimal parameter combination corresponding to the reinforcement learning agent after training, as the parameters of the micro-traffic simulation model.

[0115] In one example, the simulation parameters are set to default values, a microscopic traffic simulation is started, and the corresponding simulation results are output; the default maximum acceleration is... The vehicle's desired speed is set based on real-world data. Safe headway is Comfort deceleration is The intention to change lanes is The minimum safety clearance is .

[0116] A two-layer reinforcement learning model is trained using the Proximal Policy Optimization (PPO) algorithm to automatically adjust simulation parameters. The moving average of the reward function values ​​over the most recent 30 training epochs is calculated. If the absolute value of the change in the moving average is less than 2 in 10 consecutive evaluations, then the training is considered to have converged, the reward function no longer shows significant improvement, the optimization process can be terminated, and the strategy obtained at this time is the optimal parameter calibration strategy.

[0117] To prevent excessively long training sessions in certain scenarios where convergence is difficult, a maximum training epoch of 500 is set. When the training epoch reaches 500, regardless of whether convergence has occurred, training is forcibly terminated, and the strategy with the highest reward value in the training history is selected as the final calibration result.

[0118] To verify the effectiveness of the reinforcement learning agent based on a two-layer decision-making framework proposed in this invention, the following comparative experiments are conducted. The comparative algorithm employs a single-layer proximal policy optimization reinforcement learning algorithm. This method does not consider optimizing car-following or lane-changing parameters; instead, it simultaneously optimizes six parameters—maximum acceleration, desired vehicle speed, desired headway, comfortable deceleration, number of lane changes, and lateral velocity during lane changes—after inputting the state space. Figure 4 The diagram shows a comparison of reward curves. The convergence process of this invention (blue curve) is faster than the baseline (green curve, representing a single-layer proximal policy optimization reinforcement learning algorithm), and the reward curve fluctuates less in the later stages of training, indicating that its policy is more stable and reliable. This is due to the clear planning of long-term goals by the upper-layer modules, reducing the blind exploration of the lower-layer modules.

[0119] like Figure 5 The figure shows the comparison results of reward differences. It can be seen from the figure that the reward value of the present invention is lower than that of the comparison algorithm only in the initial exploration stage of the experiment. In the rest of the time, the reward value is better than that of the comparison algorithm. The average reward value difference is 5.09 in 69 rounds.

Claims

1. A method for intelligent parameter calibration of a microscopic traffic simulation model in a highway weaving zone, characterized in that, Includes the following steps: Step 1: Obtain road geometry data, traffic flow data, and vehicle trajectory data for the target road segment; Step 2: Based on the acquired data, establish a basic micro-traffic simulation model that includes at least one car-following behavior model and one lane-changing behavior model. Step 3: Construct a reinforcement learning agent for parameter calibration of the microscopic traffic simulation basic model. The reinforcement learning agent adopts a two-layer decision framework, including: an upper-layer decision module, which dynamically selects the type of behavior model to be optimized based on the comparison between the simulation output and the real data; and a lower-layer execution module, which outputs the adjustment amount of one or more specific parameters of the selected type of behavior model based on the selection of the upper-layer decision module. Step 4: Train the reinforcement learning agent using the proximal policy optimization algorithm; Step 5: Output the optimal parameter combination corresponding to the reinforcement learning agent after training, as the parameters of the micro-traffic simulation model.

2. The intelligent parameter calibration method for the microscopic traffic simulation model of the highway weaving area according to claim 1, characterized in that, The steps for constructing a reinforcement learning agent for parameter calibration of microscopic traffic simulation models include: Define the state space as ,in, For state space, For traffic volume, The average speed of the vehicle. The average acceleration of the vehicle. The number of times a vehicle changes lanes. This refers to the lateral speed during lane changes; Define the action space as ,in, This is the maximum acceleration adjustment value. The desired vehicle speed adjustment value. For safe headway adjustment value, For comfort deceleration adjustment value, Adjustment value for lane-changing intention. This is the minimum safety clearance adjustment value; Define the reward function as follows: ,in, As a reward value, This is the upper-level reward value. This is the reward value for the lower level. To collaboratively optimize reward values, The penalty value. This is the upper-level reward value coefficient. This is the coefficient for the lower-level reward value. To collaboratively optimize the reward value coefficient, This is the penalty value coefficient.

3. The intelligent parameter calibration method for the microscopic traffic simulation model of the highway weaving area according to claim 2, characterized in that, Step 4 includes: Step 41: Construct a two-layer actor network, consisting of an upper-layer actor network and a lower-layer actor network. The input to the upper-layer actor network is the state space. The output is the action probability distribution. , where parameters 0 represents optimized car-following parameters, and 1 represents optimized lane-changing parameters; For the lower-level actor network, its input is the state space. and parameters of the upper-level actions The output is a continuous action vector. mean and standard deviation ,Right now ,in, It is the identity matrix; Construct a network of critics, with the state space as its input. The output is a value estimate for that state. , used to evaluate the long-term expected return of a state; Step 42: In each round of training, update the parameters by interacting with the environment. ,in, For the first The state space at the next iteration. For the first Action space during the next iteration For the first The reward function value at the next iteration; The advantage function is calculated using generalized advantage estimation. The calculation formula is: ， ， In the formula, As a discount factor, For parameters to estimate the generalized dominance; Step 43, construct the actor loss function, expressed as: ， In the formula, Indicates time step Expectations This represents all trainable parameters in the two-layer actor network; This represents a clipping function that restricts input values ​​to a range. Inside, This represents the probability ratio between the old and new strategies. These are the trimming parameters; Construct the critic loss function, expressed as: ， In the formula, This indicates that the critics' network is in response to the state. The value estimate, For target value; The total loss function is calculated using the following formula: ， In the formula, For the value function loss weights, For entropy reward weight, The entropy of the policy distribution; Step 44: Determine if the current iteration count has reached the maximum iteration count. ,like If the iteration stops, the strategy with the highest reward value in the training history is selected as the final calibration result. like Calculate the nearest The moving average of the reward function values ​​over the training epochs If the moving average is in continuous In this assessment, the absolute value of its change is less than If the training is considered to have converged, the iteration stops. The policy obtained at this point is the optimal parameter calibration policy. Otherwise, the iteration count is updated. Then return to step 41 for the next iteration.

4. The intelligent parameter calibration method for the microscopic traffic simulation model of the highway weaving area according to claim 3, characterized in that, The process of obtaining the action probability distribution in the upper-level actor network includes: state space After two layers of nonlinear transformation, a two-dimensional probability distribution is output through the Softmax function, as follows: ， ， ， ， In the formula, This represents the first hidden layer after the first layer is activated; These are the weighting parameters for the first layer of nonlinear transformation; These are the bias parameters for the first layer of nonlinear transformation; This represents the second hidden layer after the second activated layer; These are the weighting parameters for the second-level nonlinear transformation; These are the bias parameters for the second-level nonlinear transformation; The raw scores for the two actions; These are the weight parameters for the output layer; These are the bias parameters for the output layer; This is the activation function.

5. The intelligent parameter calibration method for a microscopic traffic simulation model of a highway weaving area according to any one of claims 1 to 4, characterized in that, The basic model of microscopic traffic simulation is built on microscopic traffic simulation software. The car-following behavior model is the intelligent driver model, and the lane-changing behavior model is the SL2015 model.

Images