CN103942398A - Traffic simulation correction method based on genetic algorithm and generalized recurrent nerve network - Google Patents

Abstract

The invention discloses a traffic simulation correction method based on a genetic algorithm and a generalized recurrent nerve network. The method includes the following steps that firstly, average travel time of vehicles is selected and serves as an evaluation index, and a target of parameter correction is determined; secondly, need traffic data are collected so that a traffic simulation model can be established, needed parameters to be corrected and a corresponding value range are determined; thirdly, optimizing computing is performed on values of the determined parameters to the corrected with the genetic algorithm, and a combination of the corrected parameters after the iteration of the genetic algorithm is predicated with the generalized recurrent nerve network, when the combination of the corrected parameters after the iteration of the genetic algorithm is matched with the target of parameter correction, the corresponding combination of the corrected parameters is output, and if the combination of the corrected parameters after the iteration of the genetic algorithm is not matched with the target of parameter correction, iteration is conducted continuously with the genetic algorithm until the combination of the corrected parameters after the interaction is matched with the target of parameter correction after the test of the generalized recurrent nerve network. According to the traffic simulation correction method based on the genetic algorithm and the generalized recurrent nerve network, the high efficiency of parameter calibration correction is achieved, the accuracy of parameter correction is ensured, the application range is wide, and the method is safe and reliable.

Description

Traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks

Technical field

The present invention relates to a kind of parameter correcting method, especially a kind of traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks, belongs to the technical field of traffic simulation parameter correction.

Background technology

Microscopic traffic simulation software not only can be used for visual simulation and reproduce road traffic traffic status, and can also traffic circulation situation be analyzed accurately and be evaluated, and has therefore obtained application more and more widely.In simulation model of microscopic, use a large amount of independent parameters to represent traffic flow and driver driving behavior, accuracy and the reliability of the value of these parameters to simulation result has decisive role.And the default value that model itself carries depends on the traffic flow situation of model development state and driver's psychological characteristic to a great extent, thus before using the simulation software of external exploitation, first proofread and correct these parameters, thus the precision of raising realistic model.

In the algorithm research that relevant parameters was proofreaied and correct in the past, mainly taking application genetic algorithm, simulated annealing as main, and these algorithms need repeatedly to move simulation software in parameter correction process.Taking genetic algorithm as example, genetic algorithm is the global search algorithm of simulating nature circle survival of the fittest process, can only rely on the information of fitness function to realize overall searching process, it is current simulation parameter correct application correcting algorithm the most widely, but the iteration that application genetic algorithm need to be carried out about 20-30 wheel just can obtain the result of final convergence, simultaneously each is taken turns iteration and all will rely on operation simulation software to obtain corresponding Output rusults, and this process can expend a large amount of time.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks is provided, it is easy to operate, realize the high efficiency that parameter calibration is proofreaied and correct, ensure the accuracy of parameter correction, wide accommodation, safe and reliable.

According to technical scheme provided by the invention, a kind of traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks, described traffic simulation bearing calibration comprises the steps:

A, choose vehicle average travel time as evaluation index, and the target of definite parameter correction;

B, gather required traffic data, to set up Traffic Flow Simulation Models; According to the Traffic Flow Simulation Models of described foundation, determine required parameter to be corrected and corresponding span;

C, utilize genetic algorithm to carry out correction calculation to above-mentioned definite parameter to be corrected, and the correction parameter obtaining through genetic algorithm iteration is combined to the input value as generalized regression nerve networks, to utilize generalized regression nerve networks to predict the output valve of VISSIM, then whether meet parameter correction target according to the correction parameter combination of the parameter correction object judgement genetic algorithm iteration in step a; In the time predicting the output valve matching parameter correction target of VISSIM, correction parameter combination after output genetic algorithm iteration, otherwise, utilize genetic algorithm to proceed iteration, until the VISSIM output valve matching parameter correction target that the correction parameter combination after iteration obtains after generalized regression nerve networks prediction.

In described step a, parameter correction target is to differ minimum through genetic algorithm iteration and the journey time of being exported by Traffic Flow Simulation Models and the actual formation time recording.

Described parameter correction target adopts evaluates average relative error, and described evaluation average relative error is:

$\mathrm{MARE}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}|\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, it is the average travel time of Traffic Flow Simulation Models output in i time period.

In described step b, the traffic data of collection comprises basic traffic data and evaluation index data; Described basic traffic data comprises road section information and signal lamp timing information; Evaluation index data comprise the average travel time of vehicle, average overall travel speed and the crossing queue length of vehicle.

In described step b, utilize X-Y scatter diagram and one-way analysis of variance to obtain parameter to be corrected and corresponding span, described parameter to be corrected comprise the visual vehicle number in front, maximum forward sight distance, average stopping distance, safe distance extention and safe distance times fractional part.

In described step c, while utilizing genetic algorithm to carry out correction calculation, adopt and proofread and correct average relative error as fitness function, described correction average relative error is:

$\mathrm{Fitness}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}\right)}^2}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, it is the average travel time of Traffic Flow Simulation Models output in i time period.

In described step c, while utilizing generalized regression nerve networks to test, the transport function of generalized regression nerve networks adopts Gaussian function.

Advantage of the present invention: utilize genetic algorithm owing to thering is good search capability the main algorithm as parameter correction, simultaneously by utilizing generalized regression nerve networks to predict the journey time of simulation software output, can effectively reduce like this number of times of simulation software operation, and then the time of saving parameter correction, easy to operate, wide accommodation, safe and reliable.

Brief description of the drawings

Fig. 1 is process flow diagram of the present invention.

Fig. 2 is the process flow diagram that the present invention utilizes genetic algorithm to proofread and correct.

Fig. 3 is the schematic diagram of generalized regression nerve networks of the present invention.

Embodiment

Below in conjunction with concrete drawings and Examples, the invention will be further described.

As shown in Figure 1: for the high efficiency that existing parameter calibration is proofreaied and correct, ensure the accuracy of parameter correction, traffic simulation bearing calibration of the present invention comprises the steps:

A, choose vehicle average travel time as evaluation index, and the target of definite parameter correction;

In the embodiment of the present invention, parameter correction target is to differ minimum through genetic algorithm iteration and the journey time of being exported by Traffic Flow Simulation Models and the actual journey time recording.Described parameter correction target adopts evaluates average relative error, and described evaluation average relative error is:

$\mathrm{MARE}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}|\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, be the average travel time of Traffic Flow Simulation Models output in i time period, MARE is for evaluating relative error.

Particularly, adopt VISSIM realistic model, the evaluation index of proofreading and correct at VISSIM simulation parameter is chosen and will be considered following several factor:

1), these data will be convenient to actual obtaining, and can ensure to a certain extent the degree of accuracy of data;

2), in VISSIM, obtain the mode of these data identical with the mode of obtaining on the spot;

3), evaluation index can change along with the change of simulation parameter, and have certain susceptibility.Amid all these factors, in the embodiment of the present invention, the evaluation index of choosing is the average travel time of vehicle.

The target of parameter correction is that the result and the actual traffic stream situation that ensure to the full extent simulation software output match.Correction target of the present invention is that journey time and the actual journey time recording of simulation software output differs minimum.

B, gather required traffic data, to set up Traffic Flow Simulation Models; According to the Traffic Flow Simulation Models of described foundation, determine required parameter to be corrected and corresponding span;

Particularly, the traffic data of collection comprises basic traffic data and evaluation index data; Described basic traffic data comprises road section information and signal lamp timing information; Evaluation index data comprise the average travel time of vehicle, average overall travel speed and the crossing queue length of vehicle.

Utilize X-Y scatter diagram and one-way analysis of variance to obtain parameter to be corrected and corresponding span, described parameter to be corrected comprise the visual vehicle number in front, maximum forward sight distance, average stopping distance, safe distance extention and safe distance times fractional part.In the embodiment of the present invention, choosing of parameter to be corrected can and need to be selected according to actual road conditions.

In the embodiment of the present invention, needed data type can be divided into two classes, and a class is basic traffic data, comprises section essential information, signal lamp timing etc., and these data are deterministic data, can directly manually be input in VISSIM; Another kind of data are the indexs for model accuracy is evaluated, comprise the average travel time of vehicle, average overall travel speed, the crossing queue length etc. of vehicle, these indexs are used for evaluating the result of actual acquisition and the goodness of fit of simulation software Output rusults, thereby judge whether to proofread and correct model.Foundation by the above-mentioned data that collect for VISSIM realistic model, can carry out the work of traffic simulation parameter correction according to this model afterwards.

Before parameter correction, carry out sensitivity analysis to the several parameters in car-following model.Its object has two, the first, determines for evaluation index and has the parameter of appreciable impact, and it two is spans of determining this parameter.The present invention utilizes the method for X-Y scatter diagram and one-way analysis of variance, thereby determines parameter to be corrected and corresponding span.

C, utilize genetic algorithm to carry out correction calculation to above-mentioned definite parameter to be corrected, and the correction parameter obtaining through genetic algorithm iteration is combined to the input value as generalized regression nerve networks, to utilize generalized regression nerve networks to predict the output valve of VISSIM, then whether meet parameter correction target according to the correction parameter combination of the parameter correction object judgement genetic algorithm iteration in step (a); In the time predicting the output valve matching parameter correction target of VISSIM, correction parameter combination after output genetic algorithm iteration, otherwise, utilize genetic algorithm to proceed iteration, until the VISSIM output valve matching parameter correction target that the correction parameter combination after iteration obtains after generalized regression nerve networks prediction.

In the embodiment of the present invention, the output valve matching parameter correction target of prediction VISSIM refers to that the output valve of prediction VISSIM, compared with parameter correction target, meets the threshold value of setting, so the threshold value of setting can be set according to actual needs.

While utilizing genetic algorithm to carry out correction calculation, adopt and proofread and correct average relative error as fitness function, described correction average relative error is:

$\mathrm{Fitness}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}\right)}^2}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, it is the average travel time of Traffic Flow Simulation Models output in i time period.

While utilizing generalized regression nerve networks to predict, the transport function of generalized regression nerve networks adopts Gaussian function.

As shown in Figures 2 and 3, the present invention adopts the algorithm of genetic algorithm as parameter correction, and auxiliary generalized regression nerve networks is used for the Output rusults of simulation software to predict, avoid directly utilizing the realistic model of simulation software to carry out computing, thereby can avoid in genetic algorithm iterative process, needing repeatedly to move the simulation software drawback of consumption plenty of time, realize the high efficiency of parameter calibration.

About genetic algorithm

Genetic algorithm is current simulation parameter correct application correcting algorithm the most widely.Concrete genetic algorithm optimization step is as follows:

1), the generation of initial population:

The evolutionary generation counter t of genetic algorithm is made as to 0, maximum evolutionary generation is set simultaneously; And preset the size of population, generate at random initial parameter value by program.

2), coding: the parameter of above-mentioned random generation is carried out to binary coding, and corresponding genetic fragment composition for each parameter, is placed on genetic fragment corresponding all parameters in item chromosome and operates.

3), the design of fitness function:

The present invention adopts the fitness function of average relative error as genetic algorithm, and concrete formula is shown in shown in formula (2):

$\mathrm{Fitness}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}\right)}^2}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, it is the average travel time of Traffic Flow Simulation Models output in i time period.

4), the computing of genetic algorithm:

Select operation: the present invention is based on roulette algorithm and carry out the selection operation of genetic algorithm, determine that according to the size of fitness individuality enters follow-on probability, fitness is higher, be selected that to enter follow-on probability larger.

Interlace operation: the present invention selects single-point to intersect and realizes chromosomal interlace operation.

Mutation operation: the present invention selects the algorithm of basic single-point variation.

5), the judgement of end condition:

When the number of times of iteration exceedes predefined maximum algebraically, genetic algorithm is with regard to termination of iterations.

6) chromosomal decoding; The chromosome obtaining in said process is decoded, can obtain final parameter combinations.

About generalized regression nerve networks

Generalized regression nerve networks is a kind of three layers of static feedforward network, and it is made up of input layer, mode layer, summation layer and output layer.

In input layer, neuronic number is identical with the dimension of training sample input vector, and each neuron is simple linear unit, directly input variable is passed to mode layer;

Mode layer neuron number equals the number of training sample, the corresponding training sample collection of each neuron, and its transport function often adopts Gaussian function, and described Gaussian function is:

$p_i=\exp [-\frac{{(x-x_i)}^T(x-x_i)}{2{\mathrm{\σ}}^2}],i=\mathrm{1,2},\·\·\·,n$

Wherein, x is input variable, x _ifor the learning sample corresponding with input variable, σ is the spread factor of Gaussian function, is referred to as the smooth factor here.T represents to turn order computing, and in the embodiment of the present invention, input variable is process genetic algorithm iteration post-equalization parameter combinations, and n is the quantity of correction parameter combination.

Summation layer carries out read group total by two class neurons altogether and forms, and a class is that arithmetic summation is carried out in neuronic output to mode layer, and computing formula is another kind of is that neuronic output is weighted summation to mode layer, and computing formula is wherein Y _ifor sample observations, be the average travel time of prediction output.

The neuron number of output layer equals the dimension of output vector in learning sample, is that two class results that above-mentioned summation layer is obtained are divided by and are obtained.

Generalized regression nerve networks is applicable to the research of nonlinear problem, the journey time of vehicle is subject to the combined influence of different parameters, its relation be difficult to one accurately mathematical formulae express, therefore use generalized regression nerve networks can more accurately describe the Nonlinear Mapping relation of simulation parameter and journey time; Generalized regression nerve networks does not need to preset the form of model, do not need artificially to set number, the connection weights etc. of hidden layer neuron, and only need to determine smooth factor sigma, can reduce to a great extent like this because the impact of artificial factor on model Output rusults.

Generalized regression nerve networks training process is:

1), data are normalized;

2), data set is divided into two groups afterwards, be respectively training data set verification msg collection, training dataset is used for neural network to train, and verification msg collection is used for the neural network training to test;

3), the smooth factor of setting network.After training data is determined, the structure of generalized regression nerve networks is determined thereupon, is therefore in fact the training process to smooth factor sigma to the training of network.The present invention finds optimum training, checking collection and smooth factor sigma by the method for K-fold cross validation.

In the embodiment of the present invention, utilize optimum training set and the smooth factor that said process obtains to set up general regression neural network, this network can be used for the every parameter combinations producing after iteration of taking turns of genetic algorithm to test, input by these parameter values as neural network, through the propagation of neural network, finally obtain journey time as output valve, utilize this output valve to calculate this and take turns the iteration fitness value of genetic algorithm afterwards, thereby judge whether to carry out the iteration of next round.

The present invention utilize genetic algorithm owing to thering is good search capability the main algorithm as parameter correction, simultaneously by utilizing generalized regression nerve networks to carry out simulation and forecast to the journey time of simulation software output, can effectively reduce like this number of times of simulation software operation, and then the time of saving parameter correction, easy to operate, wide accommodation, safe and reliable.

Claims (7)

1. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks, is characterized in that, described traffic simulation bearing calibration comprises the steps:

(a), the average travel time of choosing vehicle is as evaluation index, and the target of definite parameter correction;

(b), gather required traffic data, to set up Traffic Flow Simulation Models; According to the Traffic Flow Simulation Models of described foundation, determine required parameter to be corrected and corresponding span;

(c), utilize genetic algorithm to carry out correction calculation to above-mentioned definite parameter to be corrected, and the correction parameter obtaining through genetic algorithm iteration is combined to the input value as generalized regression nerve networks, to utilize generalized regression nerve networks to predict the output valve of VISSIM, then whether meet parameter correction target according to the correction parameter combination of the parameter correction object judgement genetic algorithm iteration in step (a); In the time predicting the output valve matching parameter correction target of VISSIM, correction parameter combination after output genetic algorithm iteration, otherwise, utilize genetic algorithm to proceed iteration, until the VISSIM output valve matching parameter correction target that the correction parameter combination after iteration obtains after generalized regression nerve networks prediction.

2. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks according to claim 1, it is characterized in that: in described step (a), parameter correction target is to differ minimum through genetic algorithm iteration and the journey time of being exported by Traffic Flow Simulation Models and the actual formation time recording.

3. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks according to claim 2, is characterized in that: described parameter correction target adopts evaluates average relative error, and described evaluation average relative error is:

$\mathrm{MARE}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}|\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, it is the average travel time of Traffic Flow Simulation Models output in i time period.

4. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks according to claim 1, is characterized in that: in described step (b), the traffic data of collection comprises basic traffic data and evaluation index data; Described basic traffic data comprises road section information and signal lamp timing information; Evaluation index data comprise the average travel time of vehicle, average overall travel speed and the crossing queue length of vehicle.

5. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks according to claim 1, it is characterized in that: in described step (b), utilize X-Y scatter diagram and one-way analysis of variance to obtain parameter to be corrected and corresponding span, described parameter to be corrected comprise the visual vehicle number in front, maximum forward sight distance, average stopping distance, safe distance extention and safe distance times fractional part.

6. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks according to claim 1, it is characterized in that: in described step (c), while utilizing genetic algorithm to carry out correction calculation, adopt and proofread and correct average relative error as fitness function, described correction average relative error is:

$\mathrm{Fitness}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{t_i^{\mathrm{field}}-t_i^{\mathrm{sim}}}{t_i^{\mathrm{field}}}\right)}^2}{N}$

Wherein, N is the time period quantity gathering, be the actual vehicle average travel time recording in i time period, it is the average travel time of Traffic Flow Simulation Models output in i time period.

7. the traffic simulation bearing calibration based on genetic algorithm and generalized regression nerve networks according to claim 1, it is characterized in that: in described step (c), while utilizing generalized regression nerve networks to predict, the transport function of generalized regression nerve networks adopts Gaussian function.