CN106373390A - Road traffic state evaluation method based on adaptive neuro fuzzy inference system - Google Patents

Abstract

The invention relates to a road traffic state evaluation method based on an adaptive neuro fuzzy inference system, which comprises a training stage and a testing stage. In the training stage, historical data for input variables and output variables are read; the historical data are inputted to the adaptive neuro fuzzy inference system, and an output result of the adaptive neuro fuzzy inference system is obtained; based on a parameter learning rule for a back propagation thought, the output result of the adaptive neuro fuzzy inference system is combined to adjust the parameter, and an adaptive neuro fuzzy inference system after parameter adjustment is obtained; and the minimum mean square error of the adaptive neuro fuzzy inference system after parameter adjustment is calculated, whether the value achieves a specific threshold is judged, and if yes, the training stage is ended. In the testing stage, testing data are read; and the testing data are inputted to the adaptive neuro fuzzy inference system obtained through training in the training stage, and a service level value for describing the road traffic state is obtained.

Description

Traffic state evaluation method based on Adaptive Neuro-fuzzy Inference

Technical field

The present invention relates to road monitoring field, particularly to a kind of road based on Adaptive Neuro-fuzzy Inference Traffic behavior evaluation methodology.

Background technology

With the popularization of expanding economy and automobile, traffic congestion has become as the difficult problem that big and medium-sized cities generally face. The traffic behavior of road is carried out with evaluation is to alleviate congested in traffic, and vehicle accident is made with the premise of quick response.

In prior art, the achievement in research major part with regard to traffic system operation conditions evaluation is to be both for public affairs at a high speed The research on road, in terms of urban road, achievement in research is relatively fewer.HCM is in traffic system Can evaluate that aspect is more authoritative, this handbook is directed to commenting of urban highway traffic situation (including crossing and major trunk roads) Valency carried out discussion, but the acquisition methods for wherein performance indications are then to rely on formula estimation rather than field survey, HCM is the appraising model to set up performance indications for the perfect condition according to American roads simultaneously, and Be not suitable for China's national situation.Yuan Jing is self-important, Yuan Zhenzhou list of references 1 " " Methods for Traffic Capacity at Signal Junction computational methods Comparative analysiss ", highway communication technology, 2006 (5), 23-29 " for China road conditions, to existing evaluation Model is modified, and the actual effect of improved model, accuracy, objectivity and application all need to be proved further. Jungkeun yoon and brian noble is in list of references 2 " " surface street traffic estimation ", in proceedings of the 5th annual acm/usenix conference on mobile systems, Applications, and services (mobisys) .san juan, pr.june 2007.p.220 232 " it is directed to city Road traffic condition proposes a kind of new evaluation methodology, using the vehicle being provided with two-way communications capabilities gps equipment The traffic related information (mainly velocity information) obtaining urban road evaluating the traffic of certain road, but It is the traffic of single crossing and city road network cannot to be evaluated.Zhao Ming, Hou Zhongsheng are in list of references 3 " " the signalized intersections traffic condition assessment of data-driven ", in proceedings of the 27^thchinese control Conference, july 16-18,1008, kunming, china, p.559 563 " from transportation system management person with make in The angle of user is set out it is proposed that two performance index system average speeds to be commented with service level index The traffic of valency crossing, but from the angle of manager and user can not sufficiently explain and utilize this two Individual performance indications, and simply crossing is evaluated when calculating service level index, not clearly It is proposed for the evaluation methodology of road network.Additionally, Shen Yinan is in its master thesis " urban highway traffic evaluation side For the transportation network that is made up of a large amount of crossings in method and evaluation system exploitation ", for point, line, surface etc. three Aspect carries out the performance indications evaluation of urban highway traffic situation respectively, is that the overall merit of traffic network provides one Plant new thinking.

Prior art has a disadvantage in that in City road traffic system, as one kind qualitatively evaluating method, Service level (los) describe the overall operation conditions of certain road traffic system and driver and passenger to its Impression, is one important evaluation index of City road traffic system.The various los evaluation methodologys having existed, It is that quantitative assessment is carried out based on single index mostly, by contrasting the relation between this index and los grade, to traffic The available grade of service of infrastructure is evaluated.These methods do not account for the qualitative side of this evaluation problem itself Face, can not fully demonstrate subjectivity and the globality of traffic system evaluation, set up the comprehensive evaluation model of multi objective It is the important topic that road traffic los evaluates.

Content of the invention

It is an object of the invention to overcome existing urban road traffic state qualitative evaluating method order of accuarcy limited Defect, thus providing a kind of traffic state evaluation method that accuracy is high, test error value is low.

To achieve these goals, the invention provides a kind of road based on Adaptive Neuro-fuzzy Inference is handed over Logical method for evaluating state, including training stage and test phase；Wherein,

The described training stage comprises the following steps:

Step 101), read the historical data of input variable and output variable；Described input variable includes flow, accounts for There are rate and speed, described output variable includes service level value；

Step 102), by step 101) historical data that inputted is input in Adaptive Neuro-fuzzy Inference, Obtain the output result of Adaptive Neuro-fuzzy Inference；Parameter learning rule based on anti-pass thought, in conjunction with institute The output result stating Adaptive Neuro-fuzzy Inference is adjusted to the parameter of Adaptive Neuro-fuzzy Inference Whole, obtain the Adaptive Neuro-fuzzy Inference after parameter adjustment；Wherein,

The parameter of described Adaptive Neuro-fuzzy Inference includes premise parameter, consequent parameter, sets for input variable Fixed weight parameter, the weight parameter for rule settings；Described premise parameter is by the membership function shape of reasoning former piece Shape determines；Described consequent parameter is determined by the membership function shape of reasoning consequent；

Step 103), calculation procedure 102) obtained by parameter adjustment after Adaptive Neuro-fuzzy Inference Minimum Mean Square Error, judges whether this value reaches a specified threshold, if reached it is meant that the training stage terminates；No Then, execute next step；

Step 104), read one group of new historical data with regard to input variable and output variable, then re-execute Step 102)；

Described test phase comprises the following steps:

Step 201), read test data, described test data includes: the history number of flow, occupation rate and speed According to；

Step 202), test data is input to via the Adaptive Neural-fuzzy Inference system that obtains of training stage training In system, obtain the value of the service level for describing road traffic state.

In technique scheme, described Adaptive Neuro-fuzzy Inference includes five layers, is respectively as follows:

Ground floor: obscuring layer, for by accurate fuzzy inputing method；

o_1,i=μ_ai(x), i=1,2 or

$o_{1,i}={\mathrm{\μ}}_{b_{i-2}}\left(y\right),i=3,4$

Wherein, o_j,iRepresent i-th node output of jth layer, μ_ai、For the membership function of former piece, this two Individual former piece membership function is defined with generalized bell mf:

${\mathrm{\μ}}_a=\frac{1}{1+{\[{\left(\left(x-c_i\right)/a_i\right)}^2\]}^{b_i}}$ Or ${\mathrm{\μ}}_b=\frac{1}{1+{\[{\left(\left(x-c_i\right)/a_i\right)}^2\]}^{b_i}}$

A therein_i、b_i、c_iPremised on parameter；

The second layer, reasoning layer, for calculating the excitation density w of every rule_i；

$o_{2,i}=w_i=\[\stackrel{\&overbar;}{{\mathrm{\μ}}_1}+{\mathrm{\μ}}_1\×{\mathrm{\μ}}_{ai}\left(x\right)\]\×\[\stackrel{\&overbar;}{{\mathrm{\μ}}_2}+{\mathrm{\μ}}_2\×{\mathrm{\μ}}_{bi}\left(y\right)\]$

Wherein, i=1,2, μ_iRepresent the weight of each input,

Third layer, hidden layer, for calculating effective consequent mf of every rule；

o_3,i=w_iοc_i

Wherein, i=1,2；c_iReasoning consequent；" ο " represents implicit operator；

4th layer, polymer layer, for calculating the summation of strictly all rules effective consequent mf；

o₄=∑ (w_iοc_i-w_i-1οc_i-1)×τ_i

Wherein, i=1,2；τ_iRepresent the weight of every rule；Aggregation Operator ∑ (w_iοc_i-w_i-1οc_i-1)×τ_iUsing Choquet integrates；

Layer 5, mould from paste layer, for the accurate output of computing system；

o₅=f=d ο o₄

Wherein, d represents deblurring operator, and it calculates is realized using center de-fuzzy method.

In technique scheme, the parameter of Adaptive Neuro-fuzzy Inference is adjusted with reference to equation below:

δw_ji=η (d_i-x_i)·x_j·x

Wherein, δ w_ijRepresent the increment of arbitrary parameter in Adaptive Neuro-fuzzy Inference；η is Learning Step, d_i It is the desired output of node i, x_iIt is the reality output of node i, xj is the input of node i, node i is node j Last layer node, j < i；X is a multinomial, uses (x_i×(1-x_i)) expression.

In technique scheme, described Adaptive Neuro-fuzzy Inference is replaced with owa operator in reasoning layer And operator or or operator, calculate excitation density；Integrated with choquet to realize being polymerized in polymer layer；Mould from Gelatinizing operator adopts center de-fuzzy method；And set the weight of each input, use μ_iRepresent；Set every rules and regulations Weight then, uses τ_iRepresent.

It is an advantage of the current invention that:

Accuracy is high compared with the conventional method for the traffic state evaluation method of the present invention, and test error value is low.

Brief description

Fig. 1 is the reasoning process schematic diagram of the fuzzy inference system integrating-owa based on choquet；

Fig. 2 is the schematic diagram of the example of an Adaptive Neuro-fuzzy Inference；

Fig. 3 is the schematic diagram of the service level evaluation model based on agg-anfis that the present invention is set up；

Fig. 4 is the schematic diagram of sugeno-fis output curve diagram；

Fig. 5 is the error schematic diagram between the reality output of sugeno-fis fuzzy inference system and desired output；

Fig. 6 is the training data error curve diagram of anfis；

Fig. 7 be detection data desired output and training after sugeno-fis real output value between comparison diagram；

Fig. 8 (a) represents the membership function of the flow in input variable；

Fig. 8 (b) represents the membership function of the occupation rate in input flow rate；

Fig. 8 (c) is the membership function of the speed in input variable；

Fig. 8 (d) represents the membership function of output variable los；

Fig. 9 is aggfis output curve diagram；

Figure 10 is the error curve diagram of aggfis reality output and desired output；

Figure 11 is the training error curve chart of agg-anfis；

Figure 12 is the output valve of aggfis and detection data desired output comparison diagram after training；

Figure 13 is the test error curve chart of aggfis；

Figure 14 (a) represents the membership function of the flow in input variable；

Figure 14 (b) represents the membership function of the occupation rate in input flow rate；

Figure 14 (c) is the membership function of the speed in input variable；

Figure 14 (d) represents the membership function of output variable los.

Specific embodiment

Due to being related to more concept in the present invention, therefore first an explanation is done to these concepts.

Traffic flow: refer to pass through the friendship in a certain place of road, a certain section or a certain track in seclected time section Logical entity number；

Roadway occupancy: on road, current traffic flow accounts for the ratio of design traffic volume, also shouts road utilization rate.

Fuzzy inference system: fuzzy reasoning refers to draw possible inaccurate conclusion in never accurate premise set Reasoning process, also known as approximate resoning.Fuzzy inference system is with Fuzzy Set Theory and fuzzy reasoning method etc. as base Plinth, has the system processing fuzzy message ability.This system is with fuzzy logic theory for main calculating instrument, permissible Realize complicated nonlinear mapping relation, its input and output is all accurate numerical value.

Neutral net: neutral net is a kind of operational model, by mutual between substantial amounts of node (or claiming neuron) Connection is constituted.A kind of specific output function of each node on behalf, referred to as excitation function (activation function). Connection between each two node all represents one for the weighted value by this connection signal, referred to as weight.Nerve The output of network is then different according to the difference of the connected mode of network, weighted value and excitation function.

Adaptive Neuro-fuzzy Inference: Adaptive Neuro-fuzzy Inference is fuzzy inference system and nerve net The combination of network.By the study mechanism of neutral net is introduced fuzzy inference system, one will be constituted and carry mankind's sense Feel the Adaptable System with cognitive component.Neutral net is directly embedded among a structure all obscuring, and it will Unconsciously learning to training data, automatically generating, revise and high level overview is gone out optimal input and become with output The membership function of amount and fuzzy rule；And on the other hand, each Rotating fields and the parameter of neutral net are also provided with Clearly, understandable physical significance.

It is more than the description to concept involved in the inventive method, below the method for the present invention is described further.

The traffic state evaluation method of the present invention with can be detected on road speed, flow, occupation rate make For the input value of Adaptive Neuro-fuzzy Inference, after processing through Adaptive Neuro-fuzzy Inference, output is handed over Logical service level value (los).

In order to make it easy to understand, doing in detail to involved Adaptive Neuro-fuzzy Inference in the inventive method first Explanation.

In the inventive method, involved Adaptive Neuro-fuzzy Inference (abbreviation agg-anfis) is to be based on The fuzzy inference system (abbreviation aggfis) of choquet integration-owa is with feedforward neural network according to fuzzy neural Obtained from net Basic Topological combines, hereinafter will successively aggfis, agg-anfis be illustrated.

First, the fuzzy inference system of-owa is integrated based on choquet

According to the expression-form of fuzzy reasoning type and fuzzy if-then rules, most of fuzzy inference system can To be divided into following three major types: tsukamoto fuzzy model, mamdani fuzzy model and sugeno fuzzy model.

For whole calculating processes of any one fuzzy inference system, we are necessary for following fuzzy reasoning operator Determine a function:

1) with operator (and operator): for calculating the excitation density regular by AND connection former piece, generally It is t normal form；

2) or operator (or operator): for calculating the excitation density being connected former piece rule by "or", generally It is t association normal form；

3) imply operator (implication operator): for according to given excitation density, calculating every rule The mf of effective consequent, typically t normal form；

4) Aggregation Operator (aggregate operator): for being polymerized all of effective consequent mf, thus obtaining comprehensive Close output mf, typically t assists normal form；

5) de-fuzzy operator (defuzzification operator): for total output mf is converted into accurately Output valve.

In order to solve importance factor expression problem in the Problem of Universality of fuzzy reasoning operator and reasoning process, this Shen Please in depth analysis have been carried out to all kinds of Reasoning operators.Owa operator can be understood as and operator or or operator Universality expression, the choquet integral and calculating essence in fuzzy integral be one have calculating successional polymerization Operator, this two operators can solve the Problem of Universality of fuzzy reasoning operator；Meanwhile, different in owa operator The selection of weights have expressed complicated interaction relationship between object (index)；And choquet integral computer System itself, just describes the interaction relationship between two objects (index).It can be seen that, owa operator and choquet Integration can be realized considering the fuzzy reasoning of weight.

Based on above-mentioned analysis, present applicant proposes a kind of fuzzy inference system based on choquet integration-owa, should System replaces and operator or or operator in reasoning layer (inference layer) with owa operator, calculates and swashs Encourage intensity；Integrated with choquet to realize being polymerized in polymer layer (aggregation layer), rather than traditional t Association's normal form operator (max or sum)；De-fuzzy operator adopts center de-fuzzy method (coa)；And set The weight of each input, uses μ_iRepresent；Set the weight of every rule, use τ_iRepresent；Thus realizing entirely obscuring Reasoning process.

Describe in FIG to integrate the reasoning process of the fuzzy inference system of-owa based on choquet.

The fuzzy rule form of this fuzzy inference system is as follows:

r_i: if v₁is a₁and v₂is b₁and v₃is c₁,then u is d₁；

v₁、v₂、v₃For input variable, u is single output variable；a₁、b₁、c₁Represent each input variable respectively Fuzzy set；d₁Fuzzy set for output variable.

This fuzzy inference system parameter as described in Figure 1 is defined as follows:

The membership function module (membership neural module) of d: regular former piece；

d^-1: the membership function module (inverse membership neural module) of consequent；

D (x): the excitation density of every rule；

r_i: fuzzy rule module (rule neural module)；

The weight module of m-module: polymer layer；

τ_i: the weight of every rule；

u_i: the weight of each input；

[a_i,b_i]: the threshold value of every rule effectively output.

2nd, the Adaptive Neuro-fuzzy Inference of-owa is integrated based on choquet

The Adaptive Neuro-fuzzy Inference integrating-owa based on choquet is aggfis and feedforward neural network The product being combined according to fuzznet Basic Topological.This Adaptive Neuro-fuzzy Inference merges Whole reasoning processes of aggfis, therefore system have Reasoning operator universality, and enable index importance factor Expression；Simultaneously as the learning functionality of neutral net, the whole parameters in this system can be entered according to learning rules Row is reconciled, and therefore this system has the adaptability to data.

Include 5 layers based on the Adaptive Neuro-fuzzy Inference that choquet integrates-owa, Fig. 2 is adaptive for one Answer the example of neural fuzzy inference system, the system in this embodiment has only included two inputs and single output, The input of expansible system and the number exporting in actually used；

The output result of each of which layer is expressed as follows:

Layer 1:fuzzification layer (obscuring layer)

This layer of work to be done is by accurate fuzzy inputing method:

o_1,i=μ_ai(x), i=1,2 or

$o_{1,i}={\mathrm{\&mu;}}_{b_{i-2}}\left(y\right),i=3,4---\left(1\right)$

Wherein, o_j,iRepresent i-th node output of jth layer, μ_ai,For the membership function of former piece, this two Individual former piece membership function can be defined with generalized bell mf:

${\mathrm{\&mu;}}_a=\frac{1}{1+{\&lsqb;{\left(\left(x-c_i\right)/a_i\right)}^2\&rsqb;}^{b_i}}$ Or ${\mathrm{\&mu;}}_b=\frac{1}{1+{\&lsqb;{\left(\left(x-c_i\right)/a_i\right)}^2\&rsqb;}^{b_i}}---\left(2\right)$

A therein_i、b_i、c_iPremised on parameter.

Layer 2:inference layer or rule layer (reasoning layer)

This layer of work to be done is to calculate every rule using a kind of special case (and) of owa operator Excitation density w_i.

Wherein μ_iRepresent the weight of each input.

$o_{2,i}=w_i=\&lsqb;\stackrel{\&overbar;}{{\mathrm{\&mu;}}_1}+{\mathrm{\&mu;}}_1\&times;{\mathrm{\&mu;}}_{ai}\left(x\right)\&rsqb;\&times;\&lsqb;\stackrel{\&overbar;}{{\mathrm{\&mu;}}_2}+{\mathrm{\&mu;}}_2\&times;{\mathrm{\&mu;}}_{bi}\left(y\right)\&rsqb;,i=1,2---\left(3\right)$

Wherein, $\stackrel{\&overbar;}{{\mathrm{\&mu;}}_1}=1-{\mathrm{\&mu;}}_1,\stackrel{\&overbar;}{{\mathrm{\&mu;}}_2}=1-{\mathrm{\&mu;}}_2;$

Layer 3:implication layer (hidden layer)

This layer of work to be done is the effective consequent mf calculating every rule.

o_3,i=w_iοc_iI=1,2 (4)

Consequent parameter collection is by reasoning consequent (c_i) decision of membership function shape；" ο " represents implicit operator (product).

Layer 4:aggregation layer (polymer layer)

This layer of work to be done is the summation calculating strictly all rules effective consequent mf.

Wherein, the weight of every rule is by τ_iDefinition.

o₄=∑ (w_iοc_i-w_i-1οc_i-1)×τ_i, i=1,2 (5)

Aggregation Operator ∑ (w_iοc_i-w_i-1οc_i-1)×τ_iUsing choquet integration.

Layer 5:defuzzification layer (mould from paste layer)

This layer of work to be done is the accurate output of computing system, and the result obtaining after deblurring is designated as o₅:

o₅=f=d ο o₄(6)

Wherein, d represents deblurring operator, and it calculates can be realized using center de-fuzzy method (coa).

The parameter adjusting is needed to have in this Adaptive Neuro-fuzzy Inference:

1) premise parameter collection { a_i,b_i,c_i}: determined by the membership function shape of reasoning former piece；

2) consequent parameter collection: determined by the membership function shape of reasoning consequent；

3) weight of each input: use μ_iRepresent；

4) weight of every rule: use τ_iRepresent.

3rd, the parameter learning rule based on anti-pass thought being suitable in Adaptive Neuro-fuzzy Inference

Adaptive network is the network structure that a kind of overall input-output characteristic to be determined by one group of adjustable parameter.Generally, The performance of network to be measured by the difference between desired output under identical input condition and network output.This difference quilt It is defined as error criterion.A certain specific optimisation technique is applied to an assigned error index and has just obtained learning rules. Steepest descent method is the most frequently used learning rules, if applying gradient vector, produced side in steepest descent method The parameter learning rule referred to as based on anti-pass thought for the method.

The basic thought of parameter learning rule (bp) based on anti-pass thought is exactly that the entirety first defining a system is missed Poor index, then optimize this index according to learning rules.The global error index of system can be defined as follows:

$e_p=\underset{k=1}{\overset{n\left(l\right)}{\&sigma;}}{(d_k-x_{l,k})}^2$

$e=\underset{p=1}{\overset{p}{\&sigma;}}{e}_p---\left(7\right)$

Wherein, e_pIt is the error criterion to training data for the pth, e is system global error index.d_kIt is pth group K-th component of training data desired output, x_l,kIt is to apply to be produced during pth group training data input vector to network K-th component of raw reality output.Our target is exactly to make e (e_p) minimum.

Will be as follows for the definitions for error signals of l i-th node of layer:

I.e. error criterion e_pDerivative to i-th node output, i.e. one error signal of each node correspondence.We As can be seen that e_pRefer to the error criterion of last layer of output node, (output layer is tied therefore to only have last layer of node Point) error signal can be calculated with immediate derivation, -2 (d_i-x_i)；And hidden layer node (as intermediate layer node) Error signal cannot directly calculate.The error signal of hidden layer node can be derived by the error signal of its preceding layer node Obtain.That is:

${\mathrm{\&epsiv;}}_{l,i}=\frac{\&part;e_p}{\&part;x_{l,i}}=\underset{m}{\&sigma;}{\mathrm{\&epsiv;}}_{l+1,m}.\frac{\&part;f_{l+1,m}}{\&part;x_{l,i}},$ Or

${\mathrm{\&epsiv;}}_i=\frac{\&part;e_p}{\&part;x_i}=\underset{m}{\&sigma;}{\mathrm{\&epsiv;}}_m.\frac{\&part;f_m}{\&part;x_i}---\left(8\right)$

M-th node represents the node related to last layer (l+1).

Steepest descent method is applied to minimize error criterion e_pNeed to calculate gradient vector.Gradient vector is defined as: Error criterion is with respect to the derivative of each parameter.The basic conception calculating gradient vector is: from the beginning of output layer, will lead The information of number form formula is in layer propagated backward, until it reaches input layer.

If α is l layer, the parameter of i-th node, then have:

$\frac{\&part;e_p}{\&part;a}=\frac{\&part;e_p}{\&part;x_{l,i}}*\frac{\&part;f_{l,i}}{\&part;a}={\mathrm{\&epsiv;}}_{l,i}\&centerdot;\frac{\&part;f_{l,i}}{\&part;a}---\left(9\right)$

The derivative with respect to α for global error index e is:

$\frac{\&part;e}{\&part;a}=\underset{p=1}{\overset{p}{\&sigma;}}\frac{\&part;e_p}{\&part;a}---\left(10\right)$

The more new formula of general parameter α is:Learnt along negative gradient direction.η is Learning rate.

Therefore, theoretical according to classical bp, the parameter of agg-anfis more new formula is as follows in this application:

${\mathrm{\&delta;w}}_{ij}=-\mathrm{\&eta;}\&centerdot;\frac{\&part;e_p}{\&part;w_{ij}}=-\mathrm{\&eta;}\&centerdot;\frac{\&part;e_p}{\&part;x_i}\&centerdot;\frac{\&part;f_i}{\&part;w_{ij}}=-{\mathrm{\&eta;\&epsiv;}}_i.\frac{\&part;f_i}{\&part;w_{ij}}$

$w_{next}=w_{now}-\mathrm{\&eta;}\frac{\&part;e_p}{\&part;w_{ij}}---\left(11\right)$

Wherein, δ w_ijRepresent that a certain parameter (needs the parameter updating: premise parameter collection, conclusion in agg-anfis Parameter set, each input weight, the weight of every rule) increment；Node i is the last layer node of node j, J < i, i.e. x_i=f_i(∑w_ij.x_j+θ).f_iAnd x_iRepresent excitation function and the output of node i.Error signal_i Successively propagate forward from output node, the error signal of each node can be derived by upper layer error signal and be obtained.Formula In,(ask in error signal ${\mathrm{\&epsiv;}}_j=\frac{\&part;e_p}{\&part;x_j}=\underset{i}{\&sigma;}{\mathrm{\&epsiv;}}_i.\frac{\&part;f_i}{\&part;x_j}$ ) andAll can obtain, you can According to formula, parameter is updated.

If,I.e.The parameter of each node more new formula is it is known that can be to whole network Parameter (weights) be updated.

More generally, agg-anfis parameter more new formula is as follows:

δw_ji=η (d_i-x_i)·x_j·x (12)

Wherein, η is Learning Step, d_iIt is the desired output of node i, x_iIt is the reality output of node i, x_jIt is The input of node i, x is a multinomial, typically uses (x_i×(1-x_i)) expression.

It is more than the explanation to involved Adaptive Neuro-fuzzy Inference in the inventive method, below to this certainly How adaptation neural fuzzy inference system is applied to traffic state evaluation elaborates.

As a qualitative measurement index of urban road traffic state, level of service (los) reflects department The impression to traffic for the machine, describes the operating conditions of the driver in traffic flow, such as travel speed or travelling Time, the degree of freedom of driving, traffic disturbance, comfort level or degree of convenience etc..Level of service evaluation problem essence It is a class many objects (index) decision problem, the essence of this kind of decision-making (reasoning) problem is nonlinear mapping, this The target of application is exactly to set up model to describe this nonlinear mapping relation.

First, agg-anfis has the ability of Continuous Mappings, by realizing the process of approximate resoning, can be transparent The thinking model of earth's surface intelligent's class, it is possible to achieve the decision making process of traffic behavior los；Meanwhile, determining in description people During plan (assessment), in order that the reasoning results approach our objective decision result it is necessary to parameter to model It is adjusted, the adaptive ability of agg-anfis just can solve this problem.

It is specifically based on the following consideration of some:

1) service level is the value judgment of the service quality being provided.For each grade of service level, It is all to be described with natural language, that is, consequent is expressed with fuzzy set.For example: quickly, flow is very for speed Low, then los grade is a (very good) etc..

The expression of anfis consequent is a linear function, and service level reflection is the subjectivity sense to road conditions for the driver It is subject to, this model cannot really embody the implication of service level.

Agg-anfis model can solve this problem, and in its fuzzy rule consequent, service level class is used Fuzzy set (membership function) rather than simple linear equation represent, really reflect the general of service level class Read, want the demand of solve problem closer to us.

2) for a certain basic means of transportation, multiple indexs can be had to carry out los evaluation, that is, evaluation index is not Uniquely, multiple attribute synthetical evaluation problem have to be solved.

3) saltus problem between grade.For example: 0.54 belongs to b level, and 0.55 just belongs to c level, such Grade classification is very unreasonable.Service level (los) can be carried out fuzzy classification expression by agg-anfis consequent, Traffic behavior evaluation is made more to press close to human thinking's pattern.

Therefore, the present invention selects to integrate the Adaptive Neuro-fuzzy Inference of-owa based on choquet (agg-anfis) solving los evaluation problem, service level subjectivity can be solved, multiple index evaluation, etc. Many difficult problems such as order transition.

The basic data (flow, occupation rate and speed) of experiment is all obtained by detector, and these historical datas are true Real effective, can be used for many objects (index) los and evaluate；Water is serviced to traffic behavior based on existing historical data Flat (los) carries out fuzzy classification expression, a-f.All of data is all divided into training data and test data.Pass through Mapping relations between system input and los value, the present invention sets up the service level evaluation mould based on agg-anfis Type, is trained to model by historical data, and the model after training is tested.

The schematic diagram of the service level evaluation model based on agg-anfis that Fig. 3 is set up by the present invention, by this mould The evaluation to road traffic state realized by type.In this model, x, y, z represents input variable, represents respectively Flow, occupation rate and speed.a₁-a₃Fuzzy set for flow；b₁-b₃Fuzzy set for occupation rate；c₁-c₃For The fuzzy set of speed；d₁-d₆Represent six grades of output los.F is output variable, i.e. service level (los) Value.Premise parameter number in this model has 3 × 3 × 3=27 (to include a in model₁-a₃, b₁-b₃With c₁-c₃Totally 9 membership functions, each membership function comprises 3 premise parameters, and therefore premise parameter is a total of 9*3=27), consequent parameter number has 6 × 4=24 (to include d in model₁-d₆Totally 6 membership functions, Each function has 4 consequent parameters, therefore a total of 6*4=24 of consequent parameter)；Each input variable corresponding sets one Individual weight, totally 3 parameters；The every rule of correspondence sets a weight, totally 27 parameters, and initial value is all 1. These parameters (include previously mentioned 27 premise parameters, 24 consequent parameters, for 3 of input variable setting Weight parameter, 27 weight parameter for rule settings) it is nonlinear parameter, so applying to whole model Bp algorithm carries out parameter regulation.

The fuzzy rule of this model is specifically expressed as follows form:

Rule i:if speed is slow, and occupancy is high, and volume is high, then Los=f.

Above-mentioned based on the service level evaluation model of agg-anfis on the basis of, the method for the present invention includes two In the individual stage, the first stage is the training stage, and second stage is test phase.In the work that the training stage is to be done It is using existing historical data, the service level evaluation model based on agg-anfis to be trained, that is, to model Parameter set (premise, conclusion) be adjusted, when this model reaches minimum error index, training terminates.? Test phase work to be done is the service level evaluation mould based on agg-anfis obtained by the training stage On the basis of type, test data is inputted this based in the service level evaluation model of agg-anfis, is serviced Assessment of levels result.

Below specific description is done to the step that implements in this two stages.

First, the training stage

Step 101), read existing historical data, described historical data includes: input variable and output variable Historical data；Described input variable includes flow, occupation rate and speed, and described output variable includes service level Value.

Step 102), described historical data is input in model agg-anfis, with reference to formula (11) or formula (12), using the output result of model agg-anfis, the parameter of model agg-anfis is adjusted, obtains Model agg-anfis after parameter adjustment；Wherein,

The parameter of described model agg-anfis includes 27 premise parameters, 24 consequent parameters, is input variable 3 weight parameter of setting, 27 weight parameter for rule settings；

Step 103), calculation procedure 102) obtained by parameter adjustment after model agg-anfis lowest mean square Difference, judges whether this value reaches a specified threshold, if reached it is meant that model training terminates；Otherwise, execute Next step；

Step 104), read one group of new historical data, then re-execute step 102).

2nd, test phase

Step 201), read test data, described test data includes: the history containing flow, occupation rate and speed Data；

Step 202), test data is input in model agg-anfis, obtain the value of service level (los).

It is more than the step description to the method for the present invention.Below by the method for the present invention with of the prior art other Method is compared, to prove the effectiveness of the inventive method.

Initially set up sugeno-fis the and aggfis fuzzy inference system for experimental contrast analysis, and set up right The service level evaluation model based on anfis and agg-anfis answered, is analyzed as follows:

1. in anfis and agg-anfis evaluation model, in fuzzy rule former piece, speed, flow and The determination of membership function (fuzzy set) parameter of occupation rate, can obtain according to historical data analysis；For mould The determination of paste consequent service level class, can obtain according to historical data analysis.But this fuzzy class Division is not the most accurate, so the adjustment of grade to be carried out.Input historical data and according to learning rules, right It is trained based on the Adaptive Neuro-fuzzy Inference (agg-anfis) that choquet integrates-owa, that is, right The parameter set (premise, conclusion) of model is adjusted, here it is agg-anfis has learning functionality (self adaptation) Embodiment.Especially, the reasoning process of agg-anfis considers the power of object (system input and every rule) Weight, these are also intended to be learnt the parameter of (training).When system reaches minimum error index, training terminates.

2. it is based on existing historical data and specific as follows is represented to the fuzzy classification of traffic behavior service level (los):

1) imin and imax of evaluation index i (domain) of a certain los can according to historical data, be learnt；

2) for a certain index i (domain), carry out los grade classification (initialization membership function), divide altogether For abcdef6 grade.

When taking different types of membership function, we have corresponding consequent parameter collection.

3) for evaluation index i, available i (t1), i (t2) ... ... ..i (tn), represent t1 respectively, t2 ... the .tn moment History value i (ti).Be also the last string of training data, i.e. desired output.

According to the fuzzy inference system set up, Calculation Estimation index i ' (ti).Corresponding to a certain when be carved with:

$\left[\begin{array}{c}\begin{array}{cc}a\left(t_1\right),...& ,f\left(t_1\right)\end{array}\\ \begin{array}{cc}a\left(t_2\right),...& ,f\left(t_2\right)\end{array}\\ ..........................................\\ \begin{array}{cc}a\left(t_n\right),...& ,f\left(t_n\right)\end{array}\end{array}\right]$

In certain moment ti, corresponding specific Indistinct Input, according to defined fuzzy rule, above matrix can be released The value of every a line, according to center deblurring method, obtains ti moment, the exact value of evaluation index i, this value is aggfis The value being calculated, i ' (ti).

In the ti moment, i (ti) known to the value of index i.For example, we according to v/c as evaluation index i, and here Abcdef6 fuzzy set is set up on domain, then in the ti moment, known to v/c.With being worth and I known to this The value contrast released of model, obtain error signal i ' (ti)-i (ti), according to bp algorithm, to the agg-anfis setting up Model parameter be adjusted.

When system inputs it is known that desired output i (ti) is it is known that error signal is it is known that corresponding agg-anfis can be constructed Model realization los evaluates it is possible to be adjusted (study) to the parameter in model.In agg-anfis model All parameters be nonlinear parameter, application gradient descent method (steepest descent method) be adjusted, design parameter is more New formula is with reference to (formula 11).When the minimum error index of the system that reaches, training terminates.

Specific experiment step is as follows:

First, sugeno-fis and anfis experiment

First, set up sugeno-fis fuzzy inference system.The detection data that we can obtain, i.e. flow, account for There are rate and speed, these three variables input as system；And service level value (grade), it is system output.According to Training (history) data, can obtain the reality output curve of s-fis.

Secondly, set up corresponding anfis service level evaluation model, determine premise parameter collection and consequent parameter collection. Input historical data (training data), the reality output according to sugeno-fis model and training data desired output Error, is adjusted to model parameter using Hybrid learning rule, when the response times of the system that reaches (40), instruction White silk terminates.Can parameter set before and after comparative training.

Finally, input test data (less than training data space), surveys to the sugeno-fis model training Examination, it is possible to obtain system test error, can be used as the important indicator of model validation.

Sugeno-fis output curve diagram is as shown in figure 4, the abscissa of this in figure represents frequency of training, vertical coordinate table Show the reality output of sugeno-fis fuzzy inference system, this figure reflects the reality of sugeno-fis fuzzy inference system Border export, the error between the reality output of sugeno-fis fuzzy inference system and desired output as shown in figure 5, The abscissa of this in figure represents frequency of training, vertical coordinate represent the reality output of sugeno-fis fuzzy inference system with Desired output error.The training data error curve diagram of anfis is as shown in fig. 6, the abscissa of this in figure represents loud Answer number of times, vertical coordinate represents the training error of anfis.Using hybrid algorithm training, fault-tolerant index 0, response time Number 40.Error is progressively successively decreased with the increase of frequency of training it was demonstrated that the training dataset chosen in the application is that have Effect.Final training error is error=1.0038e-005.

Fig. 7 be detection data desired output and training after sugeno-fis real output value between comparison diagram, Wherein abscissa represents test data, and vertical coordinate represents that the sugeno-fis real output value after training (uses star * Represent) and detection data desired output (with cross+represent), find out the phase of detection data as seen from the figure Hope and coincide preferably between the sugeno-fis real output value after output valve and training.S-fis model after training can For system identification and prediction, it is one kind preferably None-linear approximation device.Average test error is 0.091368.Real Time of testing is the elapsed_time=8.3110 second.

(1) parameter comparison before and after the membership function (nonlinear parameter) of former piece changes:

(2) contrast before and after consequent consequent parameter (linear dimensions) changes:

Consequent parameter	Before change	After change
			mf1	[0.0002 0.006 -0.005 0.4]；	[0.0002906 0.006583 -0.00548 0.4235]
mf10	[0.0002 0.006 -0.005 0.2]	[0.0002911 0.006667 -0.005522 0.4228]
			mf 24	[0.0004054 0.01235 -0.003485 0.004147]	[0.0003725 0.01571 -0.004825 0.004552]

2nd, aggfis and agg-anfis experiment

First, we can set up aggfis model, programming realization aggfis according to the reasoning process of aggfis Reasoning, input training data can obtain the reality output of aggfis.Model parameter to be ensured has learnability, When selecting each operator, note the calculating seriality of operator.I.e. the process of whole reasoning can micro- be led, I This model of setting up just there is learnability so that it may be adjusted to its parameter.

Then, we set up corresponding agg-anfis model to carry out los evaluation, need the parameter adjusting to have:

1) parameter of former piece and consequent membership function；

2)u_i: the weight of each input, initial value is 1；

3)τ_i: the weight of every rule, initial value is 1.

The whole parameter of agg-anfis model comprises: premise parameter collection and consequent parameter collection are all nonlinear parameters, often The weight of the weight of individual input and every rule is also the object that we need to adjust, according to our aggfis models Output and the error of training data desired output, are adjusted using parameter learning rule (formula 11), when reaching During the minimum error index of system, training terminates.Can parameter set before and after comparative training.

Finally, input test data (less than training data space), tests to model, it is possible to obtain model is surveyed Examination error.

Before training, as shown in figure 8, wherein, Fig. 8 (a) represents the membership function of input variable and output variable The membership function of the flow in input variable, the abscissa of this figure represents the threshold value of flow, and vertical coordinate represents flow Membership function；Fig. 8 (b) represents the membership function of the occupation rate in input flow rate, the abscissa table of this figure Show the threshold value of occupation rate, vertical coordinate represents the membership function of occupation rate；Fig. 8 (c) is the speed in input variable Membership function, the abscissa of this figure represents the threshold value of speed, and vertical coordinate represents the membership function of speed；Fig. 8 D () represents the membership function of output variable los, the abscissa of this figure represents the threshold value of los, vertical coordinate table Show the membership function of los.

Fig. 9 is aggfis output curve diagram, and the abscissa of this figure illustrates 1429 pairs of training datas, vertical coordinate table Show aggfis output, this figure illustrates the reality output of aggfis before training.

Figure 10 is the error curve diagram of aggfis reality output and desired output, and the abscissa of this figure represents 1429 To training data, vertical coordinate represents aggfis output and the error of desired output；This figure illustrates to train front aggfis The reality output of model is larger with desired output error, but in allowed band, and be uniformly distributed up and down along 0 value.

Agg-anfis model is carried out adjust ginseng, train the cost time: 2.624 seconds.Mse (training error)= 0.00022442；Ans (test error)=0.057391.Experimental result is as follows:

Input weight matrix before training:

$\begin{array}{ccc}150& 250& 400\\ 2.5& 6& 10\\ 15& 9& 15\end{array}$

Output weight matrix before training:

0.2 0.35 0.55 0.75 0.85 0.99

Input variable mu value matrix before training

1 1 1

Regular tau value matrix before training

columns 1 through 27

$\begin{array}{ccccccccc}1& 1& 1& 1& 1& 1& 1& 1& 1\\ 1& 1& 1& 1& 1& 1& 1& 1& 1\\ 1& 1& 1& 1& 1& 1& 1& 1& 1\end{array}$

Input weight matrix after training:

$\begin{array}{ccc}149.87& 249.93& 399.9\\ 2.4567& 6.2972& 10.297\\ 7.3413& 8.3121& 11.808\end{array}$

Output weight matrix after training:

0.19158 0.33113 0.55151 0.66184 0.82544 0.90918

Input variable mu value matrix after training

0.60937 0.97412 0.59182

Regular tau value matrix after training

columns 1through 27

$\begin{array}{ccccccccc}0.49891& 0.44642& 1.1605& 0.44642& 1.1984& 0.80317& 0.81969& 0.31815& 0.10403\\ 0.44642& 0.89124& 0.80317& 0.40379& 0.99201& 0.20816& 0.61759& 0.53852& 1.0974\\ 0.18051& 1.1575& 0.31815& 0.55074& 0.59043& 0.49891& 0.49891& 0.49891& 0.49891\end{array}$

Figure 11 is the training error curve chart of agg-anfis, and the abscissa of this figure represents training sample, vertical coordinate Represent training error；This figure illustrates that training error reduces with the increase of training sample it was demonstrated that setting up Agg-anfis model is more excellent.

Figure 12 is the output valve of aggfis and detection data desired output comparison diagram after training, the abscissa of this figure Represent 640 to test sample, vertical coordinate represents detection data desired output (in figure use --- expression) and aggfis Real output value (in figure use -- *-expression)；This figure illustrates aggfis reality output and detection data after training Desired output coincide preferably, model training success.

Figure 13 is the test error curve chart of aggfis, and the abscissa of this figure illustrates 640 pairs of test samples, indulges The coordinate representation test error of aggfis, it can be seen that the test error value of aggfis is relatively low.

Figure 14 is the membership function after changing, and Figure 14 (a) therein represents being subordinate to of the flow in input variable Degree function, Figure 14 (b) represents the membership function of the occupation rate in input flow rate, and Figure 14 (c) is input variable In speed membership function；Figure 14 (d) represents the membership function of output variable los.Compared with Fig. 8, The membership function parameter of each input variable and output variable all changes.

It should be noted last that, above example is only in order to illustrate technical scheme and unrestricted.Although With reference to embodiment, the present invention is described in detail, it will be understood by those within the art that, to the present invention Technical scheme modify or equivalent, without departure from the spirit and scope of technical solution of the present invention, it is equal Should cover in the middle of scope of the presently claimed invention.

Claims (4)

1. a kind of traffic state evaluation method based on Adaptive Neuro-fuzzy Inference, including the training stage With test phase；Wherein,

The described training stage comprises the following steps:

Step 101), read the historical data of input variable and output variable；Described input variable includes flow, accounts for There are rate and speed, described output variable includes service level value；

Step 102), by step 101) historical data that inputted is input in Adaptive Neuro-fuzzy Inference, Obtain the output result of Adaptive Neuro-fuzzy Inference；Parameter learning rule based on anti-pass thought, in conjunction with institute The output result stating Adaptive Neuro-fuzzy Inference is adjusted to the parameter of Adaptive Neuro-fuzzy Inference Whole, obtain the Adaptive Neuro-fuzzy Inference after parameter adjustment；Wherein,

The parameter of described Adaptive Neuro-fuzzy Inference includes premise parameter, consequent parameter, sets for input variable Fixed weight parameter, the weight parameter for rule settings；Described premise parameter is by the membership function shape of reasoning former piece Shape determines；Described consequent parameter is determined by the membership function shape of reasoning consequent；

Step 103), calculation procedure 102) obtained by parameter adjustment after Adaptive Neuro-fuzzy Inference Minimum Mean Square Error, judges whether this value reaches a specified threshold, if reached it is meant that the training stage terminates；No Then, execute next step；

Step 104), read one group of new historical data with regard to input variable and output variable, then re-execute Step 102)；

Described test phase comprises the following steps:

Step 201), read test data, described test data includes: the history number of flow, occupation rate and speed According to；

Step 202), test data is input to via the Adaptive Neural-fuzzy Inference system that obtains of training stage training In system, obtain the value of the service level for describing road traffic state.

2. the traffic state evaluation side based on Adaptive Neuro-fuzzy Inference according to claim 1 Method, it is characterised in that described Adaptive Neuro-fuzzy Inference includes five layers, is respectively as follows:

Ground floor: obscuring layer, for by accurate fuzzy inputing method；

o_1,i=μ_ai(x), i=1,2 or

$\begin{array}{cc}{o}_{1,i}={\mathrm{\&mu;}}_{b_{i-2}}\left(y\right),& i=3,4\end{array}$

Wherein, o_j,iRepresent i-th node output of jth layer, μ_ai、For the membership function of former piece, this two Individual former piece membership function is defined with generalized bell mf

${\mathrm{\&mu;}}_a=\frac{1}{1+{\&lsqb;{\left(\right(x-c_i)/a_i)}^2\&rsqb;}^{b_i}}$ Or ${\mathrm{\&mu;}}_b=\frac{1}{1+{\&lsqb;{\left(\right(x-c_i)/a_i)}^2\&rsqb;}^{b_i}}$

A therein_i、b_i、c_iPremised on parameter；

The second layer, reasoning layer, for calculating the excitation density w of every rule_i；

$o_{2,i}=w_i=\&lsqb;\stackrel{\&overbar;}{{\mathrm{\&mu;}}_1}+{\mathrm{\&mu;}}_1\&times;{\mathrm{\&mu;}}_{ai}\left(x\right)\&rsqb;\&times;\&lsqb;\stackrel{\&overbar;}{{\mathrm{\&mu;}}_2}+{\mathrm{\&mu;}}_2\&times;{\mathrm{\&mu;}}_{bi}\left(y\right)\&rsqb;$

Wherein, i=1,2, μ_iRepresent the weight of each input,

Third layer, hidden layer, for calculating effective consequent mf of every rule；

o_3,i=w_iоc_i

Wherein, i=1,2；c_iReasoning consequent；" o " represents implicit operator；

4th layer, polymer layer, for calculating the summation of strictly all rules effective consequent mf；

o₄=σ (w_iоc_i-w_i-1оc_i-1)×τ_i

Wherein, i=1,2；τ_iRepresent the weight of every rule；Aggregation Operator σ (w_iоc_i-w_i-1оc_i-1)×τ_iUsing Choquet integrates；

Layer 5, mould from paste layer, for the accurate output of computing system；

o₅=f=d o o₄

Wherein, d represents deblurring operator, and it calculates is realized using center de-fuzzy method.

3. the traffic state evaluation side based on Adaptive Neuro-fuzzy Inference according to claim 2 Method is it is characterised in that be adjusted with reference to equation below to the parameter of Adaptive Neuro-fuzzy Inference:

δw_ji=η (d_i-x_i)·x_j·x

Wherein, δ w_ijRepresent the increment of arbitrary parameter in Adaptive Neuro-fuzzy Inference；η is Learning Step, d_i It is the desired output of node i, x_iIt is the reality output of node i, x_jIt is the input of node i, node i is node j Last layer node, j < i；X is a multinomial, uses (x_i×(1-x_i)) expression.

4. the traffic state evaluation side based on Adaptive Neuro-fuzzy Inference according to claim 2 Method is it is characterised in that described Adaptive Neuro-fuzzy Inference replaces and in reasoning layer with owa operator Operator or or operator, calculate excitation density；Integrated with choquet to realize being polymerized in polymer layer；De-fuzzy is calculated Son adopts center de-fuzzy method；And set the weight of each input, use μ_iRepresent；Set the power of every rule Weight, uses τ_iRepresent.