KR101742043B1 - Apparatus and method for travel mode choice prediction - Google Patents

Abstract

A linearization module for receiving first information, which is information on a transportation cost for each transportation means, from a traffic database and linearizing the information, and a second module for receiving the linearized first information and extracting second information, And a deep learning module for processing the first information and the second information based on supervised learning.

Description

[0001] APPARATUS AND METHOD FOR TRAVEL MODE CHOICE PREDICTION [0002] BACKGROUND OF THE INVENTION [0003]

The present invention relates to an apparatus, a method, and a storage medium for storing a program for predicting a sharing of a transportation means.

The utility function including the relation between the input variables such as the communication cost variable could not be directly calculated in the polynomial logit model for calculating the transportation share ratio of the conventional technology. In addition, in the polynomial logit model, it is a standard method to use the input data including the average linearized parameter value of the data. In addition, the utility function used for the calculation of the transportation share ratio includes complicated nonlinearities, which is not easy to calculate accurately.

On the other hand, in the art utilizing artificial neural network, the use of a sigmoid function including a nonlinear exponential function in the hidden layer eliminates the weight value between the neurons and the vanishing gradient problem, . Also, there was a problem of over-fitting according to the increase of the hidden layer. However, recently, a deep learning algorithm based on artificial intelligence has appeared to solve the tilt reduction problem described above, and a method for mitigating the above-described overfitting problem has been developed.

Hybrid approach to technology based on artificial neural network described above in the polynomial logit model for calculating the above-mentioned traffic share ratio is a basic level, and specifically, it is a level of simple comparison of individual results.

As described above, the present invention aims to provide a transportation device sharing prediction device, a method, and a storage medium for storing a program for predicting a transportation device sharing, in order to hybridize a technique based on an artificial neural network to a multinomial logit model.

 The traffic sharing contribution prediction apparatus of the present invention comprises: a linearization module for receiving first information, which is information on a traffic cost for each transportation means, from a traffic DB and linearizing the received information; A polynomial module for receiving the linearized first information and extracting second information, which is a probability density of a contribution ratio for each transportation means; And a deep learning module for processing the first information and the second information based on supervised learning; .

In addition, the method of predicting a traffic sharing contribution of the present invention may include: receiving and linearizing first information, which is information on a traffic cost for each transportation means, from a traffic database; Receiving linearized first information and extracting second information which is a probability density of a contribution ratio for each transportation means; And processing the first information and the second information based on Supervised Learning; .

In addition, the storage medium storing the program for predicting the sharing of transportation means according to the present invention receives and linearizes the first information, which is information on the transportation cost for each transportation means, from the traffic DB, receives the linearized first information, The first information and the second information are processed based on Supervised Learning, the linearized first information or the second information is received, and the backpropagation is performed using the back propagation algorithm Extracting the distribution value of the sharing ratio for each transportation means based on the optimized weight (weight value) value or the weight value, storing the optimized weight value, and optimizing the information on the traveling cost for each transportation means based on the stored weight value .

INDUSTRIAL APPLICABILITY As described above, according to the present invention, it is possible to improve the accuracy of the calculation of the share of the transportation means.

In addition, according to the present invention, the probability density and optimal utility function for the transportation means sharing can be obtained by using the traffic-related big data.

In addition, according to the present invention, it is possible to easily provide various measures for improving the share of the public transportation means based on the actual bus / public transportation stop by calculating the correct transportation means contribution ratio.

In addition, according to the present invention, it is possible to reduce the congestion of public transportation occurring in a large city by providing an accurate calculation of the transportation sharing ratio.

1 is a view for explaining an embodiment of a traffic sharing prediction system.
FIG. 2 is a diagram for explaining an embodiment of a transportation contribution sharing prediction system based on deep learning.
3 is a view for explaining an embodiment of a means for predicting a traffic share.
FIG. 4 is a diagram for explaining an embodiment of a method of predicting a traffic sharing contribution. FIG.

One embodiment of the present invention will be described in detail with reference to the accompanying drawings. In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear.

1 is a view for explaining an embodiment of a traffic sharing contribution prediction system.

1, the transportation unit contribution prediction system includes a travel cost variable processing module 100, a linearization variable module 110, a polynomial property calculation module 120, a means share ratio output module 130, a hybrid input module 140 ), A deep running calculation module 150, or a non-linearization variable module 160. The traffic sharing prediction system may further include a weight storage module 170 or a probability distribution output module 180.

The travel cost variable processing module 100 can receive information on the travel cost for each transportation means from the traffic DB 190 and process it. Specifically, the travel cost variable processing module 100 can receive and process the information on the travel cost for each transportation means according to the following Equation (1).

For example, if k = 1, a bus, k = 2, k = 3, and k = 4 may be used. M can correspond to the above-described information on the travel cost by the transportation means by the total number of information used. Also, i is information on the starting point and j is information on the arrival point. Also, Uij (k) means utility value, that is, information on the above-mentioned travel cost by the transportation means when moving from the departure point i to the arrival point j through the transportation means k.

Also, the travel cost variable processing module 100 can transmit the processed travel cost information for each transportation means to the linearization variable module 110. [ The above-mentioned means of transportation include those utilized as transportation means such as public transportation, subway, train, passenger car, and motorcycle, and are not limited to the above-mentioned examples. In addition, the information on the travel cost of each transportation means may include information on the travel time of each transportation means or cost information of each transportation means.

The above-mentioned information on the travel time of each transportation means is related to the travel time in the transportation means, the walking time in approaching the individual transportation means, the time required between the starting point and the end point, the waiting time, And is not limited to the example described above. In addition, the cost information for each transportation means includes information on necessary or consumed expenses related to utilization as transportation means such as vehicle operation cost, toll gate fee, parking fee, number of transit times, etc., and is not limited to the above-described example.

The linearization variable module 110 may linearize nonlinear information among the above-mentioned information by inputting the traffic cost information for each transportation means from the traffic cost variable processing module. In addition, the linearization variable module 110 may transmit the linearized transportation cost information for each transportation means to the polynomial building calculation module 120. The linearization variable module 110 may calculate an average value of individual information input for linearization of information and transmit the information on the average value to the polynomial calculation module 120.

The polynomial calculation module 120 receives the linearized transportation cost information for each transportation means from the linearization variable module 110 and calculates the contribution ratio for each transportation means as the probability density. Specifically, the polynomial calculation module 120 receives the linearized transportation cost information for each transportation means from the linearization variable module 110 according to Equation (1) below, and calculates the contribution ratio for each transportation means at a probability density. Also, the polynomial calculation module 120 may transmit the above-described information on the probability density to the means contribution ratio output module 130. [ In addition, the above-mentioned contribution ratio per transportation means may be 1 in total sum. For example, the share of transportation in Daejeon in 2015 is 28.8% for public transportation (24.6% for buses and 4.2% for urban railways), 55.6% for passenger cars, 9.5% for taxis and 6.1% for taxis.

The hybrid input module 140 can receive the traffic cost information by the transportation means or the information on the probability density from the above-described means contribution ratio output module 130 from the traffic cost variable processing module 100. [ In addition, the hybrid input module 140 may receive information on the travel cost of each transportation means according to individual individuals. In addition, the hybrid input module 140 calculates information on the travel cost for each transportation means, information on the travel cost of each transportation means according to the individual individual, individual personal information related to the information about the transportation means, Module 150, as shown in FIG.

The deep running calculation module 150 can process the above-described information received from the hybrid input module 140 according to Equation (2) below. For example, if k = 1, it is a bus, k = 2 is a passenger car, k = 3 is a taxi, and k = 4 is a train. M can correspond to the above-described information on the travel cost by the transportation means by the total number of information used. Also, i is information on the starting point and j is information on the arrival point. Also, Uij (k) means utility value, that is, information on the above-mentioned travel cost by the transportation means when moving from the departure point i to the arrival point j through the transportation means k. Pij (k) is the probability of using the transportation means k among all available means of transportation when an individual moves from the starting point i to the arrival point j.

In addition, the deep running calculation module 150 may process the information input from the hybrid input module 140 based on the deep running technique. The deep run calculation module may also process information received from the hybrid input module 140 by a supervised learning technique. The above-mentioned map learning technique means adjusting the value of the link strength between neurons and neurons in each layer in the artificial neural network in order to minimize the error between the calculated value and a preset value (correct answer). Specifically, the deep-run calculation module 150 may determine the value of the individual neurons and the value of the individual weights in the deep-run calculation module based on predetermined or predetermined values for the input information and the output information. Also, the deep running calculation module 150 can optimize the individual weight values in the direction of minimizing the error using a random number. Also, the deep-run calculation module 150 may optimize the input values of the deep-run calculation module 150 based on the backward propagation algorithm on weight values and output values optimized by the guidance learning techniques. A detailed description of the deep run calculation module will be described later with reference to FIG.

The non-linearization variable module 160 can extract the optimized utility function value by utilizing the optimized input value, the optimized output value, or the optimized weight value. The nonlinearization variable module 160 can extract an optimized utility function value using an optimized input value, an optimized output value, or an optimized weight value for each transportation means as shown in Equation 3 shown below.

The weight storage module 170 may store the finally optimized weight value using the back propagation algorithm in the deep run calculation module 150. [ The weight storage module 170 may also update the stored weight value as the change occurs.

The probability distribution output module 180 can use the back propagation algorithm in the deep running calculation module 150 to output the probability distribution value of the optimized traffic sharing ratio information using the finally optimized weight value. In the conventional technology, when a deep learning technique such as the present invention is used, a soft mode activation function having a sigmoid activation function and an output of 0 or 1 is used. However, , I.e., 0, 0.25, 0.5, and so on. The above-described polynomial values are not limited to those described above since design changes are possible.

FIG. 2 is a diagram for explaining an embodiment of a transportation contribution sharing prediction system based on deep learning.

Referring to FIG. 2, the deep run calculation module 150 described in FIG. 1 may be hierarchically or parallelly divided, including an input layer, a hidden layer, a soft max layer, a hybrid input layer, or an output layer. The deep running calculation module 150 includes a variable input module 200, a ReLU module, a dropout module, a hidden layer adjustment module, a soft max output module 250, or a nonlinear utility function output module 280 can do. In addition, the ReLU module may correspond to a ReLU (Retained Linear Unit) linear unit 220, the dropout module may be a dropout 230, and the hidden layer adjustment module may correspond to the hidden layer adjustment unit 240. Further, the deep running calculation module 150 may further include a means contribution ratio output module 270 or a multinomial result input module 260.

The deep running calculation module 150 can process the probability distribution value to be traveled by using any transportation means according to the following equation (4).

P (k) is a probability distribution value to be traveled by using any transportation means, k is information on the number of transportation means, and M is information on the above-mentioned travel cost in FIG. L is information about the hidden layer. Also, x (k) is information on the explanatory variable introduced in Equation (1) described above, and w (k) means information on the weight value described above with reference to FIG. Specifically, for example, when the transportation means is four buses, a car, a taxi, and a railway, the information about the travel cost is two times and the cost, and the number of hidden layers is ten in deep running, , M is 2, and L is 10. B is a bias value, and the initial value may be set to a random value or an arbitrary value and updated to an optimization value.

The per-instrumental parameter input module 200 is included in the above-mentioned input layer of the hierarchical structure, and can receive information on the respective travel expenses for the individual transit means. The above-described information on the travel cost may include information on the travel cost, travel time, or variables as described above with reference to FIG. In addition, the device-specific variable input module 200 can receive at least one input signal (M x k) in which k pieces of transportation means each have M number of travel cost variables. The number of input signals (neurons) is 2 x 2, 4 in case of bus and subway to individual transportation means, and information on traffic cost for traffic cost and travel time.

The hidden layer of the above-described hierarchical structure may include a ReLU module (ReLU linear unit 220), a dropout module (dropout 230), or a hidden layer adjustment module (hidden layer adjustment device 240). Also, the deep running calculation module 150 may use a ReLU (Retired Linear Unit). The aforementioned ReLU (Retinal Linear Unit) is effective in solving the vanishing gradient problem due to the increase in the number of hidden layers in the deep running technique. The active function used in the ReLU module is shown in Equation (5) below. The ReLU module can process the input values as 0 when x < 0 in Equation (5) below. In addition, the ReLU module can output the above-described input values when x> 0.

The dropout module may process the value of an individual neuron corresponding to an output value or output at any probability value to zero. The above-mentioned arbitrary probability value is independent of the probability value related to the traffic contribution ratio and can be arbitrarily set by the designer. Through the setting of any of the probability values described above, it is possible to determine neurons not to be used or calculated. As a specific example, when a certain probability value is set to 0.5, the dropout module sets the output (output neuron) values corresponding to the input signal (neuron) input through the above-described individual variable input module 200 to 0 Can be processed.

The soft max output module 250 may output the probability distribution value corresponding to each transportation means by using Equation (4) based on M, which is information on the above-described traffic cost. In addition, the sum of the probability distribution values corresponding to the aforementioned transportation means may be one. The nonlinear utility function output module 280 may receive information on the output of the soft max output module 250 and update the polynomial plotting module 120 described above with reference to FIG.

The polynomial result input module 260 may correspond to the polynomial plotting module 120 described above with reference to FIG. The means contribution ratio output module 270 may correspond to the means contribution ratio output module 130 described above with reference to FIG. Meanwhile, the hybrid input module 140 may receive the information about the probability density from the above-described means contribution ratio output module 130 and transmit it to the deep learning calculation module 150.

3 is a view for explaining an embodiment of a means for predicting a traffic share.

Referring to FIG. 3, the transportation contribution prediction device may include a polynomial log module 300 and a deep learning module 310. The transportation device sharing prediction device may include a linearization module 320, a hybrid input module 330, an extraction module 340, or a storage module 350.

The polynomial module 300 receives the linearized first information and extracts the second information, which is the probability density of the contribution ratio for each transportation means. The polynomial fitting module 300 may correspond to the polynomial fitting calculation module 120 and the means share ratio output module 130 described in FIG. 1, and a detailed description thereof has been given above with reference to FIG.

The deep learning module 310 may process the above-described first information and second information based on supervised learning. The deep learning module 310 may correspond to the deep learning calculation module 150 or the hybrid input module 140 described above with reference to FIG. 1, and a detailed description thereof has been given above with reference to FIG. 1 and FIG.

The linearization module 320 can receive and linearize the first information, which is information on the traffic cost for each transportation means, from the traffic DB. The above-described linearization module 320 may correspond to the linearization variable module 110 described above with reference to FIG. 1, and a detailed description thereof has been given above with reference to FIG. Also, the above-mentioned traffic DB may include the above-described travel cost variable processing module 100 in FIG. The linearization module 320 may also receive information on travel costs from the travel cost variable processing module 100 described above. Also, the first information may be information on the travel time of each transportation means or cost information of each transportation means.

The hybrid input module 330 may receive the linearized first information or the second information. The hybrid input module 330 may correspond to the hybrid input module 140 described above with reference to FIG. 1, and a detailed description thereof has been given above with reference to FIG.

The extraction module 340 may extract an optimized probability distribution value or a predicted probability distribution value for the transportation means contribution ratio based on the optimized weight (weight value) or weight value using the back propagation algorithm. The extraction module 340 may correspond to the nonlinearization variable module 160 or the probability distribution output module 180 described above with reference to FIG. 1, and a detailed description thereof has been given above with reference to FIG.

The storage module 350 may store an optimized weight value. The storage module 350 may correspond to the weight storage module 170, and a detailed description thereof has been given above with reference to FIG. The above-described traffic DB or the above-described travel cost variable processing module 100 in FIG. 1 can optimize information on the travel cost for each transportation means based on the above-described stored weight value.

FIG. 4 is a diagram for explaining an embodiment of a method of predicting a traffic sharing contribution. FIG.

Referring to FIG. 4, the method for predicting the shared transportation method includes receiving (S400) the first information, which is the information on the travel cost of each transportation means, from the traffic database (S400), receiving the linearized first information, (S410) extracting second information having a probability density, and processing (S420) processing the first information and the second information based on supervised learning. The linearization module 320 may perform step S400 of receiving and linearizing the first information, which is information on the transportation cost for each transportation means, from the traffic DB. A detailed description thereof has been given above with reference to Figs. 1, 2 and 3. The polynomial module 300 may perform the step S410 of receiving the linearized first information and extracting the second information, which is the probability density of the contribution ratio for each transportation means. A detailed description thereof has been given above with reference to Figs. 1, 2 and 3. The deep learning module 310 may perform processing (S420) of processing the first information and the second information based on Supervised Learning. A detailed description thereof has been given above with reference to Figs. 1, 2 and 3.

The embodiments disclosed in the specification of the present invention do not limit the present invention. The scope of the present invention should be construed according to the following claims, and all the techniques within the scope of equivalents should be construed as being included in the scope of the present invention.

100: Transit Cost Variable Processing Module
110: linearization variable module
120: polynomial calculation module
130: means share ratio output module

Claims (13)

A linearization module for receiving first information, which is information on a traffic cost for each transportation means, from the traffic DB and linearizing the received information;
A polynomial module for receiving the linearized first information and extracting second information, which is a probability density of a contribution ratio for each transportation means; And
A deep learning module for processing the first information and the second information based on supervised learning; And means for estimating a share of the transportation means. The apparatus of claim 1, further comprising a hybrid input module for receiving the linearized first information or the second information. 2. The traffic sharing device according to claim 1, further comprising an extraction module for extracting a weight value (weight value) optimized using a back propagation algorithm or a distribution value of a contribution ratio for each transportation means based on the weight value, Apparatus for predicting means sharing. 4. The apparatus of claim 3, further comprising a storage module for storing the optimized weight value. 5. The traffic sharing device according to claim 4, wherein the traffic sharing prediction device optimizes the information on the traffic cost for each transportation means based on the stored weight value.  The apparatus according to claim 1, wherein the first information is information on travel time for each transportation means or cost information for each transportation means. The linearization module receiving and linearizing the first information, which is information on the traffic cost for each transportation means, from the traffic database;
The polynomial module receiving the linearized first information and extracting second information, which is a probability density of the contribution ratio for each transportation means; And
The deep learning module processing the first information and the second information based on Supervised Learning; The method comprising: 8. The method of claim 7, wherein the method further comprises receiving the linearized first information or the second information. 8. The method according to claim 7, wherein the traffic sharing prediction method further comprises: extracting a weight value (weight value) optimized using a back propagation algorithm or a distribution value of a contribution ratio for each transportation means based on the weight value Method of Estimating Participation. 10. The method of claim 9, wherein the method further comprises storing the optimized weight value. 11. The method according to claim 10, wherein the traffic sharing prediction method optimizes information on a traffic cost for each transportation means based on the stored weight value.   The method according to claim 7, wherein the first information is information on a travel time for each transportation means or cost information for each transportation means. Receiving the first information, which is the information on the travel cost of each transportation means, from the traffic DB, linearizing the first information, extracting the second information which is the probability density of the transportation rate for each transportation means, Processing the second information based on Supervised Learning, receiving the linearized first information or the second information, and using the backward propagation algorithm to optimize the weight (weight) value or the weight value based on the weight And stores the optimized weight value, and optimizes the information on the travel cost of each transportation means based on the stored weight value, thereby storing a program for predicting the transportation means contribution Storage medium.

Images