Patent Application
20080097686
Preferred modes of the present invention will now be described in detail with reference to the drawings.
The travel-time realtime data 101 is time-series data formed for every road-segment unit (link) from data in a probe-car system and an information source such as a VICS (Vehicle Information & Communication System®) . The details will be described later.
Stored in the travel-time transition pattern database 104 with regard to each road-segment unit (link) are travel-time transition patterns obtained by subjecting various past index values over a prescribed time period to required statistical processing such as elimination of out-of-spec values and correlation analysis using the travel-time realtime data 101. The statistical processing is executed for every predetermined unit of time for every day type, such as day of the week, the fifth day of the month, season and weather, in the time-series data. Accordingly, travel-time transition patterns are prepared for a period of 24 hours and suitable patterns can be used in accordance with various circumstances. The unit of time is decided in accordance with prediction accuracy and the overall load of the system. Conceivable units of time are every five minutes and every 15 minutes, etc. The details of these travel-time transition patterns will be described later.
The travel-time prediction apparatus 100 includes pattern conversion means 102 and predicted-value calculation means 103 for executing prediction processing using a prediction function described later in detail. In accordance with a request from the user, the travel-time prediction apparatus 100 combines the travel-time realtime data 101 and travel-time transition patterns stored in the travel-time transition pattern database 104, obtains short-term (after 5 or 15 minutes) predicted time, mid-term (up to several hours from short-term onward) predicted time and future predicted time with respect to the road-segment unit (link) that is the target of the prediction, and outputs the predicted time. Here the road-segment unit (link) that is the target of the prediction basically is decided by being specified on the user side, and it is assumed that from several tens to several tens of thousands can be adopted as the target.
The travel-time prediction apparatus 100 is characterized by its mid-term prediction processing in order to shorten, as much as possible, the processing time needed for a prediction while the high accuracy of the prediction is maintained. The mid-term prediction processing of the travel-time prediction apparatus 100 will be described below.
[Travel-Time Realtime Data (Time-Series Data)]
The travel-time realtime data 101 used in mid-term prediction processing will be described first. Here the term “link” refers to a road segment typically having a length of from several tens of meters to several hundred meters defined between intersections, by way of example. The end of a link, such as an intersection, is referred to as a “node”.
Assume that there are d-number of prediction-target links, and let a vector obtained by arraying realtime data of each link at time t be represented by x1=(xt:1, xt:2, . . . xt:d)εD=X1×X2× . . . ×Xd. Here D is referred to as a “domain”.
Each xt:1 is assumed to represent an index indicating travel time, number of vehicles and occurrence of gridlock in link i at time t, or an index value of various attributes relating to traffic conditions, such as weather at this time. Each xt:1 is a continuous value or discrete value.
Let t be an integral value for the sake of convenience. Assume that time-series data over a predetermined time interval is constituted by a vector sequence {xt}. For example, if the predetermined time interval is five minutes, then x2 will represent the data of x1 after five minutes. Let the sequence xm . . . xn be represented by xmn(m≦n), and in particular, assume that xn=x1n holds.
[Travel-Time Transition Pattern]
Next, the travel-time transition patterns stored in the travel-time transition pattern database 104 will be described. A travel-time transition pattern at time t follows xt and is represented by wt. Here we assume that wt is obtained by recording a past average value of a quantity corresponding to xt for every time period.
Since wt differs depending upon the day type, such as day of the week, weather and whether or not the day is a holiday, wt is formed according to each day type. Accordingly, it is assumed that wt has a periodicity in which the original value is restored when time advances by 24 hours.
The problem involved in forming wt is a problem involving the learning of a regression equation that correlates (time period, day type) to travel time. Various concrete methods of forming wt are conceivable. One example that can be mentioned is a method in which the problem of how finely day type and time period should be classified is solved as an optimization problem based upon an information-quantity criterion.
[Mid-Term Prediction]
Next, a mid-term prediction method will be described in detail using the travel-time realtime data (time-series data) and travel-time transition patterns.
In mid-term prediction, it is known empirically that one of the properties of travel time is that “if gridlock starts earlier, then the travel-time transition pattern will hasten correspondingly”, and that another property is that “if travel time at a certain time is longer than usual, then a similar tendency will persist for a while”.
Such a fluctuation conforms well to a period of from 30 minutes, which is the scope of a mid-term prediction, to one or two hours. The travel-time prediction apparatus 100 according to this example uses a prediction method that formulates the above-mentioned findings.
If we assume for the sake of simplicity that either the road-segment unit (link) or day type is fixed and that the travel-time realtime data 101 travel-time transition patterns are one-dimensional time-series data comprising only one attribute “travel time”, then travel time at time t found from past data that has been stored in the travel-time transition pattern database 104 can be expressed by f(t). Further, assume that the present time is t0. Now travel time can be predicted by the prediction function
h(t|a,b)=af(t−b)
in which a and b are conversion parameters. This prediction function is a function obtained by multiplying f(t) by a constant (by a factor of a) and translating it by (−b) so as to reduce the error relative to the realtime data, as illustrated in
It should be noted that â(t0) and {circumflex over (b)}(t0), which are obtained by the equation below that minimizes the error relative to the travel-time realtime data 101, are used as a and b, respectively.
Further, travel time can be predicted by the prediction function
h(t|a,b)=f(t−b)+a
in which a and b are conversion parameters. This prediction function is a function obtained by vertically displacing f(t) by (+a) and translating it by (−b) so as to reduce the error relative to the realtime data, as illustrated in
It should be noted that â(t0) and {circumflex over (b)}(t0), which are obtained by the equation below that minimizes the error relative to the travel-time realtime data 101, are used as a and b, respectively.
In Equations (1) and (2), exp[−α(t0−u)] is a weighting coefficient that multiplies the error [xu−h(u|a,b)]2 and that acts in such a manner that the more recent the data, the more importance is attached to it. That is, if we go back in time by 1/α step from the present time t0, the weight becomes a factor of 1/e. Therefore, if we consider a case where one step is five minutes, a conversion is made using data up to data that is several times 5/α minutes in the past.
The penalty-term coefficients wa and wb of the second and third terms on the right side of Equations (1) and (2) are parameters that control how easily the function conversion tends to affect the past data.
These variables α, wa, wb are all parameters that control the nature of learning and are referred to as “hyperparameters”. A specific value of α can be decided intuitively from 5/α*3=120, etc., in a case where one step is five minutes. Further, it will suffice if wa, wb are decided to the same extent as the variance of the travel time.
Travel time after time s can be found from the present time t0 by the equation below using the prediction function of Equation (1) or (2).
{circumflex over (T)}(t0+s)=h(t0+s|â(t0),{circle around (b)}(t0)) (Eq. 3)
With regard to the above-mentioned hyperparameters, it is possible to use values that have been optimized by the concept of the information-quantity criterion “predictive stochastic complexity”. Predictive stochastic complexity is put into concrete form by the equation below, where m represents the number of records of time-series data contained in 24 to 78 hours. It should be noted that the details of “predictive stochastic complexity” are described in Non-Patent Documents 2 and 3, by way of example, the entire disclosure thereof being herein incorporated by reference thereto.
Next, the travel-time prediction apparatus 100 reads out a travel-time transition pattern wt, which corresponds to the travel-time realtime data 101, specified link and time, from the travel-time transition pattern database 104 (step S102). The above-mentioned conversion parameters that specify the conversion of the travel-time transition pattern are calculated by the pattern conversion means 102 and are output to the predicted-value calculation means 103 (step S103).
Next, the travel-time prediction apparatus 100 outputs predicted values {circumflex over (x)}t+n, {circumflex over (x)}t+n+1, {circumflex over (x)}t+n+2, . . . using the prediction function obtained by the conversion employing the above-mentioned conversion parameters (step S104).
Thus, in accordance with this example, it is possible to estimate travel time accurately using a prediction function obtained by a conversion performed so as to reduce the error between past data and a present actually measured value with regard to a specified prediction-target link.
Next, a second example of the invention obtained by modifying the first example will be described in detail with reference to the drawings.
A travel-time pattern expressed by a step-shaped function with respect to the time axis is incapable of being differentiated. In order to find a combination of (a,b) that will minimize error, it is necessary to perform calculations using all combinations of (a,b) and to select the combination for which the error is smallest. This involves an enormous amount of calculation.
Accordingly, in this example, the processing (see step S103 in
The details of processing for calculating conversion parameters will be described with reference to
In order to make sequential input of data possible, the function F is obtained by approximately converting Equation (6) below, which is the error term and penalty terms of Equation (1). A feature of this conversion is that the travel-time transition pattern is not multiplied by a constant (by a factor of a) but by exp(a).
Σ(xu−h(u|a,b))2+wa(1−a)2+wbb2) (Eq. 6)
More specifically, the travel-time prediction apparatus 100 calculates the function F in the following five patterns to which provisional fluctuation ranges d1, e1 have been applied (added to or subtracted from)/not applied to initial conversion parameters (a1, b1), as described below:
(a1,b1)
(a1+d1,b1)
(a1b1+e1)
(a1−d1,b1)
(a1,b1−e1)
The travel-time prediction apparatus 100 randomly selects combinations of the constant-multiple parameter a and translation parameter b from the following nine combinations based upon a probability proportional to the size of error from the results of calculating the above-mentioned five patterns of function F, and adopts (a2, b2) as the selected combination:
(a1,b1)
(a1+d1,b1)
(a1b1+e1)
(a1−d1,b1)
(a1,b1−e1)
(a1+d1,b1+e1)
(a1+d1,b1−e1)
(a1−d1,b1+e1)
(a1−d1,b1−e1)
The travel-time prediction apparatus 100 repeats, m times (where m is set in advance in accordance with the processing capability, etc., of the travel-time prediction apparatus 100), calculation of the function F of a plurality of patterns to which the fluctuation ranges dn, en (n=1 to m) have been applied, as described above, and selection of provisional constant-multiple parameter an and provisional translation parameter bn (n=1 to m) that are based upon the results of the calculations (step S108), and narrows down the optimum (a, b) (step S109).
Here the fluctuation ranges dn, en (n=1 to m) are assumed to be d1≧d2≧ . . . dm, e1≧e2≧ . . . ≧em and are set in conformity with the required prediction accuracy of travel time in such a manner that the steps become progressively finer as the number m of computations increases.
In a case where prediction processing is executed again, t is updated by the operation t:=t+1 (step S105) in accordance with the flow of
At the processing (step S106) for reading in the travel-time realtime data at the next time t+1, only the data updated in the time period from time t to time t+1 is read in and calculation of the function F is performed using the latest q items of data inclusive of this data (step S107). As a result, the data read in is reduced and processing speed rises.
Further, prediction accuracy is maintained by thus sequentially inputting the latest q items of data without using old data (i.e., while forgetting the old data). As a result of the foregoing, high-speed processing is realized without using differentiation and by reducing the data that is read in.
According to this example, as described above, prediction of travel time is possible with respect to a road over a broad range with a diminished amount of calculation. This means that the system is readily installed in a vehicle in which a plurality of high-performance processing devices are difficult to install because of space limitations.
Next, a third example of the invention obtained by modifying the arrangement of the first example will be described in detail with reference to the drawings. The travel-time prediction apparatus according to this example is obtained by providing the arrangement of the first example with a plurality of prediction means, namely long-term prediction means and short-term prediction means, and with a high-speed prediction function for selecting the ideal prediction means from among these prediction means and performing real-time prediction in the appropriate cycle (five minutes to one hour). Primarily the additions to and modifications of the first example will now be described in detail.
Long-Term Prediction
The long-term prediction means 110 executes long-term prediction processing using only the stored data in the travel-time transition pattern database 104 and not the travel-time realtime data 101. The reason for this is that in traffic information, the influence of the present conditions on the future is several hours at most and hence the use of realtime data is meaningless with regard to predictions farther ahead than this.
Short-Term Prediction
The short-term prediction means 112 executes short-term prediction processing that is based upon an autoregression (AR) model. Here it is assumed that the short-term prediction is one that predicts a maximum of one hour ahead using the travel-time realtime data 101 of the past one hour. Although it is possible to use various methods in short-term prediction, it is preferred that use be made of the method described in Patent Document 3 filed by the present applicant, the entire disclosure thereof being incorporated herein by reference thereto.
An overview of the method described in Patent Document 3 that uses the autoregression model will be described below as it relates to the selection of prediction means, described later.
Let the difference yt between the travel-time realtime data and the travel-time transition pattern be expressed by yt=xt−wt. The autoregression model is a statistical model that defines a probability distribution produced by the travel-time realtime data. The model can be expressed as follows:
Here εt represents a noise term and is assumed generally to be a multidimensional normal distribution the average of which is zero. Further, am is referred to as an “AR coefficient”. In order to specify one of these models, it will suffice to specify all AR coefficients and a dispersion that defines the probability distribution of ε1. These parameters are referred to collectively as θ. If θ has been specified, then travel time into the immediate future can be predicted from the past data by the following equation:
[Selection of Prediction Processing]
The travel-time prediction apparatus 100 according to this example has a function for determining an appropriate prediction method for every link by utilizing the above-mentioned three types of prediction means and the acquired read-time data, and executing effective prediction processing using this method.
First, the period of time that is the target of each prediction is decided beforehand. For example, if the time is the present time t0, then the time period is the target of short-term prediction with regard to 1≦t≦t0+6, the time period is the target of mid-term prediction with regard t0+7≦t≦t0+25, and the value of travel-time transition pattern database 104 is output as is from then onward (long-term prediction).
If the time interval is five minutes, the above-mentioned rule means that short-term prediction is made from the present time to 30 minutes hence, mid-term prediction is made from then to 120 minutes hence, and long-term prediction is made from then onward.
The travel-time prediction apparatus 100 according to this example moreover determines whether to perform short-term and mid-term prediction or use the value from the travel-time transition pattern database 104 as is with regard to the period of time that is the target of short-term and mid-term prediction.
In a case where an autoregression model is used with regard to a short-term prediction, realtime data that goes back in time by the order of the autoregression model is necessary in order to carry out the prediction. For example, in a case where an autoregression model of order m is used, travel-time realtime data in a period corresponding to t0−m≦t≦t0 is required.
When the difference between the travel-time realtime data 101 in this period and the value from the travel-time transition pattern database 104 is large, the travel-time prediction apparatus 100 according to this example activates the short-term prediction algorithm; otherwise, the apparatus makes the prediction using the value from the travel-time transition pattern database 104 as is.
For example, in a case where the quantity indicated below is greater than a predetermined threshold value Δs, the apparatus makes the short-term prediction. Otherwise, the apparatus does not make the prediction.
It will suffice if the specific value of ΔS is determined by the required accuracy of travel time. For example, if an accuracy of one minute is required, then the value is made one minute, thereby enabling a travel time based upon the above-mentioned short-term prediction to be output only when necessary.
Similarly, with regard to mid-term prediction, travel-time realtime data in a period corresponding to t0 t0−1/α≦t≦t0 is required. In this case also, whether it is necessary to execute the mid-term prediction or not can be determined depending upon whether a quantity obtained by substituting 1/α for m in Equation (7) is larger than a predetermined value ΔM. It will suffice if ΔM also is determined by accuracy in a manner similar to ΔS. However, since a mid-term prediction generally cannot be expected to have an accuracy higher than that of a short-term prediction, setting ΔM to be several times larger than ΔS (e.g., to five minutes) is appropriate.
By thus setting ΔS and ΔM appropriately, the computation cost involved in prediction processing can be controlled.
[Grouping of Prediction Processing]
By way of example, it can be expected that travel-time realtime data relating to two successive links on the same road will have statistical properties having a high degree of resemblance in many cases. The same is true with regard to links on two parallel roads. In particular, when the difference between travel-time realtime data and a travel-time transition pattern is considered, road-specific properties are smoothed out and a greater degree of correlation can be expected. The travel-time prediction apparatus 100 according to this example subjects a set of links to clustering beforehand based upon a value from the travel-time transition pattern database 104 and groups links that indicate similar tendencies.
Further, the apparatus decides a single representative link with regard to each group. If conversion parameters └â(t0),{circumflex over (b)}(t0)┘ used in mid-term prediction are found with regard solely to this representative group, then it will be possible for the apparatus to make a prediction regarding a link belonging to the group. This is advantageous, particular for mid-term prediction, in two points, namely the fact that it is possible to make a prediction also with regard to a link for which realtime data is not obtained at the present time (this in turn essentially makes it possible to apply predictions to roads throughout the entire country), and in that computation time can be curtailed.
It is necessary that this clustering be performed with regard to all links to undergo prediction. However, since there is considered to be no correlation between links that are geographically remote from each other, it will suffice to execute processing only in a geographical region that has been formed into a block. For example, clustering can be facilitated by holding travel-time transition patterns in the form of a hierarchical structure (geographical_region/secondary_mesh/linkgroup/link/) that takes these geographical relationships into consideration. Further, thus managing travel-time transition patterns in the form of a hierarchical structure is advantageous in terms of load variance and expandability.
Further, the above-described clustering processing basically need only be executed one time as pre-processing and it need not be executed in realtime. As examples of specific clustering methods, use can be made of classical methods such as the Ward Method or k-means method [e.g., “A Survey of Recent Clustering Methods for Data Mining (part 1)—Try Clustering!—” by Toshihiro Kamishima, Artificial Intelligence Society Magazine, vol. 18, no. 1, pp. 59-65 (2003), and SOM (Self-Organized Map) proposed in the publication “Self-Organizing Maps” by T. Kohonen, Springer-Verlag, Berlin, 2001], the entire disclosure thereof being incorporated herein by reference thereto.
[Scheduling of Prediction Processing]
The operation (scheduling of prediction processing) of the travel-time prediction apparatus 100 according to this example will be described next.
Next, the travel-time prediction apparatus 100 periodically executes prediction-information update processing shown in
With reference to
The travel-time prediction apparatus 100 selects a representative link from a group to which the prediction-target link belongs (step S212).
If it has been determined at step S211 that a mid-term prediction is required, then the travel-time prediction apparatus 100 executes mid-term prediction processing (step S213). Similarly, if it has been determined at step S211 that a short-term prediction is required, then the travel-time prediction apparatus 100 executes short-term prediction processing (step S214).
Finally, the travel-time prediction apparatus 100 combines the results of the predictions and outputs the result of travel-time prediction that corresponds to the prediction-target link and prediction-target time (step S215).
According to this example, as described above, the advantages of short-, mid- and long-term predictions are combined, as set forth in the section “Selection of prediction processing”. This makes it possible to obtain prediction results in which a prescribed accuracy is assured with a small amount of computation. Further, as set forth in the section “Grouping of prediction processing”, it is also possible to make predictions regarding a route that includes a link (a segment of road) over which it is substantially impossible to obtain realtime data in view of circumstances such as cost.
Further, in terms of route selection and the provision of secondary information services to users, the highly accurate prediction data calculated as set forth above is useful information to individual drivers and to various transport companies such as trucking businesses, taxi companies and bus companies that transport tourists and goods.
It is possible to perform traffic information services using a traffic information providing system having means for providing results of travel-time prediction that have been output from the travel-time prediction apparatus 100 described above. Such information content can be distributed for a fee, in view of the utility thereof, by any billing system such as fixed payment system, in which a certain distribution period has been decided, or a pay-as-you-go system that conforms to the number of times information is distributed or to the size of distribution, etc. Alternatively, by distributing such information in combination with prescribed advertisements, it is possible to distribute the information for free if the commercial sponsor of the advertisements is made to bear the system running cost.
Furthermore, it is permissible to distribute not only the results of predicting travel time but also the above-mentioned conversion parameters with the addition of explanatory notes.
Though the present invention has been described in accordance with the foregoing examples, the invention is not limited to these examples and it goes without saying that the invention covers various modifications and changes that would be obvious to those skilled in the art within the scope of the claims.
It should be noted that other objects, features and aspects of the present invention will become apparent in the entire disclosure and that modifications may be done without departing the gist and scope of the present invention as disclosed herein and claimed as appended herewith.
Also it should be noted that any combination of the disclosed and/or claimed elements, matters and/or items may fall under the modifications aforementioned.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2006-286551 | Oct 2006 | JP | national |
| 2007-033769 | Feb 2007 | JP | national |
