CN111582559B

CN111582559B - Arrival time estimation method and device

Info

Publication number: CN111582559B
Application number: CN202010315673.1A
Authority: CN
Inventors: 姜正申
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-04-21
Filing date: 2020-04-21
Publication date: 2024-02-20
Anticipated expiration: 2040-04-21
Also published as: CN111582559A

Abstract

The application relates to the technical field of intelligent navigation and the technical field of maps, and discloses a method and a device for estimating arrival time, which are used for improving accuracy of arrival time estimation. The method comprises the following steps: determining a target route between a starting point and a terminal, and at least two target road sections contained in the target route; acquiring the overall characteristic value of the target route and the local characteristic value corresponding to the target road section; determining local estimated time of each target road section according to the local characteristic value; determining preliminary estimated time of the target route according to the local estimated time of all the target road sections; and determining the final estimated time of the target route according to the overall characteristic value and the preliminary estimated time of the target route.

Description

Arrival time estimation method and device

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to a method and an apparatus for estimating arrival time.

Background

The navigation software provides the user with the function of planning a route and guides the user along the route by voice in the driving process of the user. The estimation of the arrival time (Estimated Time of Arrival, ETA) is a basic function in the map software, which is specifically to give the time required to complete a route determined on the map for that route and departure time.

At present, ETA is estimated generally by using a machine learning algorithm, and the characteristic value of the whole route is input into the machine learning algorithm for training and prediction. However, in some cases, for example, the length of the route is long, the road condition difference between different road sections is large, or a certain road section is extremely congested, and when the arrival time of the whole route is greatly affected, the accuracy of the arrival time estimation is low only by considering the whole characteristic of the route.

Disclosure of Invention

The embodiment of the application provides a method and a device for estimating arrival time, which are used for improving the accuracy of arrival time estimation.

According to a first aspect of an embodiment of the present application, there is provided a method for estimating arrival time, including:

determining a target route between a starting point and a terminal, and at least two target road sections contained in the target route;

acquiring the overall characteristic value of the target route and the local characteristic value corresponding to the target road section;

determining local estimated time of each target road section according to the local characteristic value;

determining preliminary estimated time of the target route according to the local estimated time of all the target road sections;

And determining the final estimated time of the target route according to the overall characteristic value and the preliminary estimated time of the target route.

According to a second aspect of embodiments of the present application, there is provided an apparatus for estimating arrival time, the apparatus including:

a determining unit, configured to determine a target route between a starting point and a terminal, and at least two target road segments included in the target route;

the acquisition unit is used for acquiring the integral characteristic value of the target route and the local characteristic value corresponding to the target road section;

the computing unit is used for determining the local estimated time of each target road section according to the local characteristic value;

the calculation unit is further used for determining preliminary estimated time of the target route according to the local estimated time of all the target road sections;

the calculation unit is further used for determining final estimated time of the target route according to the overall characteristic value and the preliminary estimated time of the target route.

In an alternative embodiment, the computing unit is specifically configured to:

and carrying out weighted calculation on the local estimated time corresponding to all the target road sections to obtain the preliminary estimated time of the target route.

In an alternative embodiment, the computing unit is specifically configured to:

inputting the local characteristic value of the target road section into a trained first estimation model to obtain the local estimation time of the target road section;

and inputting the overall characteristic value and the preliminary estimated time of the target route into a trained second estimated model to obtain the final estimated time of the target route.

In an alternative embodiment, the determining unit is further configured to:

determining a limiting condition corresponding to the target route;

and determining a corresponding first pre-estimated model and a corresponding second pre-estimated model according to the limiting conditions.

In an alternative embodiment, the method further includes a training unit, configured to train to obtain the first estimated model and the second estimated model according to the following manner:

obtaining a training sample, wherein the training sample comprises an integral characteristic value of a training route, a local characteristic value of a training road section contained in the training route and an actual arrival time of the training route;

inputting the local characteristic value of the training road section contained in the training route into the first estimation model to obtain the preliminary estimation time of the training route;

calculating a first loss function according to the actual arrival time and the preliminary estimated time;

Inputting the preliminary estimated time and the overall characteristic value of the training route into the second estimated model, and outputting the final estimated time of the training route;

calculating a second loss function according to the actual arrival time and the final estimated time;

and when the first loss function is smaller than a first preset threshold value and the second loss function is smaller than a second preset threshold value, determining parameters corresponding to the first pre-estimated model and parameters corresponding to the second pre-estimated model.

In an alternative embodiment, the first and second loss functions are different types of loss functions.

In an alternative embodiment, the first pre-estimated model and the second pre-estimated model are all fully connected neural networks.

In an alternative embodiment, the computing unit is further configured to:

and inputting the overall characteristic value of the target route and the preliminary estimated time into a trained second estimated model to obtain the estimated time variance of the target route.

In an alternative embodiment, the training unit is further configured to:

inputting the preliminary estimated time and the overall characteristic value of the training route into the second estimated model, and outputting an estimated time variance;

And calculating the second loss function by using a maximum likelihood estimation method according to the actual arrival time, the final estimated time and the estimated time variance.

According to a third aspect of the embodiments of the present application, there is provided a computing device, including at least one processor, and at least one memory, wherein the memory stores a computer program, which when executed by the processor, causes the processor to perform the steps of the method for estimating arrival time provided by the embodiments of the present application.

According to a fourth aspect of the embodiments of the present application, there is provided a storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the steps of the method for estimating arrival time provided by the embodiments of the present application.

In the embodiment of the application, after the starting point and the ending point are determined, a target route between the starting point and the ending point and at least two target road sections contained in the target route are determined, and the overall characteristic value of the target route and the local characteristic value corresponding to the target road sections are obtained. And determining the local estimated time of each target road section according to the local characteristic value, and determining the preliminary estimated time of the target route according to the local estimated time of all the target road sections. Here, since the preliminary estimated time of the target route is determined according to the local estimated time of each target road segment, and the local estimated time of the target road segment is determined according to the local feature value, the local estimated time is determined by the local feature of the corresponding target road segment, and the preliminary estimated time is determined by the local features of all the target road segments. And finally, determining the final estimated time of the target route according to the overall characteristic value and the preliminary estimated time of the target route. Therefore, in the process of calculating the final estimated time, not only the whole characteristics of the route but also the local characteristics of each road section are considered, so that the accuracy of the estimated time is improved, and meanwhile, the method has higher accuracy and effect aiming at the special conditions of individual road sections, such as local congestion, and the application range of the whole estimated method is enlarged.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the description below are only some embodiments of the present application.

FIG. 1a is a schematic diagram of a prior art decision tree;

fig. 1b is a schematic structural diagram of a fully-connected neural network according to an embodiment of the present application;

FIG. 1c is a schematic diagram of a neuron according to an embodiment of the present application;

FIG. 2 is a system architecture diagram of a time-of-arrival estimation system in an embodiment of the present application;

FIG. 3 is a schematic diagram of a predicted time-length reachable circle according to an embodiment of the present application;

FIG. 4 is a flowchart of a method for estimating arrival time according to an embodiment of the present application;

fig. 5 is a schematic diagram of a possible interface of a terminal device in an embodiment of the present application;

fig. 6 is a schematic diagram of a target road segment included in a target route according to an embodiment of the present application;

FIG. 7 is a schematic flow chart of ETA estimation by using two algorithm models in the embodiment of the application;

FIG. 8 is a schematic diagram of a process for estimating arrival time based on a fully connected neural network in an embodiment of the present application;

FIG. 9 is a schematic flow chart of ETA and variance estimation using a fully connected neural network in an embodiment of the present application;

FIG. 10 is a block diagram illustrating a device for estimating arrival time according to an embodiment of the present application;

fig. 11 is a block diagram showing a structure of a server according to an embodiment of the present application.

Detailed Description

For the purposes of making the objects, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the technical solutions of the present application, but not all embodiments. All other embodiments, which can be made by a person of ordinary skill in the art without any inventive effort, based on the embodiments described in the present application are intended to be within the scope of the technical solutions of the present application.

The terms first and second in the description and claims of the invention and in the above-mentioned figures are used for distinguishing between different objects and not for describing a particular sequential order. Furthermore, the term "include" and any variations thereof is intended to cover non-exclusive protection. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those listed steps or elements but may include other steps or elements not listed or inherent to such process, method, article, or apparatus.

Some of the concepts involved in the embodiments of the present application are described below.

1. Artificial intelligence

Artificial intelligence (Artificial Intelligence, AI) is the theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and extend human intelligence, sense the environment, acquire knowledge and use the knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision. Artificial intelligence techniques mainly include computer vision techniques, speech processing techniques, machine learning/deep learning, and other directions.

2. Machine learning

Machine learning is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and the like. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning typically includes techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, induction learning, and the like.

3. Supervised machine learning model

Supervised learning is a machine learning task that extrapolates functions from a labeled training dataset. Specifically, an optimal model is obtained through training an existing training sample, all inputs are mapped into corresponding outputs by using the model, and the outputs are simply judged so as to achieve the purposes of prediction and classification, and the model has the capability of predicting and classifying unknown data. The data in the supervised learning is classified in advance, and the training sample of the data contains the characteristic and the label information, so that the corresponding output is obtained according to the characteristic and the label information.

Supervised learning generally includes both classification and regression types. The target variables of the classification problem are only valued in a limited target set (nominal type), such as handwriting digital recognition problem, and the target results are in the set {0, 1, 2, 3, 4, 5, 6, 7, 8, 9 }. The target variable of the regression problem is numerical, that is, the target variable can take a value from an infinite numerical set, such as predicting commodity price, and the predicted result can be any number.

4. Fully connected neural network

The network consists of all connection layers, and each node of the all connection layers is connected with all nodes of the upper layer and is used for integrating the features extracted from the front edge. Each output of the fully connected layer can be seen as each node of the previous layer multiplied by a weight coefficient, plus an offset value. One effect of the full join is a dimension transformation, in particular, that can change the high dimension to the low dimension while preserving useful information. The method is equivalent to a feature space transformation, useful information can be extracted and integrated, and a nonlinear mapping of an activation function is added, so that a plurality of layers of fully connected layers can simulate any nonlinear transformation theoretically. Another effect of full concatenation is the expression of latent semantics (empdding), mapping the original features to individual latent semantic nodes (hidden nodes), and for the last layer of full concatenation, the classified display expression. The parameters of the fully connected layer are also generally the most due to their fully connected nature.

5. Cloud technology

Cloud technology (Cloud technology) refers to a hosting technology for integrating hardware, software, network and other series resources in a wide area network or a local area network to realize calculation, storage, processing and sharing of data.

Cloud technology (Cloud technology) is based on the general terms of network technology, information technology, integration technology, management platform technology, application technology and the like applied by Cloud computing business models, and can form a resource pool, so that the Cloud computing business model is flexible and convenient as required. Cloud computing technology will become an important support. Background services of technical networking systems require a large amount of computing, storage resources, such as video websites, picture-like websites, and more portals. Along with the high development and application of the internet industry, each article possibly has an own identification mark in the future, the identification mark needs to be transmitted to a background system for logic processing, data with different levels can be processed separately, and various industry data needs strong system rear shield support and can be realized only through cloud computing.

6、ETA

The arrival time estimate (Estimated Time of Arrival, ETA) is a basic function in map software that is to give the time required to complete a route and departure time on a map.

7、ATA

In the historical data of the map service, the actual arrival time (Actual Time of Arrival, ATA) of a route can be extracted, so that the machine learning algorithm can be trained using this data as a sample value to obtain model parameters.

8. Variance of

Variance (Var) is a measure of the degree of discretization when the probability theory and statistical variance measure a random variable or set of data. The variance in probability theory is used to measure the degree of deviation between a random variable and its mathematical expectation (i.e., the mean). The variance in statistics (sample variance) is the average of the squared values of the differences between each sample value and the average of the population of sample values. In many practical problems, studying variance, i.e., the degree of deviation, is of great importance. In the statistical description, the variance is used to calculate the difference between each variable (observed value) and the overall mean. To avoid zero sum of mean deviation, the sum of squares of mean deviation is used statistically to describe the degree of variation of the variable, as affected by the sample content.

9. Route and road section

In map applications, a route is a complete line connecting a start point and an end point, and in a practical scenario, the length of a route is typically in the range of one kilometer to several tens of kilometers.

Routes are expressed in terms of sequences of road segments (links). In map data, a road is divided into segments of several tens of meters to several kilometers, each segment is called a road segment, and is assigned a globally unique ID. Thus, a route in a map is a sequence of all segments in the route.

In the embodiment of the present application, the method for dividing a route into multiple segments is not limited, for example, the route may be divided into multiple segments according to nodes in the route, and one segment is between adjacent nodes; the road segments may also be classified according to road class, i.e., a road segment in which the road class is unchanged in the route is used as a road segment. Further, the number of links to be included in one route is not limited.

10. Global and local features

The overall characteristics are characteristics related to the whole route and are used for representing the overall traffic situation of the route, such as the overall total journey distance, the overall average speed of the departure time, the total number of the overall traffic lights, the overall congestion mileage ratio and the like.

The local features correspond to the overall features, are features corresponding to the road sections, and are used for representing the traffic situation of each road section. For example, the length of the road segment, the road class of the road segment, the speed limit of the road segment, the road condition at the current time of the road segment, the road segment speed at the current time, and the like.

The definition of the local characteristics of different road sections in the route is the same, and the local characteristics are collected according to the set road section characteristics, or the local characteristics are obtained through calculation according to road network data, and the characteristic values of different road sections are different.

11. Local estimated time, preliminary estimated time and final estimated time

The local estimated time corresponds to the road section, is the estimated time required for completing the corresponding road section, and can be determined according to the local characteristic value of the road section.

The preliminary estimated time corresponds to the route and is determined according to the local estimated time. For example, the preliminary estimated time can be obtained by accumulating or weighting the local estimated time of all the road sections.

The final estimated time also corresponds to the route, which is the time taken to walk the entire route. The final estimated time is determined according to the overall characteristic value of the route and the preliminary estimated time of the route, and the final estimated time considers both the overall characteristic of the route and the local characteristic of each road section because the preliminary estimated time is determined according to the local characteristic value.

In the embodiment of the present application, the possible prediction models include a first prediction model, a second prediction model, and a third prediction model, and for convenience of understanding, explanation is made here.

The first estimation model is an algorithm model for determining local estimation time of the road section. And inputting the local characteristic value of each road section into a first estimation model, and outputting the local estimated time of the road section.

The second pre-estimated model is an algorithm model for determining the final pre-estimated time of the route. And inputting the preliminary estimated time and the overall characteristic value of the route into a second estimated model, and outputting the final estimated time of the route.

The third pre-estimation model is an algorithm model for determining the variance of the final pre-estimation time. And inputting the preliminary estimated time and the overall characteristic value of the route into a third estimated model, and outputting the variance of the final estimated time.

In the embodiment of the application, the first pre-estimated model and the second pre-estimated model are jointly trained and predicted by two algorithm models, or the first pre-estimated model, the second pre-estimated model and the third pre-estimated model are jointly trained and predicted by three algorithm models, and when the three algorithm models are jointly trained and predicted, the input data of the second pre-estimated model and the third pre-estimated model are the same. The first pre-estimated model, the second pre-estimated model and the third pre-estimated model are all supervised machine learning models. The first, second and third pre-estimated models may be the same supervised machine learning model but different model parameters, e.g., the first, second and third pre-estimated models are all fully connected neural networks. The first pre-estimated model, the second pre-estimated model and the third pre-estimated model can also be different supervised machine learning models, for example, the first pre-estimated model is a fully-connected neural network, the second pre-estimated model is a decision tree model, and the third pre-estimated model is a support vector machine model.

The basic idea of the present application is presented below.

In the related art, the ETA estimation is generally performed by using a tree model. Decision trees (Decision trees) are a basic classification and regression method, called classification trees when used for classification and regression trees when used for regression. The decision tree is composed of nodes and directed edges. Nodes are of two types: an internal node and a leaf node, wherein the internal node represents a feature or attribute and the leaf node represents a class. Generally, a decision tree comprises a root node, a plurality of internal nodes and a plurality of leaf nodes. Leaf nodes correspond to decision results, and each of the other nodes corresponds to an attribute test. The sample set contained in each node is divided into sub-nodes according to the result of the attribute test, the root node contains the sample set, and the path from the root node to each leaf node corresponds to a decision test sequence. In the decision tree shown in fig. 1a, circles represent internal nodes and boxes represent leaf nodes. Decision trees are divided into two categories according to the types of data processed: classification decision trees and regression decision trees. The former may be used to process discrete data and the latter may be used to process continuous data.

And estimating the arrival time by using a regression decision tree. The regression decision tree segments the feature space into m different subspaces, and calculates the output value (denoted as C) of each subspace from the training samples (the training samples falling into each subspace) _m ). After such a space is generated, it is convenient to map a sample to be measured to a certain subspace R _i In (1), subspace R _i Corresponding C _i As the output value of the sample to be measured. C (C) _m The value of (1) is generally obtained by adopting a mean value algorithm, namely, all the values falling in the subspace R are obtained _i As subspace R _i Is a value of (2).

Specifically, each node (not necessarily a leaf node) of the regression decision tree will obtain a predicted value of the arrival time that is equal to the average of the arrival times of all the routes belonging to that node. Each threshold value of each feature exhaustive of branching finds the best partition point. Wherein the best criterion is to minimize the mean square error, i.e. (arrival time per route-predicted arrival time) ² Divided by the number N of routes. That is, the more routes are predicted to be wrong, the greater the degree of the wrong is, the greater the mean square error is, and the most reliable branching basis can be found by minimizing the mean square error. Branching until arrival time corresponding to route on each leaf node is unique or reaches preset termination condition If the arrival time of the route on the final leaf node is not unique (such as the upper limit of the number of the leaves), the average arrival time of all routes on the node is taken as the predicted arrival time of the leaf node.

In the method, the data input into the regression decision tree are characteristic values of the whole route, such as the whole journey total journey, the whole journey average speed at the departure time, the total journey traffic light number, the whole journey congestion mileage ratio and the like. And inputting the integral characteristic value of the training route and the actual arrival time of the training route into a regression decision tree, and training to obtain the parameters of the regression decision tree. When predicting the target route, inputting the overall characteristic value of the target route into a trained regression decision tree to obtain the estimated arrival time of the target route.

The accuracy of the obtained arrival time is lower because the whole characteristic value of the route is only used as input data of the regression decision tree, namely, the arrival time estimation of the route only considers the whole characteristic of the route, but ignores the characteristic of each road section.

Based on the above, the embodiment of the application not only utilizes the integral characteristic value of the route, but also calculates according to the local characteristic value of each road section in the route, combines the integral characteristic value and the local characteristic value, and jointly predicts the arrival time of the route, thereby integrating the common influence of the local characteristic and the integral characteristic, improving the accuracy of prediction, and particularly aiming at individual road sections and scenes with large road condition differences of the road sections in the route, and obtaining good prediction results.

In the embodiment of the present application, the predictive model may be trained by using a supervised machine learning algorithm such as an artificial neural network, LR (logistic regression), KNN (K-nearest neighbors), SVM (support vector machine), GBDT (Gradient Boosting Decision Tree ) or FC (fully connected) neural network algorithm, deep fm (Deep Factorization Machine, depth factorizer), convolutional neural network, and the like, which are all single algorithm models constructed by using machine learning. On the basis of a single model, the embodiment of the application also combines two or more algorithm models into an algorithm of a pre-estimated model, and performs combined training or pre-estimation to achieve the effect of reducing variance (Bagging), deviation (Boosting) or improved prediction (Stacking), so that the accuracy of the pre-estimated model prediction is further improved.

The pre-estimated model in the embodiment of the application is a fully-connected neural network. A fully connected neural network is a network consisting of fully connected layers. Simple fully connected layer networks are generally divided into an input layer, a hidden layer, and an output layer. Of course, a classification layer, or a special processing layer such as a loss function, can also be added behind the output layer. For fully-connected neural networks, each node of each layer in the network is connected to each node of an adjacent network layer to integrate the features extracted from the previous edge. The parameters of the fully connected layer are also generally the most due to their fully connected nature.

The structure of the fully-connected neural network is shown in fig. 1b, and the network rule is as follows:

neurons are laid out in layers. The leftmost layer is called an input layer and is responsible for receiving input data; the rightmost layer is called the output layer from which neural network output data can be obtained. The layer between the input layer and the output layer is a hidden layer, which is not visible to the outside. The hidden layer in fig. 1b is 2 layers, and in this embodiment of the present application, the number of layers of the hidden layer is not limited.

There is no connection between neurons of the same layer. Each neuron of the N-th layer is connected to all neurons of the N-1 th layer, and the output of the N-1 th layer is the input of the N-th layer. Each connection has a weight. Neurons are the basic structures that make up a fully connected neural network, where the composition of one neuron is input, linearly weighted, activated function, and output. Fig. 1c shows a schematic structure of a neuron, and as shown in fig. 1c, the input of the neuron is an n-dimensional vector x, and the output is a. The formula for linear weighting is as follows:

wherein w is _i For inputting x _i B is the bias term for the neuron.

The activation function is h (x) and is a nonlinear function. The type of the activation function can be sigmoid (S-shaped growth curve) function, tan h (hyperbolic tangent) function, relu (Rectified Linear Unit, linear rectification function) activation function and the like, and the activation function introduces nonlinear factors to neurons, so that the neural network can be arbitrarily approximated to any nonlinear function, and the neural network can be applied to a plurality of nonlinear models.

The fully connected neural network uses matrix multiplication, which is equivalent to a feature space transformation, and can extract and integrate all feature information. In addition to the nonlinear mapping of the activation functions, the multi-layer fully connected neural network can theoretically simulate any nonlinear transformation.

In the embodiment of the application, the fully-connected neural network is selected as a first pre-estimated model, a second pre-estimated model and a third pre-estimated model. The input of the first estimation model is a local characteristic value of each road section, and the output is local estimation time of each road section. Calculating to obtain preliminary estimated time by using the local estimated time of each road section, taking the preliminary estimated time as one input of a second estimated model, taking the other input of the second estimated model as the integral characteristic value of the route, and outputting the integral characteristic value as the final estimated time of the route. The input of the third pre-estimated model is the same as the input of the second pre-estimated model, and the output is the variance of the final pre-estimated time.

It should be noted that, in the implementation of the present application, the algorithms of the first estimation model, the second estimation model and the third estimation model are only supervised machine learning models, and specific algorithm models are not limited, and the fully connected neural network in the embodiment is only an example and not limited. In addition, the type of the activation function is not limited, and preferably, a Relu function may be selected.

After the design concept of the embodiment of the present application is introduced, the application scenario set in the present application is briefly described below. It should be noted that the following scenario is only for illustrating the embodiments of the present application, and is not limiting. In specific implementation, the technical scheme provided by the embodiment of the application can be flexibly applied according to actual needs.

Fig. 2 is a schematic diagram of a system for estimating arrival time according to an embodiment of the present application. The application scenario includes the terminal device 201 and the server 202, where the terminal device 201 and the server 202 may be directly or indirectly connected through a wired or wireless communication manner, which is not limited herein.

The terminal device 201 is configured to send a request to a server, and may be, but not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, or a vehicle-mounted terminal. The in-vehicle terminal is, for example, an in-vehicle terminal in an automatic delivery vehicle, or an in-vehicle navigation device in a user vehicle, or the like. The terminal device 201 is provided with a client, and the client may be a web client, a client installed in the terminal device 201, a light application embedded in a third party application, or the like.

The server 202 is a server corresponding to a client in the terminal device 201, may be an independent physical server, may be a server cluster or a distributed system formed by a plurality of physical servers, and may also be a cloud server for providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, basic cloud computing services such as big data and an artificial intelligent platform, and is applied to cloud electronic map products to cope with processing demands of big data of electronic maps.

In a cloud technology based implementation, the server 202 may process the road network data through cloud computing and cloud storage.

Cloud computing (clouding) is a computing model that distributes computing tasks across a large pool of resources, enabling various application systems to acquire computing power, storage space, and information services as needed. The network that provides the resources is referred to as the "cloud". Resources in the cloud are infinitely expandable in the sense of users, and can be acquired at any time, used as needed, expanded at any time and paid for use as needed. The cloud computing resource pool mainly comprises: computing devices (which are virtualized machines, including operating systems), storage devices, network devices.

Cloud storage (cloud storage) is a new concept that extends and develops in the concept of cloud computing, and a distributed cloud storage system (hereinafter referred to as a storage system for short) refers to a storage system that integrates a large number of storage devices (storage devices are also referred to as storage nodes) of various types in a network to work cooperatively through application software or application interfaces through functions such as cluster application, grid technology, and a distributed storage file system, so as to provide data storage and service access functions for the outside.

In a possible implementation manner, road network data are stored in a cloud storage mode, when an estimated model needs to be trained, training samples are obtained from a storage system corresponding to the cloud storage, the estimated model is trained by the training samples, when the estimated model is trained, the estimated model can be obtained through a machine learning algorithm such as a fully-connected neural network, and at the moment, calculation tasks are distributed in a large number of resource pools in a cloud calculation mode, calculation pressure is reduced, and meanwhile training results can be obtained. When the arrival time of the route is required to be estimated, the integral characteristic value and the local characteristic value are obtained from a storage system corresponding to the cloud storage, the integral characteristic value and the local characteristic value are utilized for estimating, the estimating process of the arrival time can be carried out through a machine learning algorithm, at the moment, calculation tasks are distributed in a large number of resource pools in a cloud computing mode, the calculation pressure is reduced, and meanwhile, the estimating result can be obtained.

A scenario in which the estimation process of the arrival time is applicable is described below.

The first usage scenario:

when the user needs to plan a route using navigation software, the terminal device 201 may be operated, a start point and an end point may be input in an offline map of the terminal device 201, or the start point and the end point may be directly selected, or the like. The terminal device 201, in response to a user's operation, may determine a plurality of candidate routes between the start point and the end point based on the information, calculate the estimated arrival time of each candidate route, and then select one route with the shortest estimated arrival time from among them for display to the user. In addition, if the client of the electronic map is installed in the terminal device 201, the client may also send start point and end point data to the server 202 in response to the operation of the user, the server 202 determines a plurality of candidate routes according to the start point and the end point, calculates the estimated arrival time of each candidate route, and then selects a route with the shortest estimated arrival time from the calculated arrival times, and sends the route back to the client.

Second usage scenario:

after the client in the terminal device 201 enters the navigation state, i.e. the route and the destination point are determined, the terminal device 201 sends the positioning position to the server 202 every predetermined time, the server 202 takes the received positioning position as a starting point, calculates the expected arrival time corresponding to the current moment according to the route and the destination point, further calculates the driving time required by the remaining route and feeds back the driving time to the terminal device 201, so that the user can acquire the remaining driving time, and the user can conveniently arrange the route.

Third usage scenario:

the terminal 201 sends the start point input by the user to the server 202, and the server 202 may calculate the end point position in each route after the estimated time according to the received position as the start point, thereby forming the circle with the estimated time length. Fig. 3 shows a schematic diagram of a circle of estimated duration. As shown in fig. 3, after traveling along 4 different routes in fig. 3 for half an hour, each of the two routes can reach a position a, b, c, d, and the points a, b, c, d are sequentially connected to obtain a half-hour reachable circle as shown in fig. 3. The server 202 may send the estimated time period reachable circle to the terminal 201, or send the information in the circle to the terminal 201 based on the estimated time period reachable circle, so as to facilitate the user to know the situation around the location o.

Fourth usage scenario:

when taking out the take-out orders, the take-out background server 202 can calculate route time consumption corresponding to each take-out order according to the position of the merchant as a starting point and the position of the customer as an end point, so that the take-out order can be better taken out for the distributor according to the route time consumption, the take-out result is optimized, and the delivery efficiency is improved.

Fifth usage scenario:

When a user uses a client to drive, the terminal device 201 sends the current position to the server 202, the server 202 takes the current position of each taxi driver as a starting point, the current position of the user is used as a terminal, and the time length for each taxi driver to drive to the position of the user is calculated, so that the taxi driver order receiving is better arranged, and the passenger transport efficiency is improved.

It is noted that the above-mentioned application scenario is only shown for the convenience of understanding the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. Rather, embodiments of the present application may be applied to any scenario where applicable.

The following describes a method for estimating the arrival time according to the embodiment of the present application with reference to an application scenario shown in fig. 2.

Referring to fig. 4, an embodiment of the present application provides a method for estimating arrival time, as shown in fig. 4, including:

step S400: the terminal equipment acquires the starting point and the ending point and sends the starting point and the ending point to the server.

In a specific implementation process, the terminal device obtains the start point and the end point, which may be that a user operates the terminal device, inputs an identifier of the start point and an identifier of the end point, such as a location name, in a client of the terminal device, and the terminal device or the server 202 determines a specific position of the start point and a specific position of the end point based on the identifier. As shown in fig. 5, the user may input the identifier a of the start point at the "please input the start point" position and input the identifier B of the end point at the "please input the end point" position, and the server may determine the start point and the end point, calculate the form time of each route, and recommend an optimal route to the user.

In another possible embodiment, the user may also directly select a start point and an end point in the map displayed by the client, the user clicks on the start point position in the map, and the client determines the start point in response to the clicking operation by the user.

In another possible embodiment, the terminal device is provided with a positioning unit, which may be a satellite positioning chip, and has a satellite communication function. The terminal equipment directly communicates with the positioning satellite through the positioning unit, so that the collected positioning information can be collected. The terminal equipment sends the positioning information collected in real time to the server, so that the server can also receive and acquire the positioning information of the terminal equipment. At this time, the user may take the current position as a start point or take the current position as an end point, so that the server determines a specific start point or end point.

Step S401: the server determines a target route between the start point and the end point, and at least two target segments contained in the target route.

In the implementation process, after the server determines the starting point and the ending point, a plurality of routes between the starting point and the ending point can be determined. The server may determine the final estimated time of each route with each route as a target route, or the server may select one route from a plurality of routes as a target route based on a preset rule. Or the server sends the route to the terminal device, and the user selects one route as a target route based on the display of the terminal device.

Each road segment corresponds to a unique ID, so that the target route consists of a sequence of road segments of indefinite length. In the embodiment of the application, the division manner of the road sections in the route is not limited, for example, the road sections can be divided according to the intersections, namely, the route sections between adjacent intersections are used as a road section; or for example, the road segments may be divided according to the number of lanes contained in the corresponding road, i.e., a continuous route with a constant number of lanes is used as a road segment; or the road grade can be classified according to the road grade, namely, a continuous route with unchanged road grade is used as a road section. Fig. 6 shows a schematic diagram of a target road segment comprised by a target route. As shown in fig. 6, a point G is a start point, a point K is an end point, a dotted line between the start point and the end point is a target route, and the target route is divided into a link GH, a link HI, a link IJ, and a link JK according to an intersection through which the target route passes.

Step S402: the server acquires the overall characteristic value of the target route and the local characteristic value corresponding to the target road section.

In the implementation process, the whole characteristics of the whole route comprise: road information such as total length of the whole journey, average speed limit of the whole journey, average free flow speed of the whole journey and the like, average speed of the whole journey at a departure time calculated according to GPS (Global Positioning System ) data acquired in real time, average speed of the whole journey 5 minutes (10 minutes, 15 minutes and the like) before and after the same time mined according to historical GPS data acquired in the past several months and the like.

The local features of the road segment include: the length of the current road section, the road grade, the free flow speed, the speed limit, the road condition at the current moment, the speed of the current moment and the like.

The server may obtain the required feature values from the road network data. After the road network data is obtained, the server needs to perform conventional data preprocessing such as data screening, normalization processing and the like on the road network data, and details are not repeated here. And the server obtains a corresponding characteristic value according to the preprocessed road network data.

Step S403: for each target road section, the server determines the local estimated time of the target road section according to the local characteristic value.

In the specific implementation process, the local estimated time of each target road section can be calculated by using a basic rule, for example, the simplest local estimated time of the target road section is calculated directly through a speed calculation formula according to the length of the target road section and the free flow speed, namely, the speed is equal to the displacement divided by the time. Of course, the accuracy of the calculation mode is low, and preferably, in the embodiment of the application, the machine learning model is utilized to calculate the local estimated time, namely, the local characteristic value of the target road section is used as the input of the machine learning model, and the local estimated time of the target road section is output through the model.

Step S404: and the server determines the preliminary estimated time of the target route according to the local estimated time of all the target road sections.

In the specific implementation process, the local estimated time corresponding to all the target road sections is subjected to weighted calculation, and the preliminary estimated time of the target route is obtained.

The preliminary estimated time of the target route is obtained by weighting calculation for the local estimated time of all road sections, and the road section sequence with indefinite length is converted into the preliminary estimated time, so that the indefinite number of outputs are converted into the only one factor affecting the final estimated time, namely the preliminary estimated time, and the input characteristic quantity is ensured to be fixed length when the final estimated time is calculated.

For example, if the target route includes 4 target segments, there are 4 local estimated times corresponding to the target route; if the target route includes 6 target road segments, 6 local estimated times are corresponding, that is, the number of road segments included in different target routes is different. And for a target route, weighting calculation is carried out on all local estimated time to obtain a preliminary estimated time, and the target route corresponds to an influence factor, so that the calculation is conveniently carried out by using the factor subsequently.

For calculation using the machine learning model, since the weight of each road segment has been considered when the local estimated time is calculated by the estimated model, the weights of all the target road segments can be accumulated without weighting here, i.e. the weights of all the target road segments are 1 at this time.

Step S405: and the server determines the final estimated time of the target route according to the overall characteristic value and the preliminary estimated time of the target route.

Step S406: and the server sends the final estimated time to the terminal equipment.

In the embodiment of the application, the arrival time is estimated by using a supervised machine learning algorithm, and in an alternative embodiment, a local characteristic value and a global characteristic value can be input into an algorithm model by using the algorithm model, so that the estimation of the arrival time integrates the influence of the local characteristic and the global characteristic.

In practical cases, different algorithm models need to be used under different limiting conditions. For example, the arrival times of different types of vehicles are different in the same route, and in general, the arrival time of an automobile is shorter and the arrival time of a bicycle is longer than that of a bicycle. For another example, in the case of a sunny day and a rainy day, the arrival time of the sunny day is generally shorter, and the arrival time of the rainy day is longer. Therefore, according to the embodiment of the application, the parameters of the algorithm models corresponding to different limiting conditions are different, and the corresponding algorithm models need to be determined according to the limiting conditions.

Inputting the local characteristic value of the target road section into the trained first pre-estimated model, and before obtaining the local pre-estimated time of the target route, further comprising:

determining a limiting condition corresponding to the target route;

In the implementation process, the limiting condition can be that the terminal equipment responds to the operation of the user and sends the limiting condition to the server. For example, if the user selects the vehicle as the car in the client, the terminal device sends the constraint that the car is sent to the server, and the server determines a corresponding algorithm model for the car according to the vehicle. Or the defined condition may be measured directly for the terminal device or the server. For example, the terminal device determines the current moving speed through the positioning device, compares the moving speed with a threshold value, and determines the corresponding relation between the current moving speed and the vehicle.

Another problem in the above solution is that the number of road segments included in different routes may be different, so that the accuracy of model calculation cannot be guaranteed by directly inputting the local feature value into the algorithm model. Thus, embodiments of the present application utilize at least two algorithmic models for joint prediction.

In an alternative embodiment, the first predictive model and the second predictive model are utilized in combination for prediction. The input of the first estimation model is a local characteristic value of the target road section, and the output is local estimation time of the target road section. That is, step S403, determining the local estimated time of the target road section according to the local feature value, includes:

And inputting the local characteristic value of the target road section into the trained first estimation model to obtain the local estimation time of the target road section.

In the implementation process, one target route comprises a plurality of target road segments, local characteristic values of the plurality of target road segments can be respectively input into the first prediction model for prediction, and local characteristic values of all the plurality of target road segments can be completely input into the first prediction model for prediction.

For example, the target route L includes a target link 1, a target link 2, and a target link 3. 10 local feature values of the target link 1, 10 local feature values of the target link 2, and 10 local feature values of the target link 3 are acquired. Sequentially inputting the local characteristic values of three target road sections into a first pre-estimation model, namely inputting 10 local characteristic values of the target road section 1 into the first pre-estimation model to obtain the local pre-estimation time of the target road section 1; then 10 local characteristic values of the target road section 2 are input into a first pre-estimated model to obtain local pre-estimated time of the target road section 2; and then 10 local characteristic values of the target road section 3 are input into a first estimation model to obtain the local estimation time of the target road section 3. Or a plurality of first pre-estimated models are arranged in the distributed system, the local characteristic values of the three target road sections are respectively input into the three first pre-estimated models, and the three first pre-estimated models respectively output three local pre-estimated times. Parameters among the three first pre-estimated models are shared, so that the three first pre-estimated models are the same algorithm model.

After determining the local estimated time of the target road section, because the number of road sections contained in different target routes may be different, a plurality of local estimated times cannot be directly input into the second estimated model, but all the local estimated times are weighted and calculated and then are input into the second estimated model as a characteristic value. Step S403, determining a final estimated time of the target route according to the overall feature value and the preliminary estimated time of the target route, including:

Fig. 7 is a schematic flow chart of ETA estimation using two algorithm models in the embodiment of the application. The solid arrow process is a process of estimating the arrival time. As can be seen from fig. 7, the second pre-estimation model inputs the overall characteristics of the primary ETA and the route, so as to integrate the influence of the local characteristics of the road section and the overall characteristics of the route on the arrival time, thereby improving the accuracy and the application range of calculating the final ETA.

In the implementation process, a plurality of integral characteristic values of the target route can be determined according to the road network data. And taking the plurality of integral characteristic values and the preliminary estimated time obtained by calculation according to the output of the first estimated model as the input of the second estimated model. Here, the first estimation model and the second estimation model may be the same algorithm model, or may be different kinds of algorithm models. For example, the first pre-estimated model and the second pre-estimated model are all fully connected neural networks; or the first pre-estimated model is a support vector machine, and the second pre-estimated model is a decision tree model.

In the embodiment of the application, the first pre-estimated model and the second pre-estimated model are jointly pre-estimated to obtain the final pre-estimated time, so that the training process of the first pre-estimated model and the second pre-estimated model is also joint training. In an alternative embodiment, during training, the final estimated time is optimized, that is, the actual arrival time of the training route is compared with the final estimated time, and a loss function is calculated, where the objective function is as follows:

argmin∑ _i loss(ETA _final ) … … equation 1

Where i corresponds to the number of training routes, i.e. the ith training route, loss (ETA _final ) As a loss function, the absolute value loss can be taken, namely:

loss(ETA _final )＝|ETA _final ATA| … … equation 2

Wherein ETA _final For the final estimated time of the training route, ATA is the actual arrival time of the training route. The final estimated time is calculated by the first estimated model and the second estimated model, so that the parameters of the first estimated model and the parameters of the second estimated model in the training process are required to be adjusted, and the error of the final estimated time is minimized.

Further, in the training process, not only the final estimated time is optimized, but also the local estimated time is optimized. The embodiment of the application trains and obtains a first estimated model and a second estimated model according to the following modes:

Acquiring a training sample, wherein the training sample comprises an integral characteristic value of a training route, a local characteristic value of a training road section contained in the training route and actual arrival time of the training route;

inputting the local characteristic value of the training road section contained in the training route into a first pre-estimation model to obtain the preliminary pre-estimation time of the training route;

inputting the preliminary estimated time and the integral characteristic value of the training route into a second estimated model, and outputting the final estimated time of the training route;

In the specific implementation process, the error of the final estimated time is calculated according to the formula 2, so that the final estimated time is optimized, and the local estimated time is optimized. Here, for ease of calculation, the preliminary estimated time is compared with the actual arrival time to optimize the local estimated time. For example, a first loss function of the first estimation model is calculated according to the actual arrival time and the preliminary estimation time, and the first loss function is expressed as follows:

loss(ETA _local )＝(ETA _local -ATA) ² … … equation 3

Wherein ETA _local The preliminary estimated time of the training route is obtained. The first loss function uses a square loss here, and alternatively, the first loss function may be in the form of an absolute value loss, a Huber (smoothed average absolute error) loss, or the like, which is not limited herein.

On the other hand, a second loss function of the second estimation model is calculated according to the final arrival time and the actual arrival time, and the formula of the second loss function may be as formula 2. The second loss function uses absolute value loss here, alternatively the first loss function may also be in the form of square loss, huber loss, etc., without limitation. Preferably, in order to increase the diversity and robustness of the model for the accuracy of the time estimate of interest from different sides, the first and second loss functions are different types of loss functions. For example, the first loss function is an absolute value loss and the second loss function is a square loss.

Therefore, the first loss function and the second loss function are combined, and the total objective function of the first pre-estimation model and the second pre-estimation model in the embodiment of the application can be obtained. The objective function is as follows:

argmin∑ _i [loss(ETA _local )+loss(ETA _final )]… … equation 4

From the above description, it can be seen that the more accurate local estimated time can make the final estimated time more accurate. Therefore, in the embodiment of the application, not only the error of the final estimated time is considered, but also the error of the preliminary estimated time is considered, namely the error of the local estimated time is considered, so that the accuracy of the algorithm model is further improved.

In the specific training process, a back propagation algorithm is utilized to train the first pre-estimated model and the second pre-estimated model. The back propagation algorithm is a common training method, and will not be described here in detail. And adjusting model parameters of the first pre-estimated model and the second pre-estimated model through a back propagation algorithm, so that the first loss function is smaller than a first preset threshold value, the second loss function is smaller than a second preset threshold value, and storing the model parameters at the moment to obtain the first pre-estimated model and the second pre-estimated model. Or adjusting model parameters of the first pre-estimated model and the second pre-estimated model through a back propagation algorithm, so that the value of an objective function, namely formula 4, is minimized, and the first pre-estimated model and the second pre-estimated model are obtained.

Taking the fully-connected neural network as an example of the first estimation model and the second estimation model, introducing the estimation process of the arrival time. Fig. 8 shows a schematic diagram of a process for estimating arrival time based on a fully connected neural network. As shown in fig. 8, the target route includes 6 segments, and local feature values of the 6 target segments are respectively obtained, where the number of local features of the target segments is not limited, for example, each target segment may correspond to 9 local feature values, each target segment may correspond to 10 local feature values, and local features between different target segments are the same. For each target link, all local feature values of the target link are input into one fully connected neural network (FC in fig. 8).

One target segment corresponds to one fully connected neural network. The parameters are shared between these fully connected neural networks and are therefore in fact the same fully connected neural network. Here, the structure of the fully-connected neural network may be as shown in fig. 1b, in which the number of hidden layers may be set according to the actual situation. For example, in the case where the number of training samples is millions of data, the hidden layer may be set to 3-4 layers.

The FC outputs the local estimated time (ETA in fig. 8) of the target road section, and sums the local estimated times to obtain the preliminary estimated time of the target route. Then, the preliminary estimated time and the overall characteristic value of the target route are spliced together, and the total fully-connected neural network (total FC in FIG. 8) is input, so that the final estimated time (total ETA in FIG. 8) can be obtained. Here, the structure of the total fully-connected neural network may also be a structure as shown in fig. 1b, where the number of hidden layers may be set according to the actual situation.

Further, in order to simultaneously estimate the variance of the arrival time and estimate the uncertainty, the implementation of the present application may also slightly modify the second estimation model, so as to calculate the variance. At this time, step S404, after determining the preliminary estimated time of the target route according to the local estimated time of all the target road segments, further includes:

And inputting the overall characteristic value and the preliminary estimated time of the target route into a trained second estimated model to obtain the estimated time variance of the target route.

In this case, the second pre-estimation model may be any one of supervised machine learning algorithms such as an artificial neural network, LR, KNN, SVM, GBDT or fully connected neural network algorithm, deep fm, convolutional neural network, etc.

In the embodiment of the application, the second estimation model for calculating the variance is described by taking a fully connected neural network as an example. Fig. 9 is a schematic flow chart of the joint estimation of the final estimated time and variance by using the first estimated model (FC 1 in fig. 9) and the second estimated model (FC 2 in fig. 9) in the embodiment of the application. As shown in fig. 9, the input data of the second prediction model is unchanged, that is, the overall feature value and the preliminary prediction time of the target route are also used as the input data of the modified second prediction model, and at this time, the output data of the second prediction model is the final prediction time and the variance of the final prediction time (in fig. 9, the total ETA and Var).

Further, aiming at the modified second pre-estimated model, the embodiment of the application carries out joint training on the first pre-estimated model and the second pre-estimated model. The specific training mode is as follows:

Inputting the preliminary estimated time and the integral characteristic value of the training route into a second estimated model, and outputting the estimated time variance;

calculating the second loss function based on the actual arrival time and the final estimated time includes:

and calculating a second loss function by using a maximum likelihood estimation method according to the actual arrival time, the final estimated time and the estimated time variance.

In the implementation process, a method of maximum likelihood estimation may be used to calculate a second loss function corresponding to the second estimation model.

Specifically, let y denote ATA of the training route, μ denote the final estimated time of the training route, σ denote the standard deviation of the final estimated time of the training route, and Var denote the variance of the final estimated time of the training route. Since μ and Var are both predicted by the second predictive model and are thus model parameter dependent functions, μ (θ) and Var (θ) can also be written, where θ represents the model parameters.

Assume that:

then, given N independent co-distributed data points, the log likelihood function is:

thus, maximizing the log-likelihood function is equivalent to minimizing the following objective function:

for a training route, the second loss function corresponding to the second pre-estimated model is as follows:

Thus, for the first and second pre-estimated models, the overall objective function is as follows:

and adjusting model parameters of the first pre-estimated model and the second pre-estimated model so that the first loss function is smaller than a first preset threshold value and the second loss function is smaller than a second preset threshold value, and storing the model parameters at the moment to obtain the first pre-estimated model and the second pre-estimated model. Or the first pre-estimated model and the second pre-estimated model are adjusted to minimize the value of the objective function, namely the formula 9, so as to obtain the first pre-estimated model and the second pre-estimated model.

The final arrival time and the variance of the target route can be calculated simultaneously through the algorithm model obtained through training, so that the uncertainty of the final arrival time can be determined simultaneously. For example, if the calculated final arrival time is 600 seconds and the variance is 100 seconds squared, the final arrival time is (600±10) seconds.

The above-described flow is described in detail below with reference to specific embodiments, which follow.

And the terminal equipment responds to the operation of the user and sends the specific running mode, the positioning position at the current moment and the end position to the server.

The server determines a plurality of routes according to the positioning position and the end position.

For each route, the server predicts the ETA for that target route with that route as the target route.

The specific prediction process is as follows:

and determining the overall characteristic value of the target route according to the road network data, wherein the target route comprises P target road sections and the local characteristic value of each target road section.

And determining a first pre-estimated model and a second pre-estimated model according to the specific driving mode.

And inputting the local characteristic value into a first estimation model aiming at the local characteristic value of each target road section, and calculating to obtain the local estimation time of the target road section.

And accumulating all the local estimated time to obtain the preliminary estimated time of the target route.

And inputting the overall characteristic value and the preliminary estimated time of the target route into a second estimated model, and calculating to obtain the final estimated time of the target route.

And the server sends the multiple routes between the positioning position and the terminal position and the final estimated time corresponding to each route to the terminal equipment. Therefore, the user can select a route which can reach the end point fastest according to the final estimated time.

The arrival time estimation method is applied to the mobile phone map APP product, estimates driving time for a user, and can effectively improve estimated accuracy. Specific predicted effect pairs are shown in table 1:

TABLE 1

	Traditional method for lifting decision tree based on gradient	Method for estimating arrival time of application
			MAPE	12.83％	12.7％
MPE	-0.91％	0.2％
			Yield of good quality	78.33％	79.10％
Poor evaluation rate	4.03％	3.81％

Among them, MAPE (Mean Absolute Percent Error, average absolute percent error) and MPE (Mean Percentage Error, average percent error) are two commonly used measures of predictive effect. MAPE measures the accuracy of the prediction, and the lower the value, the more accurate the prediction is; MPE measures the deviation of the predicted result, and a value greater than zero indicates that the predicted result is larger, and a value smaller than zero indicates that the predicted result is smaller, and the index is closer to 0 and indicates that the predicted result deviation is smaller. The excellent rate and the poor evaluation rate are two indexes defined based on user experience, the excellent rate can reflect the satisfaction degree of users, and the poor evaluation rate reflects the injury degree of the users.

As can be seen from Table 1, the present application has a boost in each index compared to the conventional gradient boost decision tree. Not only is the MAPE index lower, but the MPE is also closer to 0. From the aspects of the excellent rate and the poor evaluation rate, the method and the device can remarkably improve user experience. Compared with the traditional ETA estimation method, the estimation method of the arrival time in the embodiment of the application can improve the prediction accuracy and reduce the prediction deviation at the same time.

The following is an embodiment of the device of the present application, and for details of the device embodiment that are not described in detail, reference may be made to the foregoing one-to-one method embodiment.

Referring to fig. 10, a block diagram of a device for estimating arrival time according to an embodiment of the present application is shown. The cross-chain data processing apparatus is implemented as all or part of server 202 in fig. 2 by hardware or a combination of hardware and software. The device comprises: a determining unit 101, an acquiring unit 102, a calculating unit 103, and a training unit 104.

A determining unit 101 for determining a target route between a start point and a terminal, and at least two target segments included in the target route;

an obtaining unit 102, configured to obtain an overall feature value of a target route and a local feature value corresponding to a target road segment;

a calculating unit 103, configured to determine, for each target road segment, a local estimated time of the target road segment according to the local feature value;

the calculating unit 103 is further configured to determine a preliminary estimated time of the target route according to the local estimated times of all the target road segments;

the calculating unit 103 is further configured to determine a final estimated time of the target route according to the overall feature value and the preliminary estimated time of the target route.

In an alternative embodiment, the calculating unit 103 is specifically configured to:

In an alternative embodiment, the determining unit 101 is further configured to:

determining a limiting condition corresponding to the target route;

In an alternative embodiment, the training unit 104 is further configured to train to obtain the first estimated model and the second estimated model according to the following manner:

In an alternative embodiment, the first and second penalty functions are different types of penalty functions.

In an alternative embodiment, the computing unit 103 is further configured to:

In an alternative embodiment, the training unit 104 is further configured to:

Referring to fig. 11, a block diagram of a server according to an embodiment of the present application is shown. The server 1100 is implemented as the server 202 in fig. 2. Specifically, the present invention relates to a method for manufacturing a semiconductor device.

The server 1100 includes a Central Processing Unit (CPU) 1101, a system memory 1104 including a Random Access Memory (RAM) 1102 and a Read Only Memory (ROM) 1103, and a system bus 1105 connecting the system memory 1104 and the central processing unit 1101. The server 1100 also includes a basic input/output system (I/O system) 1106, which helps to transfer information between various devices within the computer, and a mass storage device 1107 for storing an operating system 1113, application programs 1114, and other program modules 1115.

The basic input/output system 1106 includes a display 1108 for displaying information and an input device 1109, such as a mouse, keyboard, or the like, for user input of information. Wherein both the display 1108 and the input device 1109 are coupled to the central processing unit 1101 through an input-output controller 1110 coupled to the system bus 1105. The basic input/output system 1106 may also include an input/output controller 1110 for receiving and processing input from a number of other devices, such as a keyboard, mouse, or electronic stylus. Similarly, the input output controller 1110 also provides output to a display screen, a printer, or other type of output device.

The mass storage device 1107 is connected to the central processing unit 1101 through a mass storage controller (not shown) connected to the system bus 1105. Mass storage device 1107 and its associated computer-readable media provide non-volatile storage for server 1100. That is, mass storage device 1107 may include a computer-readable medium (not shown) such as a hard disk or CD-ROM drive.

Computer readable media may include computer storage media and communication media without loss of generality. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes RAM, ROM, EPROM, EEPROM, flash memory or other solid state memory technology, CD-ROM, DVD or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the ones described above. The system memory 1104 and mass storage device 1107 described above may be collectively referred to as memory.

According to various embodiments of the present application, the server 1100 may also operate by a remote computer connected to the network through a network, such as the Internet. That is, the server 1100 may be connected to the network 1112 through a network interface unit 1111 connected to the system bus 1105, or the network interface unit 1111 may be used to connect to other types of networks or remote computer systems (not shown).

The memory also includes one or more programs, one or more programs stored in the memory, the one or more programs including instructions for performing the estimation of arrival times provided by embodiments of the present application.

Those skilled in the art will appreciate that all or part of the steps in the method for estimating arrival time according to the above embodiments may be implemented by a program for instructing relevant hardware, and the program may be stored in a computer readable storage medium, where the storage medium may include: read Only Memory (ROM), random access Memory (RAM, random Access Memory), magnetic or optical disk, and the like.

The foregoing embodiment numbers of the present application are merely for describing, and do not represent advantages or disadvantages of the embodiments.

The foregoing description of the preferred embodiments of the present application is not intended to limit the invention to the particular embodiments of the present application, but to limit the scope of the invention to the particular embodiments of the present application.

Claims

1. The method for estimating the arrival time is characterized in that a system to which the method is applied comprises a terminal device and a server, and the method comprises the following steps:

the terminal equipment acquires a starting point and a terminal point and sends the starting point and the terminal point to the server;

the server determines a target route between a starting point and a destination point and at least two target road sections contained in the target route;

the server acquires the integral characteristic value of the target route and the local characteristic value corresponding to the target road section; wherein the overall characteristic value includes: the total length of the whole journey, the average speed limit of the whole journey, the average free flow speed of the whole journey, the average speed of the whole journey at the departure time calculated according to the GPS data acquired in real time, and the average speed of the whole journey within preset time before and after the same time mined according to the historical GPS data; the local feature values of the road segments include: the length of the current road section, the road grade, the free flow speed, the speed limit, the road condition at the current moment and the speed at the current moment;

Determining a limiting condition corresponding to the target route; determining a corresponding trained first pre-estimated model and a corresponding trained second pre-estimated model according to the limiting conditions;

for each target road section, the server inputs the local characteristic value of the target road section into the trained first pre-estimated model to obtain the local pre-estimated time of the target road section, wherein the first pre-estimated model is a fully-connected neural network;

the server performs weighted calculation on the local estimated time corresponding to all the target road sections to obtain the preliminary estimated time of the target route;

the server inputs the overall characteristic value and the preliminary estimated time of the target route into a trained second estimated model to obtain the final estimated time of the target route, wherein the second estimated model is a decision tree model;

meanwhile, the server inputs the overall characteristic value and the preliminary estimated time of the target route into a trained third estimated model, so that the estimated time variance of the target route is obtained while the final estimated time of the target route is obtained, and the third estimated model is a support vector machine model;

and the server determines uncertainty of the final arrival time according to the final estimated time and the estimated time variance.

2. The method of claim 1, wherein the first predictive model and the second predictive model are trained in accordance with:

3. The method of claim 2, wherein the first and second loss functions are different types of loss functions.

4. The method of claim 1, wherein the first and second predictive models are both fully connected neural networks.

5. The method of claim 1, wherein the training to obtain the first predictive model and the third predictive model further comprises:

inputting the preliminary estimated time and the overall characteristic value of the training route into the third estimated model, and outputting an estimated time variance;

the calculating a third loss function according to the actual arrival time and the final estimated time includes:

and calculating the third loss function by using a maximum likelihood estimation method according to the actual arrival time, the final estimated time and the estimated time variance.

6. An apparatus for estimating time of arrival, the apparatus comprising:

a determining unit, configured to obtain a start point and an end point through a terminal device, and determine a target route between the start point and the end point, and at least two target road segments included in the target route;

the acquisition unit is used for acquiring the integral characteristic value of the target route and the local characteristic value corresponding to the target road section; wherein the overall characteristic value includes: the total length of the whole journey, the average speed limit of the whole journey, the average free flow speed of the whole journey, the average speed of the whole journey at the departure time calculated according to the GPS data acquired in real time, and the average speed of the whole journey within preset time before and after the same time mined according to the historical GPS data; the local feature values of the road segments include: the length of the current road section, the road grade, the free flow speed, the speed limit, the road condition at the current moment and the speed at the current moment; determining a limiting condition corresponding to the target route; determining a corresponding trained first pre-estimated model and a corresponding trained second pre-estimated model according to the limiting conditions;

The computing unit is used for inputting the local characteristic value of each target road section into a trained first estimated model to obtain the local estimated time of the target road section, wherein the first estimated model is a fully-connected neural network;

the calculation unit is further used for carrying out weighted calculation on the local estimated time corresponding to all the target road sections to obtain the preliminary estimated time of the target route;

the calculation unit is further used for inputting the overall characteristic value and the preliminary estimated time of the target route into a trained second estimated model to obtain the final estimated time of the target route, wherein the second estimated model is a decision tree model; meanwhile, inputting the overall characteristic value and the preliminary estimated time of the target route into a trained third estimated model, so that the estimated time variance of the target route is obtained while the final estimated time of the target route is obtained, wherein the third estimated model is a support vector machine model; and simultaneously determining uncertainty of the final arrival time according to the final estimated time and the estimated time variance.

7. The apparatus of claim 6, further comprising a training unit configured to train to obtain the first predictive model and the second predictive model according to:

8. A computer device, comprising:

at least one processor, and

a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor, the at least one processor implementing the method of any one of claims 1-6 by executing the memory stored instructions.

9. A computer readable storage medium storing computer instructions which, when run on a computer, cause the computer to perform the method of any one of claims 1 to 6.