Distributed Multi-Task Machine Learning for Traffic Forecasting

Abstract

A traffic prediction method based on distributed machine learning is provided for predicting traffic on a road, where the traffic prediction method based on distributed machine learning includes: distributing a global multi-task traffic model by a learning server to a learning agent to locally train a traffic model, uploading the locally trained traffic model to the learning server by the learning agent, updating the global multi-task traffic model by the learning server using locally trained traffic model parameters obtained from the learning agent, generating a time-dependent global traffic map by the learning server using the well-trained global multi-task traffic model, distributing the time-dependent global traffic map to vehicles traveling on the road, and calculating an optimal driving route with the shortest driving time by the vehicles using the time-dependent global traffic map based on a driving plan.

Description

[Technical field]

[0001] The present invention relates generally to machine learning for vehicular traffic systems, and more specifically to a method and apparatus for distributed multi-task machine learning for vehicular traffic prediction and its application to route planning. [Background technology]

[0002] With the rapid increase in connected and autonomous vehicles, intelligent transportation systems are becoming increasingly important for smart cities. Unlike traditional vehicles, connected and autonomous vehicles are much more intelligent. Not only can they collect various vehicle and traffic data, but they can also run advanced algorithms to guide their own mobility. In addition to this, much more traffic data is collected by data collectors and mobile devices in the transportation infrastructure. For example, the California Department of Transportation's Performance Measurement System (PeMS) has installed hundreds of thousands of data collectors across the state to collect various traffic data. Each data collector collects traffic data every five minutes. In addition, mobile devices such as smartphones can also collect traffic data and provide crowdsourced information. As a result, a huge amount of traffic data is available. It is crucial to leverage vehicle intelligence and abundant traffic data to improve driving safety, travel time, energy efficiency, air pollution reduction, and more.

[0003] However, realizing intelligent traffic is a very difficult problem. Physical roads form a complex road network. Most importantly, traffic conditions such as congestion in one place may propagate and affect traffic conditions in other places. In addition, unexpected events such as traffic accidents and driver behaviors may make traffic conditions even more dynamic and uncertain. Therefore, how to accurately predict traffic and apply the prediction to route planning is difficult but demanding because route planning is probably the most popular application in vehicles. However, a combination of existing technologies cannot solve this problem.

[0004] There are existing route planning methods. However, existing route planning methods that rely on crowdsourcing information may make misleading decisions. For example, if one vehicle is carrying 100 mobile devices, the crowdsourcing-based route planning method may consider the vehicle to be 100, resulting in a much higher traffic density. As a result, the corresponding road may be considered congested. In addition to this, existing route planning methods perform route planning based on current traffic information without making predictions. However, the vehicle traffic system is a highly dynamic system. Traffic conditions change from moment to moment. Therefore, an intelligent traffic prediction method is needed to perform optimal route planning. The intelligent traffic prediction method relies heavily on machine learning techniques.

[0005] Machine learning techniques are being applied in vehicle mobility management. For example, autonomous vehicles utilize different types of sensors to collect data and then use different machine learning algorithms to learn and analyze the data to control and guide the vehicle's movements. In connected vehicles, machine learning techniques can be used in infrastructures such as the cloud to achieve centralized learning and thus remotely control vehicle mobility.

[0006] To plan a route, a prediction technique must be able to perform short-term, medium-term, and long-term predictions. However, on-board learning in autonomous vehicles and centralized learning in the cloud are not practical for route planning. First, machine learning algorithms need to be trained with a large amount of different data types. Vehicles can only see scenes within their own field of view, and cannot see objects occluded by objects such as buildings and other vehicles, even if the objects are close to the vehicle, so vehicles can only collect limited data. The limited data can train and predict instantaneous local traffic for instantaneous motion planning. However, instantaneous predictions are not suitable for longer-term route planning. Second, although the cloud has the potential to collect a sufficient amount of data to train machine learning algorithms, there are several challenges. That is, 1) cloud or centralized learning relies on communication, but since communication bandwidth is limited, it is impractical for data collectors such as vehicles to send all data to the cloud; 2) data privacy policies may prevent data collectors from transferring data to the cloud because machine learning algorithms can learn the privacy of data collectors and personal information of drivers; and 3) security policies may prevent data collectors from sending their own data to the cloud or a centralized server, which may allow, for example, a security attacker to intercept the data, identify the driver's location, and take dangerous actions.

[0007] Efficient transportation requires efficient road utilization. To do so, route planning algorithms must perform optimal route planning to minimize traffic congestion and reduce travel time. To this end, traffic prediction techniques are crucial for efficient transportation. As mentioned above, traditional centralized learning and the emerging in-vehicle learning cannot perform feasible traffic prediction. Existing route planning methods cannot plan optimal routes because they do not perform traffic prediction. Therefore, it is desirable to provide an accurate and practical vehicle traffic prediction mechanism to perform optimal route planning for intelligent transportation.

[0008] Therefore, there is a need to provide a method for accurate vehicle traffic prediction and a route planning method that uses the predicted traffic to plan an optimal route. Summary of the Invention

[0009] One objective of some embodiments is to provide a distributed machine learning based vehicle traffic prediction method that can accurately predict short-term, medium-term and long-term traffic on both city roads and highways. In addition, another objective of some embodiments is to use accurate traffic predictions for optimal route planning of vehicles to minimize traffic congestion and travel times.

[0010] Some embodiments are based on the recognition that vehicle traffic data is widely collected by data collectors and vehicles in traffic infrastructure. How to utilize the available data to optimize vehicle traffic becomes a challenge to be addressed. Due to facts such as communication bandwidth limitations, privacy protection, and security, it is impractical to transfer all data to a central server for centralized data analysis and traffic prediction. Meanwhile, with the limited amount of data of individual data collectors, it is not feasible to train machine learning algorithms and make large-scale traffic predictions for a city or state. This is because a data collector does not know the traffic conditions of other locations. For example, a vehicle cannot predict traffic in a location where the vehicle has not yet traveled. Therefore, traditional machine learning approaches are not suitable for making traffic predictions for optimal route planning.

[0011] To that end, some embodiments of the present invention utilize distributed machine learning techniques, such as federated learning, to build a robust traffic model for accurate traffic prediction. Here, infrastructure devices, such as IEEE DSRC / WAVE Roadside Units (RSUs) and / or 3GPP C-V2X eNodeBs and / or remote servers, act as learning servers, and data collectors act as learning agents. The learning server coordinates distributed learning among a set of data collectors. First, the learning server designs and distributes a traffic model, such as a neural network, to a set of data collectors for a first round of distributed training. Then, each data collector independently trains the received traffic model by using its own data without sharing it with other data collectors and the learning server. After a certain number of iterations of the training process, each data collector sends the trained traffic model to the learning server, which then aggregates the traffic models received from all data collectors to generate a common global traffic model. Once the model aggregation is completed, the learning server redistributes the aggregated traffic model to the data collectors for a second round of training process. This training and aggregation process continues until a robust traffic model is constructed.

[0012] In a vehicular environment, it must be recognized that even though the data collectors are distributed at different geometric locations, the traffic data collected between the data collectors may be correlated since the traffic conditions at one location may propagate to and affect the traffic conditions at other locations. Thus, data collected at one location may also affect data collected at other locations. In addition to this, the traffic patterns at different locations may be different. For example, the traffic pattern at an intersection is different from the traffic pattern on a highway.

[0013] To that end, it is desirable to have different but correlated traffic models trained by distributed data collectors. To that end, some embodiments are based on the recognition that multi-task distributed learning techniques are suitable for predicting large-scale traffic for route planning. Thus, the learning server designs different but correlated traffic models to be distributed to data collectors, so that the traffic models of closer data collectors have a closer relationship. Taking a traffic model based on neural networks as an example, closer data collectors may share some weight parameters. However, the neural network models of data collectors far away from each other do not share parameters.

[0014] Some embodiments are based on the recognition that data collectors are distributed in different locations and therefore reflect location information. In addition to location factors, there are other factors that can affect the traffic environment, such as time, weather, road conditions, and special events. Traffic conditions in the same location vary at different times, in different weather, etc. Traffic conditions during rush hours are very different from traffic conditions during off hours. Traffic conditions on a snowy day are very different from traffic conditions on a sunny day.

[0015] For that purpose, it is desirable for the data collectors to divide their data into different clusters based on time, weather, etc. Therefore, the learning server defines a set of rules and distributes the rules to the data collectors to cluster their data. As a result, the data collectors train different traffic models by using different data clusters. The data collectors do not train traffic models for which the data collectors do not have suitable data. Therefore, the data collectors send only the trained traffic models to the learning server.

[0016] Thus, the learning server builds a common global traffic model by aggregating locally trained traffic models, taking into account information including location, time, weather, etc.

[0017] According to some embodiments of the present invention, there may be provided a distributed machine learning based traffic prediction method for predicting traffic on a road. In this case, the distributed machine learning based traffic prediction method may be a computer-implemented distributed machine learning based traffic prediction method for predicting traffic on a road. The method may include distributing a global multi-task traffic model by a learning server to a learning agent to locally train the traffic model, and uploading the locally trained global multi-task traffic model to the learning server, the locally trained global multi-task traffic model being trained by the learning agent, and the method may further include updating the global multi-task traffic model by the learning server using the locally trained global multi-task traffic model uploaded from the learning agent, generating a time-dependent global traffic map by the learning server using the updated global multi-task traffic model, and distributing the time-dependent global traffic map to each of the vehicles traveling on the road.

[0018] Further, according to some embodiments of the present invention, a local traffic prediction agent may be provided for providing a locally trained traffic model to a learning server. The local traffic prediction agent may be a hardware device or software that may be referred to as a local traffic prediction agent stored in a device including at least one memory and at least one processor. The local traffic prediction agent may include an interface configured to collect local traffic data from sensors disposed on a road network, the interface configured to obtain a multi-task traffic model and data cluster rules from the learning server via a communication network, the local traffic prediction agent may further include a memory configured to store the local traffic data, the data cluster rules and the trained traffic model, a traffic prediction neural network, and a processor, the processor configured in association with the memory to locally train the traffic prediction neural network using the local traffic data based on the data cluster rules to update the obtained multi-task traffic model of the traffic prediction neural network, and to upload the updated locally trained multi-task traffic model to the learning server via the interface using the communication network.

[0019] And further, some embodiments of the present invention may provide a distributed machine learning based traffic prediction system for providing traffic predictions to vehicles traveling on a road network, said system may include at least one local traffic prediction agent as described above, at least one learning server as described above, and a communication network configured to connect the at least one local traffic prediction agent and the at least one learning server, at least one roadside unit, and vehicles traveling on the road network.

[0020] The at least one local traffic prediction agent may include an interface configured to collect local traffic data from sensors disposed on a road network, the interface configured to obtain a multi-tasking traffic model and data cluster rules from the learning server via a communication network, and the at least one local traffic prediction agent may further include a memory configured to store the local traffic data, the data cluster rules and a trained traffic model, a traffic prediction neural network, and a processor, the processor configured in association with the memory to locally train the traffic prediction neural network using the local traffic data based on the data cluster rules to update the obtained multi-tasking traffic model of the traffic prediction neural network, and to upload the updated locally trained multi-tasking traffic model to the learning server via the interface using the communication network.

[0021] The at least one learning server may include a transceiver configured to obtain trained multi-tasking parameters of a traffic prediction neural network from the local traffic prediction agent via a communications network, the local traffic prediction agent being disposed at a location on the road network, and the at least one learning server may further include a memory configured to store traffic data, a global time-dependent map, a traffic prediction neural network, a trained multi-tasking traffic model, and a map of the road network, and one or more processors, the processors configured in association with the memory to perform the steps of updating the traffic prediction neural network using the trained multi-tasking parameters, generating an updated global time-dependent traffic map based on the trained multi-tasking traffic model, distributing the updated global time-dependent traffic map to the vehicles traveling on the road network, and distributing data clustering rules to the local traffic prediction agents.

[0022] The system may include an input interface configured to update model parameters (learned model) of a traffic prediction neural network in a learning server by obtaining trained parameters from a learning agent via the input interface, wherein each learning agent is deployed at a location on the road network, and each learning agent is configured to train a multi-task traffic model by collecting traffic data (patterns) at the location where it is deployed, generate a global time-dependent traffic map based on the fully trained multi-task traffic model, determine a driving plan for vehicles traveling on the road network, and calculate an optimal route with the shortest travel time for vehicles based on the driving plan and the global time-dependent map.

[0023] Some embodiments are based on the recognition that each application in a vehicular environment has different requirements. Therefore, different techniques must be developed for each application. Route planning requires large-scale traffic predictions with different time horizons, including short-term, medium-term and long-term predictions.

[0024] Therefore, some embodiments of the present invention provide a multi-horizon traffic forecast, where traffic is forecasted in a multi-horizon in the time domain for each short-term forecast or medium-term forecast or long-term forecast. The forecast time horizon consists of multiple forecast periods. For example, the short-term horizon may consist of 5 forecast periods, the medium-time horizon may include 20 forecast periods, and the long-term horizon may consist of 50 forecast periods, where the forecast period represents a Δt time interval, e.g., if Δt=5 minutes, traffic is forecasted every 5 minutes. As a result, in the short-term horizon, traffic is forecasted 5 times, in the medium-time horizon, traffic is forecasted 20 times, and in the long-term horizon, traffic is forecasted 50 times. A longer time horizon provides more traffic forecasts, but a shorter time horizon provides a higher accuracy of traffic forecasts.

[0025] Some embodiments are based on the realization that route planning is about finding the optimal route to travel on a real road network, based on criteria such as travel time and energy consumption.

[0026] Therefore, some embodiments of the present invention formulate the route planning problem as an optimization problem of minimizing the travel time so that even other metrics such as energy consumption and driving comfort can be optimized. The real road map is converted into a time-dependent graph, in which a vertex is the intersection or connection point of any two adjacent road sub-segments, and an edge is a road sub-segment that connects two adjacent vertices. There is at least a data collector on each edge. Unlike traditional traffic graphs that are built based on road structure and distance, an edge may consist of multiple road segments, and most importantly, the length of an edge is the travel time on the edge. As a result, when the traffic situation changes, the length of the edge also changes, and thus the shape of the graph also changes.

[0027] Some embodiments are based on the recognition that there is uncertainty in the vehicle environment. Therefore, the traffic model must be trained to respond to unexpected events such as traffic accidents. It is impractical for the data collector to capture every type of unexpected event. However, the vehicle can capture these events as it travels along the road.

[0028] Thus, the route planning model and the traffic prediction model can interact with each other to enhance the real-time traffic model.

[0029] The presently disclosed embodiments are further described with reference to the accompanying drawings, which are not necessarily to scale, emphasis instead being placed generally on illustrating the principles of the presently disclosed embodiments. [Brief description of the drawings]

[0030] [Figure 1]FIG. 1 illustrates an intelligent vehicle transportation system according to some embodiments of the present invention. [Diagram 2] FIG. 2 illustrates an example of a data clustering method used to divide data from each data collector into clusters according to some embodiments of the present invention. [Figure 3A] FIG. 1 illustrates a traffic prediction architecture using distributed multi-task learning techniques according to some embodiments of the present invention. [Figure 3B] FIG. 2 illustrates a prediction time horizon with six prediction periods according to some embodiments of the present invention. [Figure 3C] FIG. 1 is a schematic diagram illustrating an example of a traffic prediction architecture including a learning server connected to distributed data collectors (learning agents) via a communication network (multiple communication networks) according to an embodiment of the present invention. [Figure 4] FIG. 2 illustrates a time-dependent graph for planning a route from a start point v1 to a destination point v12 according to some embodiments of the present invention. [Figure 5A] FIG. 2 illustrates exemplary edges and running time calculations of a time-dependent graph, where running time represents the length of an edge of the time-dependent graph, according to some embodiments of the present invention. [Figure 5B] FIG. 2 illustrates travel time calculations for road segments with short edges in a time-dependent graph such that a vehicle can traverse the road segment within one prediction period according to some embodiments of the present invention. [Figure 5C] FIG. 2 illustrates a travel time calculation for a road segment with long edges in a time-dependent graph, where the vehicle requires multiple prediction periods to traverse the road segment, according to some embodiments of the present invention. [Figure 6] FIG. 2 illustrates functional blocks and interactions between components of an intelligent transportation system according to some embodiments of the present invention. [Figure 7] FIG. 2 illustrates a multi-task federated learning algorithm for traffic speed prediction according to some embodiments of the present invention. [Figure 8A]FIG. 2 illustrates a modified A* algorithm for optimal route planning based on time-dependent graphs according to some embodiments of the present invention. [Figure 8B] FIG. 2 is a diagram of an optimal route calculation algorithm used by a modified A* algorithm according to some embodiments of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0031] Description of the embodiments

[0032] The following description provides exemplary embodiments only and is not intended to limit the scope, applicability, or configuration of the disclosure. Rather, the following description of exemplary embodiments will provide those skilled in the art with an enabling description for implementing one or more exemplary embodiments. It is contemplated that various changes may be made to the function and arrangement of elements without departing from the spirit and scope of the disclosed subject matter as set forth in the appended claims.

[0033] Specific details are given in the above description to provide a thorough understanding of the embodiments. However, those skilled in the art will appreciate that the embodiments can be practiced without these specific details. For example, systems, processes, and other elements in the disclosed subject matter may be shown as components in block diagram form in order to avoid obscuring the embodiments in unnecessary detail. In other instances, well-known processes, structures, and techniques may be shown without unnecessary detail in order to avoid obscuring the embodiments. Additionally, like reference numbers and names in the various drawings refer to like elements.

[0034] Also, individual embodiments may be described as a process that is depicted as a flowchart, a flow diagram, a data flow diagram, a structure diagram, or a block diagram. Although a flowchart may describe operations as a sequential process, many of the operations may be performed in parallel or simultaneously. In addition, the order of operations may be rearranged. A process may be terminated when its operations are completed, but may have additional steps not discussed or included in the diagram. Moreover, not all operations in any process specifically described may occur in all embodiments. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, or the like. When a process corresponds to a function, the end of the function may correspond to a return of the function to the calling function or to the main function.

[0035] Additionally, embodiments of the disclosed subject matter may be implemented, at least in part, either manually or automatically. The manual or automated implementation may be performed or at least assisted through machine, hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segments to perform the necessary tasks may be stored on a machine-readable medium. A processor(s) may perform the necessary tasks.

[0036] To facilitate the development of Intelligent Transportation Systems (ITS), it is essential to have accurate predictions of traffic conditions, such as traffic flow and traffic speed. This is because such knowledge can help drivers make effective driving decisions to alleviate traffic congestion, increase fuel efficiency, and reduce air pollution. With these promising benefits, traffic prediction can play a key role in ITS goals of advanced passenger information systems, advanced traffic management systems, and commercial vehicle operations.

[0037] To reap all the above benefits, traffic prediction must process real-time and historical traffic data and observations collected by data collectors and mobile devices. For example, induction loops can measure driving speed by reading inductance changes over time, and such data can be used for traffic speed prediction. In addition to this, the widespread availability of mobile devices (e.g., global positioning systems in vehicles and phones) allows for crowdsourcing of mobility data from the general public, further facilitating traffic data collection like never before. Such emerging big data can significantly increase data availability in terms of coverage and fidelity, significantly enhancing data-driven traffic prediction. Prior art on traffic prediction can be mainly divided into two categories. The first category focuses on using parametric approaches such as autoregressive integrated moving average (ARIMA) models and Kalman filtering models. When dealing with traffic that only exhibits regular changes, such as recurring traffic jams during morning and evening rush hours, parametric approaches can achieve promising prediction results. However, due to the stochastic and nonlinear nature of road traffic, traffic predictions using parametric approaches may deviate from actual values, especially for sudden traffic. So, instead of fitting traffic data to a mathematical model as done by parametric approaches, alternative methods use nonparametric approaches using machine learning (ML) based methods. For example, stacked autoencoder models can be used to learn the overall traffic flow characteristics for prediction. Long short-term memory (LSTM) recurrent neural networks (RNNs) can be used to predict traffic flow, speed and occupancy based on data collected by a data collector and its upstream and downstream data collectors. Along with the use of RNNs, convolutional neural networks (CNNs) can also be utilized to capture potential traffic evolution patterns in the underlying road network.

[0038] Prior art focuses on using advanced deep learning models for traffic prediction, but they all consider traffic changes with single-task learning (STL) models. In reality, traffic patterns on roads can vary greatly under different circumstances due to weather changes, road condition changes (such as road works and accidents), and special events (e.g., soccer matches and concerts). Therefore, using STL models cannot capture such diverse and complex traffic situations. Moreover, due to the limited on-chip memory available in data collectors and mobiles, local training data may be extremely insufficient and promising prediction performance cannot be achieved. In addition to this, prior art assumes that data collected by a data collector can be shared with other data collectors or with a centralized unit such as a data collector that accesses data from its upstream and downstream data collectors. However, the collected data may contain personal information such as vehicle license plates and mobile phone users' historical trajectories captured by cameras. In this case, sharing traffic data directly between data collectors may cause privacy issues. Meanwhile, communication costs are also one of the major concerns.

[0039] FIG. 1 illustrates an intelligent vehicle transportation system 100 and its components, as well as interactions between the components of the system, according to some embodiments of the present invention. The transportation system 100 includes a road network 105, distributed data collectors 110, a learning server 115, and vehicles 120. The distributed data collectors 110 may be referred to as local traffic prediction agents (or learning agents) 110. The data collectors 110 are learning agents and are deployed in the road network 105. The road network 105 consists of roads, roadside units, edge computing devices, etc. The vehicles 120 travel on the roads of the road network 105. The learning server 115 may be located at a remote location or along the road. The learning server 115 designs / updates traffic models, data clustering rules, and aggregates trained traffic models and traffic predictions. The learning server 115 distributes the data clustering rules to the data collectors 110 to divide the data of each data collector into different clusters via a communication network 112. The communication network 112 may be a wired or wireless network and is configured to connect any / all of the following: between the distributed data collector 110 and the learning server 115, between the learning server 115 and the vehicles 120, between the distributed data collector 110 and the road network 105, between the road network 105 and the vehicles 120, or a combination thereof. Furthermore, the communication network 112 may be a 4th generation core network, or a 5th generation core network, or a 5th generation or later core network, and may connect to roadside units (RSUs / DSRC (Dedicated Short Range Communications) transceivers) (not shown) located along roads or pedestrian walkways, or / and edge computing systems (not shown). In some examples, the roadside units may include cameras located on roads to monitor vehicles traveling on the roads, or cameras located at intersections to monitor pedestrians.

[0040] The learning server 115 also distributes the traffic model to the data collectors 110 for distributed training. Each data cluster is used to train a traffic model, for example, rush hour data is used to train a rush hour traffic model. The data collectors use their local data to train a traffic model and send the locally trained traffic model to the learning server to build a common global traffic model. The learning server 115 distributes the final traffic model to the data collectors 110 for traffic prediction. Each data collector predicts the traffic for the road segment where the data collector is located. The data collectors send their traffic predictions to the learning server, which then combines all traffic predictions to build a global traffic prediction for route planning. The global traffic prediction is distributed to the vehicles 120 for route planning. When the vehicles 120 travel the planned route, they can provide information such as traffic accidents to the learning server. The learning server 115 can then adjust the data collectors to update the traffic model. Multi-horizon traffic speed prediction

[0041]

number

[0042]

number

[0043] To address the insufficient local training data and protect data privacy, some embodiments of the present invention apply distributed machine learning techniques to solve problem (1). There are various distributed machine learning techniques, and some embodiments of the present invention use federated learning (FL) as an example to illustrate the distributed machine learning approach. Traffic data collected by a data collector may have strong spatial-temporal dependencies. To capture different traffic situations present in the collected traffic data, a multi-task FL model is provided, in which different learning models are designed for each traffic situation, e.g., the model for rush hour is different from the model for off-hour. Furthermore, these learning models may be correlated, e.g., off-hour traffic may affect rush hour traffic.

[0044] To facilitate multi-task learning, the training data is divided into different clusters, with each cluster corresponding to a learning model, for example, rush hour data is used to train a rush hour model. Data clustering is important for various reasons, for example, off-hour data is not desirable for training a rush hour traffic model, and local traffic data is not suitable for training a highway traffic model. There are various methods for data clustering. FIG. 2 shows a data clustering method 200 for traffic prediction, where location information is reflected in the location of the data collector 110. Different data collectors are located in different places. The data collector data 210 is divided into summer data 220 and winter data 230. This level of data clustering reflects the weather. Each of the summer data and winter data is divided into weekend data 240 and weekday data 250. This level of data clustering reflects the day of the week. Each of the weekend data 240 and weekday data 250 is further divided into rush hour data 260 and off-hour data 270 .

[0045]

number

[0046] To solve problem (2), the multi-task FL algorithm uses an iterative update scheme. Figure 7 shows a multi-task federated learning algorithm for traffic speed prediction according to some embodiments of the present invention.

[0047]

number

[0048] The global model parameters are then sent to the data collector for the next round of training. After each round of training, the next round of training is performed, and the same process is repeated between the training server and the data collector each time until the total loss function Fm(wm) for each cluster m is small enough.

[0049] FIG. 3A illustrates a traffic prediction architecture 300 using distributed multi-task learning techniques according to some embodiments of the present invention, where an infrastructure device acts as a learning server 115 to coordinate a set of data collectors 110 acting as learning agents. In FIG. 3A, there are N data collectors, where data collector 1 is a toll booth, data collector 2 is a loop detector, and data collector N is a camera. The learning server 115 designs M traffic models 310 such that each traffic model targets a specific task, e.g., a model for rush hour traffic and another for off-hour traffic, a model for city traffic and another for highway traffic. Thus, these traffic models are different but correlated, since traffic at one location may affect traffic at other locations and traffic at one time may affect traffic at another time. Thus, the learning server 115 defines data clustering rules 320 that divide the data into M clusters 330. The learning server 115 then distributes the rules 320 and traffic models 310 to the data collectors 110. Correspondingly, each data collector 110 divides its data into M clusters 330. Some of the data clusters may be empty. In that case, the data collector does not train the corresponding traffic model. In FIG. 3A, data collector 1 has all data clusters, data collector 2 has no cluster-M data, and data collector N has no cluster-1 data. As a result, data collector 1 trains all traffic models, data collector 2 does not train traffic model M, and data collector N does not train traffic model 1. After clustering its data, each data collector 110 starts the first distributed training. Each data collector 110 independently trains the received traffic model by using its own data without sharing it with other data collectors and the learning server. After some training iterations, each data collector sends the locally trained traffic model 340 to the learning server, which then aggregates the traffic models received from all data collectors to generate a common global traffic model 350.Once the model aggregation is complete, the learning server 115 redistributes the aggregated traffic model 350 to the data collectors for a second training run. This training and aggregation process continues until a robust traffic model is built. Once the final traffic model is built, the learning server 115 distributes it to the data collectors 110 for traffic prediction. The predicted traffic varies depending on the application and can be speed, traffic flow, number of specific vehicles, etc. For route planning applications, traffic speed is predicted. Finally, the data collectors send their local traffic predictions 360 to the learning server to build a global traffic prediction model, i.e., a time-dependent graph, which is then distributed to the vehicles 120 for route planning.

[0050] A traffic forecast is characterized by two parameters: the forecast time horizon and the forecast period. The forecast time horizon represents the furthest time for which traffic is forecasted, and the forecast period indicates how often traffic is forecasted within the forecast time horizon. The forecast period constitutes the forecast time horizon. Figure 3B shows an example of a forecast time horizon 370 consisting of six forecast periods 380 according to some embodiments of the present invention. A longer time horizon provides more traffic forecasts, but a shorter time horizon provides a more accurate traffic forecast, since a longer stretch of traffic further into the future needs to be forecast. This widening of the forecast time horizon necessarily leads to a decrease in forecast performance.

[0051] FIG. 3C is a schematic diagram illustrating an example of a traffic prediction architecture 300 including a learning server 115 connected to a distributed data collector (learning agent) 110 via a communication network (multiple communication networks) 112 according to some embodiments of the present invention.

[0052] The learning server 115 may be referred to as a traffic prediction server based on distributed machine learning. The learning server 115 is configured to provide map information related to traffic prediction to vehicles 120 traveling on the road network. The learning server 115 may include one or more processors 121, a memory 140, a memory unit / storage 200 configured to store a traffic prediction neural network 132, traffic data cluster rules 134, a global time-dependent map 135, a trained multi-task traffic model 136, and a global map of the road network 137, and an input interface (or transceiver) 150 configured to communicate with the learning agent 110 via the communication network 112 and update model parameters of the traffic prediction neural network 132. The trained multi-task traffic model 136 may be a traffic model 310 shown in FIG. 3A.

[0053] The learning server 115 is configured to update the parameters of the traffic predictive neural network 132 by obtaining the trained parameters of the trained multi-task traffic model 173, which has been trained by the learning agent 110, via the input interface 150 and the communication network 112. This updating process continues iteratively at predetermined elapsed periods.

[0054] In this case, each of the learning agents 110 is configured to be deployed at a location on the road network 105 and locally train a multi-task traffic model 174 by collecting traffic data (vehicle traffic patterns) at the deployed location and clustering the traffic data. Furthermore, the learning server 115 generates a global time-dependent traffic map 135 based on the updated trained multi-task traffic model 136 and distributes the global time-dependent traffic map 135 to the vehicles 120, which then distribute the modified A traffic map 135 shown in FIG. 8A and FIG. 8B. * Using an algorithm and a global time-dependent traffic map 135, the vehicle determines an optimal route based on its driving plan, which is the driving route on the road network 105 that takes the shortest travel time from the vehicle's current location to the destination.

[0055] Each of the learning agents 110 may include an interface / transceiver 151 configured to perform data communication with the learning server 115 via the communication network 112. Each learning agent 110 further includes one or more processors 160 and a memory 180 connected to a memory unit / storage 170 that stores traffic data 171, a traffic predictive neural network 173, a trained multi-task traffic model 174, a local map of the road network 175, and a local time-dependent map 172.

[0056] In some examples, a computer-implemented distributed machine learning based traffic prediction method may be provided for predicting traffic on a road by using one or more hardware including one or more processors in association with a memory / memory unit / storage storing instructions / programs, the instructions / programs causing the one or more processors to execute steps. The steps may include distributing the global multi-task traffic model 136 by the learning server 115 via the communication network 112 to the learning agents 110. Each of the learning agents 110 is configured to locally train the traffic model 136 (310) based on data signals acquired from the road sub-segments 190, the edge computing devices 185, and the vehicles 120 traveling on the road. The steps further include uploading / retrieving the locally trained traffic model 173 trained by the learning agent 110 from the learning agent 110 to the learning server 115, and updating the global multi-task traffic model 136 by the learning server 115 using locally trained traffic model parameters of the trained multi-task traffic model 174. The above steps further include generating a time-dependent global traffic map 135 by the learning server 115 using the fully trained global multi-task traffic model 136, distributing the time-dependent global traffic map 135 to each of the vehicles 120 traveling on the road, and calculating an optimal driving route with the shortest driving time for each of the vehicles 120 using the time-dependent global traffic map 135 based on the driving plan of each of the vehicles 120. Optimal route planning based on predicted travel time

[0057] Traffic speeds on the road network vary with time, for example, traffic speeds during rush hours are generally lower than traffic speeds during off-hours. Thus, by using the physical road network and predicted traffic speeds, the traffic map can be modeled as a time-dependent graph. To construct the time-dependent graph, the learning server uses multi-horizon speed predictions from traffic data collectors and divides road segments into multiple subsegments such that traffic on each road subsegment is predicted by a unique data collector located only on that subsegment. Then, the road network is modeled as a time-dependent graph G=(V,ε,W), where the set of vertices V includes the intersections and junctions of any two adjacent road subsegments, whereby the set of edges ε is the road subsegment connecting the two adjacent vertices, and W is the set of weights. For an edge e∈ε, the weight we(t)∈W is modeled as the travel time on edge e at time t, calculated as the ratio of the length of the road subsegment to the predicted speed. For example, a road sub-segment 190 in the road network 105 may be a road section on one road, or multiple connected road sections on multiple roads. Unlike a static graph where the weight associated with each edge is a constant value, its counterpart in the graph G is a time-varying variable due to the time-varying speed through each road sub-segment, e.g., piecewise linear speed as shown in FIG. 5C. Unlike a static graph that is constructed based on a road map and distance, vertices can be dynamically selected, so that they are not fixed points in the road network, as long as there is at least a data collector between any two vertices. Thus, an edge in the time-dependent graph may consist of multiple road segments. Most importantly, the length of an edge is not a physical distance, but a travel time on the edge. As a result, when traffic conditions change, the length of the edge also changes, e.g., the same edge length may be much longer during rush hour than during off-hours. Finally, the learning server distributes the time-dependent graph to vehicles within its communication coverage for route planning.

[0058] 4 illustrates an example of a time-dependent graph 400 according to some embodiments of the present invention, where vertices 410 are represented as v1, v2, etc., and between any two vertices there is a data collector 420, and each edge 430 has length w x,y (t x ) 440, where x indicates the start vertex, y indicates the end vertex, and time t x indicates the time when the vehicle leaves vertex x. In this figure, the vehicle leaves vertex v1 and arrives at vertex v 12 From vertex v1 to vertex v 12 It can be seen that there are many different routes to v1 and v2. Route planning is about finding the route with the minimum travel time. In this figure, route 450 represents a minimum travel time route, where a vehicle departs v1 at time t1 and arrives at v3 at time t3, departs v3 at time t3 and arrives at v7 at time t7, departs v7 at time t8 and arrives at v9 at time t9, and arrives at v1 at time t10. 10 ni v 10 and finally at time t 10 ni v 10 Departing at time t 12 Destination v 12 Arrives at Data Collector c 1,3 is located on the edge from v1 to v3, and the data collector c 3,7 is located on the edge from v3 to v7, and the data collector c 7,10 From v7 to v 10 Located on the edge of the data collector 10,12 is v 10 From v 12 It is located on the side of.

[0059]

number

[0060] FIG. 8A illustrates a modified A-Z diagram for optimal route planning based on a time-dependent graph according to some embodiments of the present invention. * 8A shows a modified A algorithm according to some embodiments of the present invention. * The optimal route calculation algorithm used by the algorithm.* The algorithm is designed to find the optimal route with the shortest travel time. In the search algorithm, the arrival time g v and the heuristic total running time l v , v∈V is g s =t s and l s =g s +h d (g s ), which is initially set to infinity. The heuristic total travel time is calculated by multiplying the arrival time and the heuristic travel time to the destination, h d The heuristic travel time to the destination, h d is calculated as the ratio of the Euclidean distance to the destination and the maximum speed. Then, the corrected A * The search process in the algorithm starts from the starting point s and extends to the neighboring vertices that have road sub-segments adjacent to s. For these neighboring vertices, the arrival time g is updated by comparing the most recently assigned arrival time with the arrival time when taking the route from the starting point s. Meanwhile, their heuristic total travel time l is also updated. Then, the vertex with the shortest heuristic total travel time among the nearby vertices is selected to continue the search process. The same process is repeated. Finally, when the destination point d is reached, the search process stops and returns the optimal route selection and its travel time estimate.

[0061] The key to route planning based on time-dependent graphs is to calculate the length of the route, i.e., the travel time on the route. Unlike existing route planning methods that use current traffic conditions to plan routes, the embodiment of the present invention uses traffic prediction to perform optimal route planning. For edges of the time-dependent graph 400, the corresponding data collector predicts the traffic speed every Δt period. Thus, the length of the edge, i.e., the travel time, is dynamically calculated using the traffic prediction. As a result, the length of the edges of the time-dependent graph changes as time changes, and thus the shape of the graph also changes, which indicates that the travel time on the route also changes with time.

[0062] There are various ways to calculate the travel time on a road segment. Figures 5A, 5B and 5C show a method for calculating the travel time by using a predicted piecewise linear traffic speed function according to some embodiments of the present invention. Figure 5A shows an edge e 500 of a time-dependent graph 400, whose starting vertex is v s 510, and the end vertex of this edge is v e 520. Edge e consists of two road segments R1 530 and R2 540. These two road segments are connected by a data collector 550 that predicts the traffic speed on both road segments R1 and R2. Road segment R1 is short such that a vehicle can traverse R1 in less than one Δt period, its physical distance being D1 560. Road segment R2 is long such that a vehicle needs more than one Δt period, for example seven Δt periods, to traverse R2, its physical distance being D2 570. If a vehicle travels past R2 at time t s At vertex v s Assume that the vehicle arrives at and passes through R1 within one forecast period Δt at time t s The expected traffic speed at is s0. Therefore, the travel time on the road segment R1 is D1 / s0. As a result, the vehicle c =t s +D1 / s0. Therefore, the length of edge ε of the time-dependent graph is w vs,ve (t s )580 is w vs,ve (t s )=w vs,c (t s )+w c,ve (t c ) can be calculated as w vs,c (t s ) = D1 / s0. w c,ve (t c ) is shown in FIG. 5C, where the vehicle travels at time t c It arrives at the data collector at time t c is assumed to be within the forecast period. c,ve (t c) is calculated using a piecewise rate function, where s1 is the time t c s is the predicted speed at time t c +αΔt, and s3 is the predicted speed at time t c +(α+1)Δt, s4 is the predicted speed at time t+(α+2)Δt, s5 is the predicted speed at time t+(α+3)Δt, s6 is the predicted speed at time t+(α+4)Δt, and finally s7 is the predicted speed at time t+(α+5)Δt, where the vehicle travels only βΔt in the last prediction period (where 0<β<1). Thus, the travel time w on road segment R2 is c,ve (t c ) is equal to (α+5+β)Δt. Therefore, the mileage in the first prediction period is d1=αΔt * s1, and the mileage in the second prediction period is d2 = Δt * s2, and the mileage in the third prediction period is d3 = Δt * s3, and the mileage in the fourth prediction period is d4 = Δt * s4, and the mileage in the fifth prediction period is d5 = Δt * s5, and the mileage in the sixth prediction period is d6 = Δt * s6, and the distance traveled in the final forecast period is d7 = βΔt * The sum of these distances is equal to the physical distance of road segment R2, i.e., d1 + d2 + d3 + d4 + d5 + d6 + d7 = D2. Finally, the length of side ε, w vs,ve (t s )=D1 / s0+d1 / s1+d2 / s2+d3 / s3+d4 / s4+d5 / s5+d6 / s6+d7 / s7=D1 / s0+(α+5+β)Δt.

[0063] FIG. 6 shows functional blocks and interactions between components of the intelligent transportation system 100 shown in FIG. 1 according to some embodiments of the present invention. Here, each data collector 110 has a local database 600, and the learning server 115 designs a traffic model and corresponding rules for clustering data (605). Then, the learning server distributes the data clustering rules and traffic models to the data collectors and coordinates multiple rounds of distributed training (610). Upon receiving the rules and traffic models, each data collector clusters its own local data and trains a traffic model using its own local data (615), as shown in FIG. 2. After the distributed training is completed, the learning server builds a global traffic model by aggregating the locally trained traffic models (620). Then, the learning server distributes the global traffic model to the data collectors (625), and the data collectors make local traffic predictions (630). The local traffic predictions are sent to the learning server (635). The learning server constructs a time-dependent graph by using the road network 105 and traffic predictions from the data collectors, and distributes the time-dependent graph to the vehicle 120 for route planning (640). The vehicle uses the time-dependent graph to plan its route to minimize travel time (645). The vehicle travels the planned route to reach the destination in the minimum travel time (650). As the vehicle travels its planned route, unexpected events such as new road constructions and traffic accidents may occur. If the observed events cause a discrepancy with the planned travel time, the vehicle can feed these events back to the learning server to update the traffic predictions (655).

[0064] The above-described embodiments of the present disclosure can be implemented in any of numerous ways. For example, the embodiments may be implemented using hardware, software, or a combination thereof. When an ordinal term such as "first," "second," or the like is used in the claims to modify a claim element, it does not in itself imply a priority, precedence, or order of one claim element over another, nor does it imply a chronological order of operations in a method of performing, but is merely used as a label to distinguish one claim element having a particular name from another element having the same name (but using an ordinal term) to distinguish the claim elements.

[0065] Although the present disclosure has been described with reference to certain preferred embodiments, it must be understood that various other adaptations and modifications can be made within the spirit and scope of the disclosure. It is therefore the object of the appended claims to cover all such variations and modifications that come within the true spirit and scope of the disclosure.

[0066] Also, embodiments of the present disclosure may be embodied as a method or computer-implemented method, an example of which is provided. The order of operations performed as part of this method may be determined in any suitable manner. Thus, embodiments may be configured to perform operations in an order different from that illustrated, which may include performing some operations simultaneously, even though in the illustrated embodiment they are shown as a sequence of operations.

Claims

【Request 1】 【Number 1】 【Request 2】 【Number 2】

3. The method according to claim 1 or 2, wherein a learning server designs the global multi-task traffic model in the form of a traffic prediction neural network, and the learning server can be a roadside unit or an eNodeB or a cloud.

4. 3. The method of claim 1 or 2, wherein distributed learning agents train the global multi-task traffic model by using their local data, the learning agents may be infrastructure data collectors, each data collector being located on a road network and configured to collect traffic data at its location.

5. The method of claim 1 or 2, wherein the learning server is configured to design data cluster rules and distribute the data cluster rules to the learning agents.

6. The method of claim 5 , wherein each learning agent is configured to cluster its local data using the distributed rules.

7. 3. The method of claim 1 or 2, wherein the learning server and learning agents cooperatively construct the time-dependent global traffic map, the time-dependent global traffic map being modeled as a time-dependent graph, each edge of the time-dependent graph being a travel time indicating a predicted direction on a road segment connecting two vertices of the graph.

8. 8. The method of claim 7, wherein the learning server distributes the well-trained global traffic model to learning agents for travel time prediction, and each learning agent calculates the travel time on a road segment by predicting traffic speed using the global traffic model and its own local data, and the traffic speed prediction is multi-horizon by predicting short-term traffic speed, medium-time traffic speed, and long-term traffic speed.

9. 3. The method of claim 1 or 2, wherein the road network is generated by transforming a map stored on a map server, the map being provided to cover a large geometric area such as a city.

10. 10. The method of claim 9, wherein the vertices of the time-dependence graph include vertices corresponding to intersections or junctions of two adjacent road segments, and the edges of the time-dependence graph are the road segments connecting adjacent the vertices.

11. The method of claim 9 , wherein the road network is transformed into a time-dependent graph based on the well-trained traffic predictive neural network.

12. The method of claim 1 or 2, wherein the driving plan includes at least a portion of a current location, a destination, a time, a date, waypoints on the road network, or a combination thereof.

13. The method of claim 1 or 2, wherein each learning agent trains a traffic prediction neural network corresponding to a different traffic task.

14. The method of claim 10 , wherein an edge is configured to include at least one learning agent.

15. The method of claim 7 , wherein travel times on edges of the time-dependent graph are calculated using segment traffic speed predictions based on a prediction time interval Δt.

16. The method of claim 1 or 2, wherein the distributed machine learning is achieved by a Federated Learning (FL) technique.

Images