Annotation of the results obtained in 2023
The scientific objectives of the third stage of the project were: 1) develop recommendations for increasing the efficiency of using the transport system with the proposed solutions; 2) develop a research software package for the adaptive traffic signal control and CAV and CV vehicles using big data processing approaches, high-performance processing systems and modern software systems/packages for traffic modelling. The key scientific task of the third stage was to evaluate the effectiveness of the proposed solutions using various scenarios for the interaction of vehicles and infrastructure objects (traffic signals), first of all, to analyze the effectiveness of the developed methods and algorithms in a scenario with mixed traffic flows, including with connected (CV), autonomous (CAV) and regular (human-driven) vehicles (RV). For the applicability of the methods and algorithm developed in the first and second stages of the project in a mixed traffic flow consisting of connected and human-driven vehicles, a modification of the developed solutions was carried out. Modified methods and algorithms for joint coordinated control of phases/cycles of traffic signals and CAV and CV vehicles in a mixed flow were developed using simultaneous control of the trajectories of CAV and CV vehicles and phases/cycles of traffic signals to increase the capacity of the transport network by reducing waiting time at intersections and maximizing the use of the green cycle of traffic signals. The developed methods and algorithms as separate stages include: 1) Methods and algorithms for short-term prediction of travel time along the route of CV/CAV vehicles in a scenario with adaptive traffic signal control using general information about the load of the transport network and available information from CAV/CV vehicles, including in scenarios with mixed traffic flows. 2) A vehicle routing algorithm in a time-dependent traffic network that combines a mixed traffic flow prediction method and an adaptive traffic signal control method. 3) Modified adaptive traffic signal control methods using general information about the load of the transport network and available information from CAV/CV vehicles in scenarios with a mixed traffic flow, including CV/CAV and RV vehicles. The method is based on maximizing the weighted flow of vehicles passing through an intersection using available information from connected and observed vehicles and predictive information about the movement parameters of connected and observed vehicles obtained using an artificial neural network model. To develop recommendations for improving the efficiency of using transport infrastructure with the proposed solutions, experiments were conducted using various scenarios for the interaction of vehicles and infrastructure facilities (traffic signals). In particular, we estimate the impact on the efficiency of models describing the state and distribution of traffic flows, vehicle movement dynamics, adaptive traffic signal control, as well as methods of interaction between traffic signals and CAV vehicles for their joint coordinated control. Experimental results include: 1) Results of experimental studies of the method and algorithm for short-term prediction of travel time along the route of CV/CAV vehicles in a scenario with adaptive traffic signal control using general information about the load of the transport network and available information from CAV/CV vehicles. The proposed travel time prediction algorithm based on the ANN model taking into account the waiting time at the intersection showed better results compared to basic approaches based on the mean absolute error criterion. The reduction in mean absolute error ranged from 45% for a small urban area traffic simulation scenario to 63.6% for a large-scale transport network simulation scenario compared to a deterministic travel time prediction model. 2) Results of experimental studies of a vehicle routing algorithm in a time-dependent transport network with a mixed traffic flow. The proposed algorithm made it possible to reduce the average waiting time to 10%, the average waiting time to 6.6%, and the average fuel consumption to 4.9% compared to the basic algorithm for solving the routing problem. 3) Results of experimental studies of a modified adaptive traffic signal control method using general information about the load of the transport network and available information from CAV/CV vehicles in scenarios with a mixed traffic flow, which confirmed the effectiveness of the proposed approaches. Average travel time decreases up to 1.9% and average waiting time to 14.6%. The use of additional information about the movement of observed vehicles also reduces the values of the criteria under consideration. Increasing the duration of the phase, on the contrary, increases the duration of transport correspondence. 4) Results of experimental studies of methods and algorithms for joint coordinated control of phases/cycles of traffic signals and CAV and CV vehicles in a mixed traffic flow using simultaneous control of the trajectories of CAV and CV vehicles and phases/cycles of traffic signals. Experimental studies in a simulation environment show that the proposed methods can reduce average fuel consumption (up to 4.2%), average driving time (up to 5.3%), and average waiting time (up to 27%) compared to state-of-the-art methods for solving the problem of adaptive traffic signal control. The conducted studies confirm that control efficiency increases both with an increase in the ratio of connected vehicles in the modeling scenarios under consideration, and with the consistent use of methods of adaptive control of only traffic signals and a method of joint control of the trajectories of connected vehicles and adaptive traffic signal control. As a recommendation for increasing the efficiency of using transport infrastructure with the developed solutions with the increase in the ratio of connected vehicles in the flow, it is proposed to consistently use - method of adaptive traffic signal control based on maximizing the weighted traffic flow; - method of adaptive traffic signal control based on reinforcement learning; - method of jointly control of the vehicle trajectories and traffic signals. As part of the second task, considered at the third stage, the development of the architecture and implementation of a research software of a cooperative intelligent transport system was carried out, designed to solve the problems of controlling traffic signals and the trajectories of vehicles of the CAV and CV type, using approaches to big data processing , machine learning methods, high-performance processing systems and modern vehicle traffic simulation software systems/packages, which allow combining real-time streaming data processing with batch analytics results. The architecture of the research software and the components support horizontal scaling in order to process large-scale vehicle movements. The following components are used as components of the research package: - distributed software message broker Apache Kafka for data exchange between system components; - Apache Spark framework for implementing distributed processing of semi-structured big data; - NoSQL database Apache Cassandra for storing data processing results; - web application that implements an application programming interface (API) for interacting with users, receiving requests and returning results of analytical processing.

 

Publications

---

Annotation of the results obtained in 2021
In the first year of the current project, it was planned to develop methods and algorithms for traffic signal control, including coordinated control in a spatial area, on the basis of the traffic flow distribution and available information from CAV and CV vehicles. To solve this problem, a general information and computation technology was developed in the project, consisting of the following stages: Stage 0. Collect information on the state of traffic flows in the network (using data from standard data sources such as CCTV cameras or loop detectors and information from CV and CAV). Stage 1. Calculate a short-term forecast of the mixed traffic flows parameters using machine learning methods taking into account general information about the traffic flow distribution and available information from CAV and CV vehicles. Stage 2. Create a description (feature vector) of the current transport situation at a separate intersection or in a spatially compact area of the transport network. Predict the arrival time of specific vehicles at intersections using the specified information and the predicted dynamics of vehicle movement obtained at the first stage. Stage 3. Optimize the objective function selected to solve the traffic control problem using the machine learning method - reinforcement learning and/or using the developed control algorithm based on a deterministic predictive model. At the first stage, to solve the problem of short-term forecasting of a mixed traffic flow (consisting of CAV and CV vehicles and human-driven vehicles), we developed a short-term mixed traffic flow prediction algorithm that combines an algorithm for traffic flow prediction based on a graph neural network with an algorithm for anticipatory routing of connected vehicles. On Stage 2, we perform the following steps: - create a feature vector describing the transport situation at a separate intersection or in a spatially compact area of the transport network; - predict the arrival time of specific vehicles at intersections using the specified information and predicted dynamics of vehicle movement. To create a description (feature vector) of the current transport situation at a separate intersection or in a spatially compact area of the transport network, the following characteristics are used: the current traffic signal phase at the intersection and the duration of the current phase, the number of moving vehicles in each traffic lane regulated by the traffic signal, the queue length, the distribution of vehicle positions in traffic-light-controlled lanes and “pressure” at an intersection, which is defined as the number of vehicles on the lanes entering the intersection minus the number of vehicles on the corresponding outgoing lanes. To predict (estimate) the arrival time of vehicles at intersections, we developed methods and algorithms for predicting the arrival time of the CAV and CV vehicles using information from these vehicles. In particular, a deterministic predictive model was developed for estimating the time interval required to reach the intersection by the vehicle under consideration. The travel time required for the vehicle to reach the intersection generally consists of three components: time of uniformly accelerated rectilinear motion of the vehicle (until the maximum speed is reached), time of uniform rectilinear motion, and time delay at the start of motion. Finally, at Stage 3 of the proposed information and computation technology, the traffic signal control problem at an isolated intersection and coordinated control of a group of traffic signals in a given spatial area was solved using a feature vector containing current and predicted information about the traffic state and available information from CAV and CV vehicles. In this project, we developed and investigated two approaches to solving the traffic signal control problem: using the machine learning method - reinforcement learning and using the developed control algorithm based on a deterministic predictive model. Thus, during the first year of the project, the following scientific results were obtained: 1) A method and algorithm for predicting traffic flow parameters based on general information about traffic flows and available information from CAV and CV vehicles that combines an algorithm for traffic flow prediction based on a graph neural network with an algorithm for anticipatory routing of connected vehicles. 2) Methods and algorithms for predicting the arrival time of CAV and CV vehicles at intersections using information from the specified vehicles (in particular, the position and speed of vehicles) and models of the vehicle movement dynamics. 3) Methods and algorithms for traffic signal control at an isolated intersection and coordinated control of a group of traffic signals (in a given spatial area) using predictive information about the traffic states and available information from CAV and CV vehicles based on the machine learning method - reinforcement learning. 4) A method and algorithm for adaptive traffic signal control at an isolated intersection and coordinated control of a group of traffic signals (in a given spatial region) using predictive information about traffic states and available information from CAV and CV vehicles based on deterministic prediction models. 5) Results of experimental studies of the developed methods and algorithms: 5.1) An experimental study of the short-term mixed traffic flow prediction algorithm in the microscopic traffic simulation package SUMO confirms the effectiveness of the proposed solution in comparison with the non-CV traffic flow prediction. The experiments were conducted depending on the proportion of connected vehicles in the scenario (with available information about the departure time and the route). An increase in the penetration rate of CV vehicles from 0.2 to 1 made it possible to reduce the prediction error from 8% to 65% with a prediction horizon of one minute and from 6% to 39% with an increase in the prediction horizon to 5 minutes. 5.2) The experimental results of methods and algorithms for traffic signal control at an isolated intersection and coordinated control of a group of traffic signals in a given spatial area based on reinforcement learning, which showed the advantage of the developed algorithms in comparison with baseline control methods. Experimental studies have shown that the reinforcement learning approach is superior to classical traffic signal control methods. Compared to the algorithm that uses a predefined cycle and phase time plan, the gain in average travel time was 20.6%. In comparison with the algorithm that aimed to balance the queue length between adjacent intersections, the gain of the proposed algorithm was 14.4%. In addition, the proposed double-Q learning approach performed better than the basic reinforcement learning algorithms: 3.6% on a synthetic grid scenario with 36 intersections and 10.3% on a dataset containing 196 intersections with information about traffic flow from open data on taxi trips. 5.3) The experimental results of methods and algorithms for adaptive traffic signal control, based on a deterministic prediction model, which showed an advantage of the developed control method over the most effective existing solutions. Experimental studies were carried out on five scenarios: two synthetic and three real-life control scenarios that allow us to evaluate different traffic control aspects: signal control of an isolated intersection, coordinated signal control of multiple intersections along an arterial, and coordinated signal control of multiple intersections in a city area. We compared the algorithms using two metrics: average waiting time and average travel time. The proposed method showed the best result by the average waiting time criteria in four out of five scenarios. Moreover, the advantage over the second result in these four scenarios varies from 3% to 44%. According to the average travel time criteria, similar results were obtained, which allow us to conclude that the proposed algorithm performed better in most traffic scenarios. Based on the obtained experimental results, we can conclude that the method proposed in this work is superior to the known classical methods and reinforcement learning-based algorithms in most traffic scenarios. Moreover, reinforcement learning -based algorithms cannot provide robust performance for each scenario, which differs significantly in both the training dataset and test episodes. It should also be noted that the proposed algorithm does not require preliminary training, which is an important advantage over algorithms based on reinforcement learning.

 

Publications

---

Annotation of the results obtained in 2022
In the second year of the current project, it was planned to develop and study methods and algorithms for the cooperative coordinated control of traffic signals and connected vehicles (CV) and / or autonomous vehicles (CAV) using general information about the traffic flow distribution and available information from CAV and CV vehicles. As part of solving this problem, an approach to control the signals / phases of traffic lights, developed at the first stage of the project, was elaborate. According to the research work plan, the solution to the problem included the following sub-stages: I) Solution of the problem of controlling the trajectory of the CAV and CV vehicle types, which optimizes the key quality indicators of the movement of vehicles at the intersection, taking into account the structure of the intersection, the current predefined traffic signal cycles, and the current dynamics of vehicles. II) Solution of the problem of controlling traffic signal phases taking into account the structure of the intersection and information about the formed trajectories of the CAV and CV vehicle types and the current dynamics of vehicles. III) Solution of the problem of cooperative coordinated control of traffic signal phases and CAV and CV vehicle types, using simultaneous control of the trajectories of CAV and CV vehicles and traffic signal phases to increase the capacity of the transport network by reducing the waiting time at intersections and maximizing the use of the traffic signals green phase cycle. At stage I, the problem of controlling the trajectories of vehicles in a scenario with fixed (predefined) phases of traffic signals was considered. As part of the stage, an algorithm for constructing the trajectory of a vehicle at a controllable multi-lane intersection was developed. According to the algorithm, a vector containing the initial location of the vehicle, its speed, and the current time is formed for all vehicles entering into a considered intersection. Further, a physically realizable trajectory was built from the initial location to the intersection with a predefined set of parameters, including acceleration and speed limit. The obtained trajectory is the resulting trajectory if the movement is allowed at the moment of entry into the intersection for the considered lane. Otherwise, the second stage of the algorithm is performed, within which the trajectory is rebuilt so that the vehicle enters the intersection when the green phase of the traffic signal is on. To increase the efficiency of movement simulation, an approach based on multi-agent modeling with “grouping” of moving vehicles is used. Trajectories are calculated only for vehicles nearest to the intersection in each lane. The rest of the vehicles are controlled with the use of a leader-following model. At stage II, the problem of controlling phases of traffic signals was considered, taking into account the structure of the intersection and information about the generated movement trajectories of the CAV and CV vehicle types and the current dynamics of vehicles. Under stage II of the research plan, the modification of the solutions obtained in the first year of the project was carried out. In particular, — the elaboration of estimating techniques of the predicted weighted traffic flow passing through the intersection during the selected phase of the traffic signal cycle was developed using machine learning methods; — the modification of the solutions proposed at the first stage of the project was carried out: the method of maximum weighted flow, the approach based on the method of reinforcement machine learning (RL) to take into account predictive data in decision making. As part of this sub-stage, the maximum flow method was modified to take into account the waiting time at the intersection. To estimate the time required to achieve an intersection by a vehicle, it is proposed to use a model based on an artificial neural network. The characteristics of the traffic flow, which directly or indirectly describe the traffic situation on the current and adjacent road segments, as well as the movement of the considered vehicle are used as input parameters of the model. The next phase of the traffic signal cycle is chosen similarly to the basic method: the phase for which the weighted traffic flow is maximum is selected. In addition, the adaptive control method based on the reinforcement learning approach was also modified. In the project, it is proposed to apply predictive information about the number of vehicles passing the intersection for a given time interval to improve the efficiency of traffic signal control. Consequently, the hybrid approach proposed in the project consists of the following steps: 1) Each RL agent monitors the current state of the traffic flow and a) predicts the number of vehicles that pass the intersection during a given time interval, b) creates a feature vector describing the traffic situation. 2) Each RL agent chooses an action for the next time interval applying a double Q-learning approach using a generated feature vector. 3) The selected set of actions of all agents is sent to the system. At stage III, the problem of cooperative coordinated control of traffic signal phases and CAV and CV vehicle types was considered with the simultaneous control of the trajectories of CAV and CV vehicle types and phases of traffic signals to increase the capacity of the transport network by reducing the waiting time at intersections and maximizing using the green cycle of traffic signals. As part of solving this problem, the results obtained at sub-stages I and II of the project were used. In particular, the cooperative control method was developed using the algorithm for constructing the trajectory of CV and CAV vehicle types and the method for adaptive control of traffic signals based on maximizing the weighted traffic flow. The proposed algorithm for trajectory construction, considering the fixed phases of a traffic light, was modified to work with a traffic signal controlled by the adaptive method. The results of experimental studies of all the developed methods and algorithms carried out at each of the described stages confirmed the effectiveness of the proposed solutions and their advantage over the most effective existing approaches on synthetic and real scenarios of vehicle traffic simulation.

 

Publications

---