Multi-agent Traffic Simulation Research Articles

Distributional equity impacts of automated vehicles: A disaggregated approach

Better understanding of the equity impacts of automated vehicles (AV) is needed to design equitable deployment plans for AV technology. The goal of this study was to develop and demonstrate a modeling framework to support distributional equity assessments of AV systems prior to their wide adoption. This modeling framework takes a disaggregated approach that integrates agent-based simulation, diverse transportation outcomes, and equity analyses. The framework uses individual-level data to capture detailed disparities in transportation outcomes, applies a state-of-the-art multi-agent traffic simulation model (MATSim) to simulate privately-owned and shared AVs simultaneously, and considers distributional equity from multiple perspectives. To test and demonstrate the framework, a case study was conducted for the Tampa Bay region. Five scenarios were considered with different AV market shares and integration strategies based on scenario planning by the U.S. Federal Highway Administration. Results reveal that high AV penetration rates were required for substantial reductions in inequality. The introduction of AVs led to a more even distribution of accessibility, but slightly more uneven distribution of traditional mobility and affordability. Impacts of disparities in outcomes for disadvantaged groups were mixed. These results suggest that AVs will likely perpetuate existing inequity in the transportation system as long as its fundamental structure remains the same as today. Results highlight the importance of planning and design strategies that directly address the distributional impacts to ensure that AV technology is deployed equitably.

Learning from Demonstrations of Critical Driving Behaviours Using Driver's Risk Field

In recent years, imitation learning (IL) has been widely used in industry as the core of autonomous vehicle (AV) planning modules. However, previous IL works show sample inefficiency and low generalisation in safety-critical scenarios, on which they are rarely tested. As a result, IL planners can reach a performance plateau where adding more training data ceases to improve the learnt policy. First, our work presents an IL model using the spline coefficient parameterisation and offline expert queries to enhance safety and training efficiency. Then, we expose the weakness of the learnt IL policy by synthetically generating critical scenarios through optimisation of parameters of the driver's risk field (DRF), a parametric human driving behaviour model implemented in a multi-agent traffic simulator based on the Lyft Prediction Dataset. To continuously improve the learnt policy, we retrain the IL model with augmented data. Thanks to the expressivity and interpretability of the DRF, the desired driving behaviours can be encoded and aggregated to the original training data. Our work constitutes a full development cycle that can efficiently and continuously improve the learnt IL policies in closed-loop. Finally, we show that our IL planner developed with less training resource still has superior performance compared to the previous state-of-the-art.

A Nationwide Impact Assessment of Automated Driving Systems on Traffic Safety Using Multiagent Traffic Simulations

The objective of this paper is to propose a methodology to estimate nationwide traffic safety impacts of automated vehicle technologies using multi-agent traffic simulations. The influence of three levels of driver trust in the automation system (appropriate, over trust, distrust) is considered in the simulation and takes different transition modes of control between the driver and the system into account. The nationwide estimation of crashes is obtained by projecting results of the simulations using traffic data for three different and representative municipalities. Results indicated that Automated Driving Systems and Advanced Driver Assistance Systems significantly reduced the number of casualties and fatalities compared to manual driving. Simulation results in consideration of the influence of driver trust also found that this reduction may be negatively affected by over- and under-trust parameters. However, even with the introduction of these parameters, the reduction rate was still significant compared to manual driving. The proposed methodology using multi-agent traffic simulations may thus address concerns surrounding the deployment of automated driving systems which is a feature not found in conventional simulations, provide useful insight for interested parties to develop research and policy making strategies that accelerate traffic safety improvements, and to support social acceptance efforts.

Compatibility of Automated Vehicles in Street Spaces: Considerations for a Sustainable Implementation

Automated Vehicles (AVs) will bring a fundamental change in the mobility sector in the coming years. Whereas many studies emphasize opportunities with AVs, studies on the impacts of AVs on travel behavior particularly show an overall increase in traffic volume. This increase could impair the needs of other uses and users within street spaces and decrease the permeability of the street space for pedestrians and cyclists. However, only a few studies, so far, have looked at the changes of traffic volume due to AVs at the street level, and to what extent these impair the needs of other uses and users within different street spaces was not in the focus at all. This paper investigates the compatibility of AVs in street spaces, building on different modeling results of scenarios with AVs based on the Multi-Agent Traffic Simulation (MATSim) framework. Using the so-called compensatory approach and the whole street network of Vienna, Austria, as a case study, we examine how compatible AVs and their related changes in traffic volume are with the needs of other uses and users, i.e., pedestrians and cyclists, within different street spaces, by specifically considering the various characteristics of the latter. Results show that the effects of AVs on the compatibility of street spaces would be unevenly distributed across the city. For Shared Automated Vehicles (SAVs), a deterioration in compatibility is observable, especially in inner-city dense areas, because of an increase in traffic volume and an already high amount of competing uses. In contrast, especially (on main roads) in the outskirts, improvements in compatibility are possible. This particularly applies to SAVs with a stop-based service. However, private AVs interlinked with an overall capacity increase would lead to a deterioration in compatibility, especially in parts of the higher-level street network that already have incompatible traffic volumes, further increasing the separating or barrier effect of such streets. The results can provide insights for policymakers and stakeholders about where and how to facilitate AVs, to reach an implementation that is compatible with the different uses and needs of users within street spaces: While SAVs should be implemented particularly in the outskirts, as a complement for public transport, an implementation of AVs in the lower-level street network in inner parts of the city should not be facilitated, or it should at least be linked to measures that make street spaces more compatible with the needs of pedestrians and cyclists, e.g., implementation of walking and cycling infrastructure.

Methodology for Determining Charging Strategies for Urban Private Vehicles based on Traffic Simulation Results

A major part of urban private vehicles needs to be electrified in order to meet the German climate protection goals in the transport sector. Therefore, charging strategies must be developed and the geographical distribution of infrastructure has to be planned accordingly. We propose a new method to combine a multi agent traffic simulation (MATSim) with real-world vehicle distribution and measured charging behavior. The results of our simulation were analyzed for the exemplary use case of Berlin’s private passenger cars. Three different charging strategies were developed and analyzed: charging at home, at work and during leisure activities. Charging at home was suitable to serve all agents’ charging demand with low charging power (7.4 kW). Charging at work only served 56.23 % of the agents with a charging power of 11 kW, but required fewer charging points. Charging during leisure activities served from 46.36 % to 81.92 % of the agents, depending on the chosen charging power and user behavior.

Assessing the effect of different parking pricing policies on DRT demand using multiagent traffic simulation, case study of St. Petersburg

Abstract This paper presents the method to perform the sensitivity analysis of demand responsive transit demand to the parking pricing policies using microsimulation. Both the disutility of parking search cruising and parking fees are considered in the mode choice model. The preliminary results suggest that parking pricing policies may affect the DRT demand, reduce the overall congestion and vehicles kilometers travelled by car. In the St. Petersburg scenario up to 50% of DRT users, that have switched from the car, have previously parked the car in the city center, mostly for the prolonged time period (4 to 8 hours).

Multi-agent traffic simulations to estimate the impact of automated technologies on safety

Objective: The objective of this article was to develop a multi-agent traffic simulation methodology to estimate the potential road safety improvements of automated vehicle technologies.Methods: We developed a computer program that merges road infrastructure data with a large number of vehicles, drivers, and pedestrians. Human errors are induced by modeling inattention, aimless driving, insufficient safety confirmation, misjudgment, and inadequate operation. The program was applied to simulate traffic in a prescribed area in Tsukuba city. First, a 100% manual driving scenario was set to simulate traffic for a total preset vehicle travel distance. The crashes from this simulation were compared with real-world crash data from the prescribed area from 2012 to 2017. Thereafter, 4 additional scenarios of increasing levels of automation penetration (including combinations of automated emergency braking [AEB], lane departure warning [LDW], and SAE Level 4 functions) were implemented to estimate their impact on safety.Results: Under manual driving, the system simulated a total of 859 crashes including single-car lane departure, car-to-car, and car-to-pedestrian crashes. These crashes tended to occur in locations similar to real-world crashes. The number of crashes predicted decreased to 156 cases with increasing level of automation. All of the technologies considered contributed to the decrease in crashes. Crash reductions attributable to AEB and LDW in the simulations were comparable to those reported in recent field studies. For the highest levels of automation, no assessment data were available and hence the results should be carefully treated. Further, in modeling automated functions, potentially negative aspects such as sensing failure or human overreliance were not incorporated.Conclusions: We developed a multi-agent traffic simulation methodology to estimate the effect of different automated vehicle technologies on safety. The crash locations resulting from simulations of manual driving within a limited area in Japan were preliminary assessed by comparison with real-world crash data collected in the same area. Increasing penetration levels of AEB and LDW led to a large reduction in both the frequency and severity of rear-end crashes, followed by car-to-car head-on crashes and single-vehicle lane departure crashes. Preliminary estimations of the potential safety improvements that may be achieved with highly automated driving technologies were also obtained.

Driver profiling – Data-based identification of driver behavior dimensions and affecting driver characteristics for multi-agent traffic simulation

This paper focusses on the role of driver individuality in the field of cognitive driver behavior modeling for the prospective safety impact assessment of advanced driver assistance systems (ADAS) and automated driving functions. Virtual traffic simulation requires valid models for the environment, the vehicle and the driver. Especially modeling human driver behavior is a major challenge, which in recent years has already led to the development of various driver models for the purpose of virtual simulation. Modeling human behavior in traffic with a precise representation of human cognition, capability and individuality, are crucial demands, which require thorough investigation and understanding of the human driver. Current driver behavior models often leave aside the aspect of driver individuality and lack the consideration of differences in driving behavior between different drivers. To take into account all the aspects from complex human cognitive processes to individual differences in action implementation, the Stochastic Cognitive Model (SCM) was developed. The SCM is based on five subcomponents: gaze control, information acquisition, mental model, action manager and situation manager (=decision making process) and action implementation. The aim of the present study is to provide a basis for establishing a solid logic for the integration of driver individuality into the current structure of the SCM by creating a new submodule that takes into account several behavior affecting driver characteristics. This subcomponent controls the stochastic variance in several driver behavior parameters, such as velocity or comfort longitudinal acceleration. In a representative driving simulator study with 43 participants, driver behavior on the highway was investigated and thoroughly analyzed. Information about several relevant driver characteristics and personality traits of the participants was collected and a logical hierarchical model was set up to cluster several dependent and independent variables into four layers: independent manifest driver variables, such as age or gender (Level 1), latent driver personality factors, such as thrill seeking or anxiety (Level 2), driver behavior dimensions, such as dynamics and law conformity (Level 3), and various dependent driver behavior parameters, such as velocity, acceleration or speed limit violation (Level 4). Multiple linear regression analyses were run to find the individual driver characteristics and personality traits, by which most of the stochastic variance in the measured driver behavior parameters can be explained. Subsequently, a principal component analysis (PCA) was run to test, if the previously clustered driver behavior parameters were loading on the presumed behavioral dimensions on the third level of the model to identify significant components of driver behavior, such as dynamics or law conformity. Results of the present study show significant correlations between driver characteristics and driver behavior parameters. According to the results of the PCA, variability in driver behavior can be explained to a great extent by three largely independent components: (1) Speed and cruise control, (2) Dynamics and (3) Driver performance. With the consideration of driver individuality in driver behavior models for the agent-based traffic simulation, validity of the results from prospective safety impact assessment analyses of automated driving functions can be enhanced. Beyond that, the findings of the current study can be used as a solid basis for the development of adaptive functions in the field of vehicle automation, considering the different driving skills and preferences of drivers with different individual profiles.

Investigation of the Potential of Multi-Agent Traffic Simulations to Find Good Arrangement of Taxies

Investigation of the Potential of Multi-Agent Traffic Simulations to Find Good Arrangement of Taxies

Reinforcement Learning of Driver Receiving Traffic Signal Information for Passing through Signalized Intersection at Arterial Road

Abstract Real-time traffic signal information has recently become available thanks to further developments in Intelligent Transportation System (ITS) technology. Drivers may change their driving behavior by receiving information on the status of upcoming traffic signals. The behavior of a driver who reacts to such information with the purpose of smoothly passing through an upcoming signalized intersection is described as reinforcement learning in this study. The influence of reactive drivers on traffic flows is analyzed by a multi-agent traffic flow simulation on an arterial road composed of four signalized intersections. Furthermore, the relationship between the ratio of reactive drivers and traffic flows is clarified, under the assumption that there is a mixture of reactive and nonreactive drivers on the road. The results of the multi-agent traffic flow simulation showed that the average stop time decreased as the ratio of reactive drivers increased. Moreover, the threshold of the ratio of reactive drivers to begin to influence largely vehicle stop situations was around 50%. The average travel time was almost constant regardless of the ratio of reactive drivers. It was also demonstrated that reacting to the traffic signal information and decelerating accordingly did not cause time delays. This suggests that the provision of the traffic signal information and appropriate reaction to the provided information may help reduce the amount of CO2 emissions from a vehicle approaching a signalized intersection and probably contribute to alleviating progress of global warming.

A novel multi-agent traffic simulation model to predict the impact of automated vehicles on traffic safety

Real traffic and proving ground tests are crucial to develop and evaluate automated vehicles. However, comprehensive evaluations of the potential impact that these vehicles may have on traffic safe...

Two distribution methods for multiagent traffic simulations

Modeling and simulation play an important role in transportation networks analysis. With the widespread use of personalized real-time information sources, the behavior of the simulation depends heavily on individual travelers reactions to the received information. As a consequence, it is relevant for the simulation model to be individual-centered, and multiagent simulation is the most promising paradigm in this context. However, representing the movements of realistic numbers of travelers within reasonable execution times requires significant computational resources. It also requires relevant methods, architectures and algorithms that respect the characteristics of transportation networks. In this paper, we define two multiagent simulation models representing the existing sequential multiagent traffic simulations. The first model is fundamental diagram-based model, in which travelers do not interact directly and use a fundamental diagram of traffic flow to continuously compute their speeds. The second model is car-following based, in which travelers interact with their neighbors to adapt their speeds to their surrounding environment. Then we define patterns to distribute these simulations in a high-performance environment. The first is agent-based and distributes agents equally between available computation units. The second pattern is environment-based and partitions the environment over the different units. The results show that agent-based distribution is more efficient with fundamental diagram-based model simulations while environment-based distribution is more efficient with car following-based simulations.

A framework for a multi-agent traffic simulation using combined behavioural models

The task of simulation of urban processes often concerns city traffic and road dynamics. Many assumptions in an urban science may be made by analysing traffic flows and vehicle behaviour on city streets. Since road traffic is a dynamic and possibly spatially large system that depends on many external conditions, approaches like visual analysis and computational steering may be used. There are many traffic modelling frameworks presented in the field, but there are also several drawbacks most of them have: they are limited to a specific model or case; have abstracted logic which disregards some factors affecting vehicle movement; they ignore possible physical interactions which are not described in a model or use too many simplifications in vehicle interaction simulations. In this paper, a novel agent-based traffic modelling framework is presented, combining different classes of traffic models into a single vehicle agent and allowing the user to set a specific model for each supported class. The framework was made using a game engine, Unreal Engine 4, which allows users to model realistic interaction between vehicles and make realistic visualisations for purposes of visual analysis and computational steering.

Bi-objective evolutionary optimization of level of service in urban transportation based on traffic density

This work investigates levels of service in urban transportation coupling a multi-objective evolutionary algorithm with the multi-agent traffic simulator MATSim. The evolutionary algorithm searches...

Uncertainty Quantification of Traffic Volume in Multi-Agent Traffic Simulation

Uncertainty Quantification of Traffic Volume in Multi-Agent Traffic Simulation

Charging Equipment Layout Method for Electric Vehicles based on Multi-Agent Traffic Flow Simulation

Charging Equipment Layout Method for Electric Vehicles based on Multi-Agent Traffic Flow Simulation

Agent-based simulation of autonomous taxi services with dynamic demand responses

Abstract: An agent-based simulation approach is presented, which makes it possible to capture the dynamic interplay between a supply of autonomous vehicle fleets with distinct operational schemes and a population of artificial persons based on an established multiagent traffic simulation framework. The simulation is able to show how agents react to the new travel options and make consistent decisions based on a well-defined framework of utility scoring.

A Versatile Multi-Agent Traffic Simulator Framework Based on Real Data

Among real-system applications of AI, the field of traffic simulation makes use of a wide range of techniques and algorithms. Especially, microscopic models of road traffic have been expanding for several years. Indeed, Multi-Agent Systems provide the capability of modeling the very diversity of individual behaviors. Several professional tools provide comprehensive sets of ready-made, accurate behaviors for several kinds of vehicles. The price in such tools is the difficulty to modify the nature of programmed behaviors, and the specialization in a single purpose, e.g. either studying resulting ows, or providing an immersive virtual reality environment. Thus, we advocate for a more exible approach for the design of multi-purpose tools for decision support. Especially, the use of geographical open databases offers the opportunity to design agent-based traffic simulators which can be continuously informed of changes in traffic conditions. Our proposal also makes decision support systems able to integrate environmental and behavioral modifications in a linear fashion, and to compare various scenarios built from different hypotheses in terms of actors, behaviors, environment and ows. We also describe here the prototype tool that has been implemented according to our design principles.

A Hybrid Learning Algorithm for Generating Multi-Agent Daily Activity Plans

This paper proposes a hybrid learning algorithm based on the competing risk model and the cross entropy method for generating complete one-day activity plans for multi-agent traffic simulations. An agent’s activity plan generation process is modeled using the Markov decision process. As generating complete activity plans of agents using a reinforcement learning approach is computationally expensive and inefficient, we propose a hybrid method that first estimates the activity type of agents and the scheduled ending time sequences from empirical data based on a competing risk model. The activity plans obtained are then completed by the cross entropy method for the optimal destination choice learning of agents. The performance of the proposed method is compared with the Q-learning algorithm. The numerical result shows that the proposed method generates consistent daily activity plans for multi-agent traffic simulations.

A GPU-Based Programming Framework for Highly-Scalable Multi-Agent Traffic Simulations

Highly detailed reproducibility of multi-agent simulations is strongly demanded. To realize such highly reproducible multi-agent simulations, it is important to make each agent respond to its dynamically changing environment as well as scale the simulation to cover important phenomena that could be produced. In this paper, we present a programming framework to realize highly scalable execution of them as well as detailed behaviors of agents. The framework can help simulation developers utilize many GPGPU-based parallel cores in their simulation programs by using the proposed OpenCL-based multi-platform agent code conversion engine. We show our prototype implementation of the framework and how our framework can help simulation developers to code, test, and evaluate their agent codes which select actions and path plants reactively in dynamically changing large-scale simulation environments on various hardware and software settings.