Appointments
- Project Scientist, National Robotics Engineering Center, CMU, Pittsburgh, April 2019 –
- Principal AI/ML Consultant, Machine Learning, Auris Health Inc.
- Project Scientist, The Robotic Institute, Carnegie Mellon University, Pittsburgh, August 2016 –
- Cofounder and Scientist, Rapid Flow Technologies, LLC, 2015
- Special Faculty, The Robotics Institute, Carnegie Mellon University, Pittsburgh, April 2014 – July 2016
SELECTED PUBLICATIONS
Multiagent sensor fusion for connected & autonomous vehicles to enhance navigation safety
Abstract
Today, autonomous vehicle (AV) navigation systems rely solely on local sensor data feed for safe & reliable navigation. However, it is not uncommon for sensor data to contain erroneous measurements resulting in false predictions, classified as either false positives (predict non-existent obstacle) or false negatives (e.g., missed obstacle). In this paper, we propose a methodology to identify and minimize false negatives in autonomous vehicle navigation, since these are arguably the most dangerous. According to the methodology, each autonomous agent simultaneously localizes and maps its local environment. This map, in turn, is encoded into a low-resolution message and shared with nearby agents via DSRC, a wireless vehicle communication protocol. Next, the agents distributively fuse this information together to construct a world interpretation. Each agent then statistically analyzes its own interpretation with respect to the world interpretation for the common regions of interest. The proposed statistical algorithm outputs a measure of similarity between local and world interpretations and identifies false negatives (if any) for the local agent. This measure, in turn, can be used to inform the agents to update their kinematic behavior in order to account for any errors in local interpretation. The efficacy of this methodology in resolving false negatives is shown in simulation.
Hierarchical Bayesian Framework for Bus Dwell Time Prediction
Abstract
In many applications, uncertainty regarding the duration of activities complicates the generation of accurate plans and schedules. Such is the case for the problem considered in this paper – predicting the arrival times of buses at signalized intersections. Direct vehicle-to-infrastructure communication of location, speed and heading information offers unprecedented opportunities for real-time optimization of traffic signal timing plans, but to be useful bus arrival time prediction must reliably account for bus dwell time at near-side bus stops. To address this problem, we propose a novel, Bayesian hierarchical approach for constructing bus dwell time duration distributions from historical data. Unlike traditional statistical learning techniques, the proposed approach relies on minimal data, is inherently adaptive to time varying task duration distribution, and provides a rich description of confidence for decision making, all of which are important in the bus dwell time prediction context. The effectiveness of this approach is demonstrated using historical data provided by a local transit authority on bus dwell times at urban bus stops. Our results show that the dwell time distributions generated by our approach yield significantly more accurate predictions than those generated by both standard regression techniques and a more data-intensive deep learning approach.
Accommodating High Value-of-Time Drivers in Market-Driven Traffic Signal Control
Abstract
In this paper, we propose a market-driven approach to traffic signal control. In contrast to traditional traffic engineering approaches, our approach gives agency and decision-making influence to individual drivers and exploits auction mechanisms to make traffic control decisions. Drivers make payments to their corresponding movement managers (each responsible for a particular directional flow through the intersection), and movement managers then compete for control of the signal. These financial transactions, if treated literally provide an alternate source of funding transportation infrastructure. Previous work with this model has demonstrated the ability to achieve better overall traffic flow performance than actuated control, a simple adaptive traffic signal control strategy based on detection and monitoring of waiting vehicles. Here we consider the design and analysis of bidding strategies capable of factoring in a given driver’s value of time (VOT), as indicated by the amount of voluntary contributions that are made on top of the fixed fee that every driver is charged. We analyze the potential for expediting high VOT drivers without undue disruption of overall traffic flows.
Synthesizing Route Travel Time Distributions Considering Spatial Dependencies
Abstract
Estimation of route-level travel time distributions (or travel rates) from segment-level data is of great interest today. This paper shows how the random variable properties of comonotonicity and independence can be used in combination to develop such distributions for a wide range of operating conditions. Efficacy of the technique is illustrated using Bluetooth-based travel times collected from individual vehicles on I-5 in Sacramento. The technique offers a way that agencies and other service providers can provide credible travel time (travel rate) distributions to route guidance devices, freight dispatchers and other users who want to know about the reliability of travel time estimates for the routes they utilize.
A Computationally Scalable Bayesian Sequential Learning Framework For Time-Series Forecasting
Abstract
Time-series forecasting is a widely used data science technique for predicting the future state of stochastic mechanisms. Application domains that benefit from such practices include stock markets, inventory planning, supply chain management, healthcare, and resource allocation under uncertainty. In recent years, deep learning has increasingly become the method of choice in time-series forecasting applications where historical data is abundant. Alternatively, frequentist approaches are quite popular in applications where historical data is limited, but their underlying assumption of stationary data tends to restrict their performance. This paper considers an alternative approach to small data applications, extending a Bayesian sequential learning technique to overcome this shortcoming. Historically, Bayesian learning approaches have suffered from scalability problems when applied to time-series forecasting due to the Bayesian estimation step’s complexity. Contemporary Bayesian approaches utilize Markov Chain Monte Carlo methods such as Metropolis-Hastings and Hamiltonian Monte Carlo for this purpose. While these methods have proved useful for offline time-series analysis, their computational requirements limit their utility in online, high temporal frequency forecasting. We propose substituting Nested Sampling, another method for estimating Bayesian evidence, as a means of achieving a scalable time-series forecasting framework. We demonstrate our approach’s effectiveness via comparative experimental analysis on three prediction problems of practical interest.
An Auction-based Scheduling Approach to the Dynamic Dial-a-Ride Problem
Abstract
The dial-a-ride problem (DARP) involves transporting a set of customers from respective origins to destinations within requested pick-up and drop-off time windows, using a fixed fleet of transport vehicles. In over-subscribed problem settings, the main objective is generally to maximize the number of requests that can be accommodated. In settings where there is sufficient vehicle capacity to service all requests, the objective shifts to some combination of minimizing overall travel time and minimizing number of vehicles used. One common application of DARP is found in door-to-door transportation services offered to elderly or disabled travelers.
Most prior research into solution approaches to DARP has focused on the static problem (where all requests are known in advance), and emphasized the use of offline optimization techniques and extended computation. These techniques are not viable however in dynamic DARP domains, where new requests continue to emerge through the day and unexpected events continually force changes to preplanned trips.In this paper, we address this gap and focus on solving the dynamic DARP formulation. We propose an online, auction-based scheduling strategy that constructs a solution incrementally over time and hence is capable of efficiently incorporating new requests as they arise (while also accounting for resource usage constraints and minimizing vehicle travel times). The performance of the proposed algorithm is evaluated on both a set of DARP feasibility benchmark problems from the literature and a large-scale real-world paratransit dataset more recently introduced by another incremental scheduling approach designed to maximize the number of requests that can be serviced in an oversubscribed context. In addition to obtaining favorable comparative results, the real-world paratransit data set is also used to analyze the impact of temporal uncertainty on solution quality, and to evaluate the potential for utilizing asymptotically different cost functions as a basis for introducing customer priority.
Additional Listings
- “Novel Perception Algorithmic Framework for Object Identification and Tracking In Autonomous Navigation” This paper introduces a novel perception framework that has the ability to identify and track objects in autonomous vehicle’s field of view. The proposed algorithms don’t require any training for achieving this goal. The framework makes use of ego-vehicle’s pose estimation and a KD-Tree-based segmentation algorithm to generate object clusters. In turn, using a VFH technique, the geometry of each identified object cluster is translated into a multi-modal PDF and a motion model is initiated with every new object cluster for the purpose of robust spatio-temporal tracking. The methodology further uses statistical properties of high-dimensional probability density functions and Bayesian motion model estimates to identify and track objects from frame to frame. The effectiveness of the methodology is tested on a KITTI dataset. The results show that the median tracking accuracy is around 91% with an end-to-end computational time of 153 milliseconds
- “Simulated Basic Safety Message: Concept & Application“: In this paper, we introduce a concept called Simulated – BSM that leverages the combination of rich vehicular sensor feed (speed, object proximity, SLAM maps etc.), and emerging guaranteed low-latency communication protocols (DSRC, 5G) to enhance efficiency & safety of transport facilities. The main idea is that an ego-vehicle with local sensor feed capability creates a basic safety message for any identified rogue vehicles in its field of view, and broadcasts it to infrastructure for implementing efficient and safe control actions. System architecture details are discussed. The efficacy of S-BSM framework was successfully demonstrated in the context of enhancing safe & efficient operations at rural high-speed signalized intersection.
- “Analysis of Trends in Transit Bus Dwell Time Data” Transit vehicles create special challenges for urban traffic signal control. Signal timing plans are typically designed for the flow of passenger vehicles, but transit vehicles—with frequent stops and uncertain dwell times—may have different flow patterns that fail to match those plans. Transit vehicles stopping on urban streets can also restrict or block other traffic on the road. This situation results in increased overall wait times and delays throughout the system for transit vehicles and other traffic. Transit signal priority (TSP) systems are often used to mitigate some of these issues, primarily by addressing delay to the transit vehicles. However, existing TSP strategies give unconditional priority to transit vehicles, exacerbating quality of service for other modes. In networks for which transit vehicles have significant effects on traffic congestion, particularly urban areas, the use of more-realistic models of transit behavior in adaptive traffic signal control could reduce delay for all modes. Estimating the arrival time of a transit vehicle at an intersection requires an accurate model of dwell times at transit stops. As a first step toward developing a model for predicting bus arrival times, this paper analyzes trends in automatic vehicle location data collected over 2 years and allows several inferences to be drawn about the statistical nature of dwell times, particularly for use in real-time control and TSP. On the basis of this trend analysis, the authors argue that an effective predictive dwell time distribution model must treat independent variables as random or stochastic regressors.
- “Comparing Actuated And Bid-based Control Strategies”. In this paper we explore some fundamental differences between control strategies modeled based on economic theory principles as opposed to more traditional control theory principles. More specifically, we explored differences between our bid-based control model and various actuated control models. The comparisons include situations where maximum greens are imposed in actuated control, and situations where they are not. To incorporate these ideas, two bid-based (b1, and b2) and four actuated (a1, a2, a3, and a4) control options are considered. Simulation experiments have been conducted to test these ideas for an intersection with two one-way, one-lane approaches, one eastbound and one northbound. Here are some of the inferences that can be drawn: 1) with actuated control, vehicles on the NB approach, with the larger volumes, always experience higher delays than the EB approach, while the differences between the NB and EB approaches are less significant for bid-based control; 2) Actuated control scenarios where maximum green are imposed (a2, a4) produced lower delay distributions than the scenarios where they are not imposed (a1, a3); 3) it is possible to find a combination of max-greens for a given flow condition that allow the actuated controller to match the performance (at least in terms of delays) of bid-based control; 4) bid-based control produced cycle length distributions that are smaller than those produced by various actuated control models; 5) Actuated control scenarios where maximum green are imposed (a2, a4) produced lower cycle length distributions than the scenarios where they are not imposed (a1, a3).
- “Analyses of Arterial Travel Times Based on Probe Data” This paper presents an analysis of arterial travel times based on AVI (automatic vehicle identification) data from vehicles that were equipped with toll tags. The source is a six-month experiment conducted on a small arterial network in upstate New York. Data were collected using wireless, solar-powered toll tag readers. The paper explores and examines trends by time of day, day of the week, and as affected by weather and other conditions. The results point toward the value of using such data for travel time prediction, travel time reliability monitoring, incident detection, and overall performance monitoring.
- “Bid-based signal control with all passive players”. In this paper, we present a realization of bid-based control strategy in which all drivers are modeled as passive players, and movement managers develop bidding strategies based on state-observer system principles. Their bidding strategies blend engineering attributes (length of dynamic queue, and number of turns since last win, which is analogous to delay) with economic attributes (the account balances of the movement managers). Movement managers bid for green time for their respective turning movements. Arriving motorists pay fees so the movement managers can bid for discharge slots. Movement managers pay the municipality what they bid when use of the intersections space is contested; otherwise, they pay a nominal fee. An intersection between two one-way streets has been used to test these ideas. To provide benchmarks against which to compare the results from bid-based control, a model of an actuated controller is employed. The results suggest that bid-based control strategy produces lower delay and cycle length distributions than those produced by actuated control strategy.
- “Using Travel Time Reliability Measures With Individual Vehicle Data”. The assessment of travel time reliability for segments and routes is a rapidly advancing frontier. The increasing availability of probe data is making it possible to monitor reliability in real-time based on individual vehicle data as opposed to ex-post-facto based on averages. This paper examines metrics that can be used to monitor reliability based on probe data. The merits of traditional metrics like the planning time index, buffer index, and travel time index are compared with newer ideas like complete cumulative distribution functions and mean/variance combinations. The question is: what is the quality of information about real-time reliability provided by these various options? This paper compares these metrics in the context of probe-based observations of travel times and rates. Also, a new idea for a pairwise metric, the root mean square travel rate τ rms in conjunction with the standard deviation σ τ . These two measures in combination seem to provide a picture of reliability that is nearly as complete as the underlying Cumulative Density Function (CDF) and better than the simpler metrics. These ideas are examined in the context of probe data from I-5 in Sacramento, CA.
- “Measuring Cruising For Parking in Washington, DC Using Dense, Ubiquitous, AVI Sensor Networks“ Drivers searching extensively for parking, a behavior known as cruising, leads to excess congestion and pollution. Many smart parking interventions have attempted to address this issue in recent years, including the parkDC initiative of the District Department of Transportation (DDOT) in Washington, DC. However, the problem of effectively detecting and measuring cruising remains largely unsolved. This paper describes a recently developed approach for continuous measurement of cruising for parking that is currently deployed in DDOT’s Penn Quarter/Chinatown Multimodal Value Pricing Pilot for metered curbside parking. This approach uses dense, ubiquitous networks of Bluetooth automatic vehicle identification (AVI) sensors to reconstruct and classify vehicle routes. Bluetooth AVI sensors are widely used for travel time measurement on freeways and arterials, as most new vehicles contain Bluetooth devices that can be passively detected and tracked across multiple stationary sensors. These techniques have been extended here to urban environments for vehicle route reconstruction and travel time measurement. Cruising routes are sufficiently distinct from normal travel routes, allowing reliable classification of vehicles as traveling or cruising. Results from the first several months of this deployment show that cruising rates for this network are very high, averaging 20-40% of trips depending on the time of day. Cruising vehicles contribute the majority of vehicle miles traveled in this network.
- “Cost-effective network topology for ubiquitous Bluetooth reader deployment in urban networks”. Travel time, an important performance measure for transportation systems, has traditionally been studied indirectly, but new technologies have made it possible to observe travel times directly. An increasingly popular method for travel time estimation is the use of a network of Bluetooth MAC address readers, where sampled addresses can be matched and travel times estimated. While most studies have addressed deployment for single roads, usually freeways or large arterials, the ubiquitous deployment of Bluetooth readers in a dense urban network raises new questions about network topology. In this paper, the authors explore the pros and cons associated with node (intersection) versus monument (mid-block) deployment of readers for dense, urban networks. A low-cost Bluetooth reader design developed for these experiments is described, and they present findings from four different deployments using these readers. The authors conclude that locating MAC readers at nodes is easier and more cost effective than deploying at monuments (where previous studies have recommended that readers should be located), without loss of data quality. The results in this paper show that MAC readers located at nodes are able to capture turning movements effectively, and in many cases outperform readers located at monuments.
- “Viewing Traffic Signal Control as a Market-Driven Economy”. In this paper, economic principles and the paradigm of a game are used to create a signal control strategy. The game structure is not formal (as in game theory), but the idea of a game is used nonetheless. That is, instead of using the standard techniques of minimum greens, maximum greens, and gaps to control the signal indications, an economically based game structure is employed. The intersection’s space is viewed as a scarce commodity whose use is determined through a bidding process. Movement Managers manage the vehicle departures for specific turning movements. Arriving motorists pay the Movement Managers an initial fee, and make voluntary contributions as they perceive necessary to arrange times of entry for them. Movement Managers submit bids for use of the intersection’s space and the highest bidders win. Distributed processing and connected vehicle technology are seen as the mechanisms by which implementation would be feasible. The value in such an idea is that one can study and reach an understanding of the economics that underlie effective traffic control.
- “Agent based framework for modeling operations at isolated signalized intersections”. An agent-based model of intersection control is presented. Drivers make payments to movement managers so that they can pass through the intersection. Movement managers collect fees from arriving drivers, receive voluntary contributions from those same drivers, and participate in a bidding process overseen by the municipality which determines which movement mangers get to discharge vehicles. Winning movement managers get to discharge vehicles from their queues for the duration of time associated with the win. A realization of the model is presented to illustrate these ideas. It involves two approaches, arriving drivers, two movement managers, and a municipality. The movement managers control the conflicting one-way traffic streams. They determine what bids to submit given the status of the system at each point in time, recalibrate those strategies based on past experience with their use, forecast the performance of those strategies, take actions based on those analyses, and then repeat the process in each time step as the simulation unfolds.
- “Simulating Adaptive Control Strategies in Large Urban Networks”. This paper describes a scalable approach to simulation of decentralized adaptive signal control systems, motivated by our interest to provide a basis for assessing the benefit of the Surtrac adaptive signal control system at a potential deployment site in advance of installation. The approach centers around a simulation controller interface called VISCO, which links the VISSIM microscopic traffic simulator to a set of externally hosted local intersection control processes. Local control processes are free to communicate with each other and exchange control information in the same manner that they would in a field implementation. VISCO coordinates all interaction with the simulator process to create a distributed software-in-the-loop simulation architecture. To illustrate and analyze the efficacy of the approach, we summarize a simulation analysis that was conducted of the downtown triangle area of Pittsburgh PA. A 63-intersection VISSIM model of this site is described and analyses are presented to characterize both the efficiency of the distributed architecture and the potential utility of Surtrac adaptive control. With respect to the former, the distributed simulation of local Surtrac control processes is found to run in roughly 4.4 times faster than real-time, in comparison to the 14.4 times faster than realtime speed that a conventional VISSIM simulation of this model with fixed timing plans performed. Experiments also show that the VISCO distributed architecture is effective in significantly reducing the cost associated with VISSIM’s external COM interface. With respect expected improvement of adaptive signal control in the downtown triangle area of Pittsburgh, the simulation analysis shows strong benefit of Surtrac over both the existing timing plans in use in this area and Synchro optimized plans that were generated with perfect knowledge of traffic volumes and turning counts.
- “Synthesizing Route Travel Time Distributions From Segment Travel Time Distributions“. This paper examines a way to synthesize route travel time probability density functions (PDFs) on the basis of segment-level PDFs. Real-world data from I-5 in Sacramento, California, are employed. The first finding is that careful filtering is required to extract useful travel times from the raw data because trip times, not travel times, are observed (i.e., the movement of vehicles between locations). The second finding is that significant correlations exist between individual vehicle travel times for adjacent segments. Two analyses are done in this regard: one predicts downstream travel times on the basis of upstream travel times, and the second checks for correlations in travel times between upstream and downstream segments. The results of these analyses suggest that strong positive correlations exist. The third finding is that comonotonicity, or perfect positive dependence, can be assumed when route travel time PDFs are generated from segment PDFs. Kolmogorov–Smirnov tests show that travel times synthesized from the segment-specific data are statistically different only under highly congested conditions, and even then, the percentage differences in the distributions of the synthesized and actual travel times are small. The fourth finding, somewhat tangential, is that there is little variation in individual driver travel times under given operating conditions. This is an important finding, because such an assumption serves as the basis for all traffic simulation models.
- “Effectiveness of Different Signal Control Strategies For Dilemma Zone Protection On High-Speed Approaches To Traffic Signals“. High-speed, rural intersections are challenging facilities to instrument and control. They often demand special attention to ensure safe operation. Different practices nationwide and worldwide have been created to minimize the number of collisions that occur because main street green terminates without sufficient yellow time to allow for safe stops. This paper compares the effectiveness of volume-density control with two strategies that involve advance detection: an NQ4 device that generates mainline green holds when long vehicles at high speeds are detected and a Detection-Control System (D-CS) developed by Bonneson et al. Simulation studies are conducted for five intersections, and field tests are conducted for three of these. The results indicate that volume-density control is surpassed in safety by both of the other strategies.
- “Hardware-In-the-Loop Simulation: Challenges and Solutions“. This paper discusses some complex Hardware-In-the-Loop configurations used for evaluating three distinct control strategies aimed at minimizing dilemma zone problems at high-speed, rural intersections. Five intersections were studied; three configurations were tested. The authors show that effective Hardware-In-the-Loop Simulations can be created for complex signal control configurations. They demonstrate that it is important to become familiar with the electrical details of the components and then create software and hardware splices that tie the devices together in such a way that each one functions in the manner intended. The authors show that for speed traps, the time step in simulation models dictates longer-than-field-based separations, about 29 m (95′) instead of 6.6 m (20′) (as in the field). This clearly suggests that time step duration has a significant impact on the results obtained for HILS, and hence suggests it worthwhile to push that frontier further, striving for even smaller time steps so that HILS can be effective.
- “Three Dilemma Zone Strategies For High-Speed Rural Intersections: Comparison of Field Results“. High-speed, rural intersections with both passenger car and truck flows present dilemma zone challenges. Three signal-timing strategies were tested in the field at three intersections of varying geometries with the goal of minimizing the occurrence of dilemma zones without sacrificing efficient operation. The strategies focused on hardware changes impacting vehicle sensing as well as mainline green phase extension and termination. The base strategy was volume-density control. The second one (NQ4) added advance detection of high-speed trucks triggering a fixed green extension. The final strategy replaced the second system with a sophisticated advance detection and control system (D-CS) logic based on the work of Bonneson et al. (2002). Results indicate simple volume-density control is surpassed in safety by both other strategies; moreover, the detection-control system dramatically reduced the number of vehicles trapped in dilemma zones at the onset of amber. The second strategy works fairly well in the field. The drawbacks to the NQ4 system are most noticeable at high-volume intersections — it does not actually find times when no vehicles are in dilemma zones, vehicle speeds are not directly used for computing main street hold times (which is accomplished in the D-CS control strategy resulting in improved efficiency), and it is a bit cumbersome and expensive to install. Based on our findings, the addition of the DC-S algorithm into a controller is a worthwhile investment for mixed-traffic, high-speed, rural intersections with dilemma zone issues.
- List, G. F., Demers, A., Isukapati, I.K., & Hauser, E., (2010), “Placement of Detection Loops On High Speed Approaches to Traffic Signals”. HWY 2007-13 Final Report Prepared For North Carolina Department of Transportation.
- Wallace, W., Wojtowicz, J., Murrugarra, R., List, G. F., Demers, A., & Isukapati, I. (2010), “Electronic Toll & Traffic Management Project. Task Report -28 & 29”, Prepared For New York State Department of Transportation
- Isukapati, I.K., & List, G.F., (2012), “Use Case Analysis (Supplement D), Establishing and Monitoring Travel Time Reliability (SHRP-2, L-02)”. Prepared for National Academy of Sciences
- Isukapati, I. K. (2014), “Intersection Control as a Shared Decision Process.” North Carolina State University
- “Analyses of Arterial Travel Times Based on Probe Data” In Advances in Dynamic Network Modeling in Complex Transportation Systems (pp. 115-142). Springer New York. This paper presents an analysis of arterial travel times based on AVI (automatic vehicle identification) data from vehicles that were equipped with toll tags. The source is a six-month experiment conducted on a small arterial network in upstate New York. Data were collected using wireless, solar-powered toll tag readers. The paper explores and examines trends by time of day, day of the week, and as affected by weather and other conditions. The results point toward the value of using such data for travel time prediction, travel time reliability monitoring, incident detection, and overall performance monitoring.