The ATMS Laboratories include a prototype system for modeling and evaluating ATMIS. The system was developed as part of the Testbed initiative with the intention that the prototype itself can be an operating ATMIS system that can be used to optimize, control, and manage real-world traffic, as well as allow for the investigation of ATMIS technologies without relying on field implementation of the detection and sensor hardware. In this latter mode, a core real-time simulation model drives the system, substituting simulated data for any real-world data from hardware contemplated to be part of a future ATMIS configuration. Current components and capabilities of the ATMIS Laboratories include:

• a hybrid simulation system;
  
• adaptive intersection signal control and fast traffic prediction techniques;
  
• adaptive freeway ramp control techniques;
  
• real-time O-D trip demand estimation modules;
  
• freeway incident detection algorithms based on neural networks;
  
• incident management expert systems for freeways and arterials;
  
• real-time optimal/equilibrium traffic assignment techniques;
  
• a comprehensive system of fault-tolerant traffic control based on distributed processing;
  
• dynamic assignment models;
  
• distributed processing framework for network optimization;
  
• image processing algorithms for vehicle-tracking and incident detection.

These modules are incorporated as a single framework where they communicate in real-time both with each other and with the Testbed, with appropriate transfer of information over ethernet. Flexible protocols have been developed for adding and deleting modules, and for interfacing with field devices. Details of each of these modules of the prototype system are provided below.

TRICEPS Hybrid Simulation System
TRICEPS, the Testbed Research Implementation Control and Evaluation Prototype System, has evolved from a first generation Testbed Workbench (TW), a software platform which facilitates the testing and evaluation of a wide range of algorithms for traffic control, advanced traffic management strategies (ATMS), and advanced traveler information systems (ATIS) with simulated or real world data. The current structure of TRICEPS can be divided into two major sections: an analysis section and a real-time control section. The analysis section is structured to allow the coordinated execution of multiple analysis modules on a network of UNIX workstations. The modules are driven by sensor data (from either a simulation or the real world) and by data generated from other modules' analyses. The backbone of this portion of TRICEPS is a midlevel distributed processing communications library called The UC Irvine Distributed Algorithm Testing Environment (ELUCIDATE) which takes a graph of interconnected processes, spawns the processes, establishes socket-based (low-level) communications between them, and provides a convenient interface for the communication between processes. On top of the ELUCIDATE library sits the Transportation Algorithm Interface Library (TAIL) which provides a more sophisticated interface for the connection of modularized transportation-related algorithms including tools for data interpretation and network representation translation, and a detailed message passing. Virtually any transportation algorithm can be connected with relative ease to any number of other algorithms via the TAIL/ELUCIDATE library to receive and/or provide data.

Adaptive Intersection Signal Control Module
The algorithm used here is based on minimizing the queue delays at individual intersections, with real-time information on turning fractions provided by the real-time traffic assignment module, and the traffic arrival patterns provided by the fast traffic prediction module. The hybrid simulation module waits for the signal indications from the signal control algorithm before proceeding with the simulation of each time-step, and thus this algorithm operates in a synchronous mode. The algorithm is based on a rolling horizon concept, with the solution found over a short immediate future horizon and repeated with updated arrival flows and turning fractions after each simulation time-step. The algorithm has been tested with simulated arrival patterns and has resulted in as much as 15 percent less delay than for actuated control.

Adaptive Freeway Ramp Control Techniques
This module uses freeway occupancy and flow data coming from the simulation system (or from an actual freeway, such as the SR-91 which is instrumented to send such data to UCI) to control the ramps in real-time. This module focuses on the algorithm side, and currently incorporates the ALINEA control algorithm. The module does not focus on the physical implementation of the control to separate computers for fault-tolerance purposes, which is done in a separate module described below. The control algorithm has been field-tested (Papageorgiou et al., 1991) and has been found to perform well.

Real-time O-D Estimation Techniques
This algorithm uses detector data from the simulation and the predicted equilibrium link flow values from the network assignment module to update Origin-Destination zonal demands based on a simple smoothing model and a seed O-D matrix. The algorithm operates on an asynchronous mode and is constantly repeated in real-time. The algorithm currently has the drawback that it does not have a dynamic path-flow assignment model incorporated. This would be expected since such dynamic assignment models in the literature (including the one developed in the testbed) are not fast-enough for real-time applications.

Incident Detection Algorithms
The algorithms used for freeway incident detection are based on neural network and decision tree approaches. Two different neural network approaches are currently available, one based on a feed-forward/back-propagation algorithm, and another based on newer paradigms from neuro-physiological research. The neural network models have undergone extensive testing with both simulated and real data from the SR-91 freeway and have been found to perform better than other existing algorithms in terms of the detection and false alarm rates. An incident detection algorithm is also available for arterials and is based on a modular neural network approach. A special-purpose algorithm for low-volume freeway incidents has also been developed based on a decision tree approach.

Incident Management Expert Systems
The incident management system (TCM) evolved out of two different modules, one for freeway incident management (FRED) and one for arterial incident management (ARTIST). The TCM is developed on the G2 real-time expert system shell. The existing model has a knowledge-base developed for the area around Disneyland in the City of Anaheim. The incident management actions include downloading new signal plans and showing CMS signs based on simple routing rules. The system receives information from the incident detection module whenever an incident is detected, and it also constantly receives information from the signal control module so that the knowledge-base is aware of the signal settings in the immediate past (which is used in deciding the signal settings to download under an incident management scenario). The link-level congestion information is received in real-time from the hybrid simulation. The expert system operates in an asynchronous mode.

Real-time Assignment Module
The assignment module is capable of executing both equilibrium and optimal assignment of origin-destination traffic demand to network paths. The assignment uses current levels of intersection delays and the latest estimated O-D table as input. The outputs are the link flow values and the intersection turning fractions. A path-flow based algorithm (Gradient Projection) is used for the assignment. The research on the testbed has demonstrated that this algorithm is one to two orders of magnitude faster than the conventional Frank-Wolfe algorithm. Since the algorithm is based on path-flow variables, it is easy to find the turning fractions at all intersections without adding any artificial turning links as required by link-flow algorithms such as Frank-Wolfe.

Fault-Tolerant Freeway Control System
The Fault-Tolerant Freeway Traffic Control System (FFTCoS) is an important component of TRICEPS. Its primary function is to perform metering for freeway on-ramps. The basic structure consists of distributed processing units. Unlike conventional traffic monitoring systems in which aggregate detector measures (e.g., volumes and occupancies at 30-second intervals) are provided to the other components of the monitoring system, raw loop detector data is broadcast to the distributed processing nodes of the FFTCoS.

The FFTCoS consists of the following components:

regional processors, each of which interfaces with standard sensors and actuators such as loop-detectors, collector stations, and ramp controllers in one real or simulated freeway segment.
processing components for global monitoring, control, and data logging.
Under this general structure, a wide variety of research and algorithmic testing is possible. In the current implementation the modules in the analysis section include a hybrid traffic simulation via the mesoscopic simulation model DYNASMART and more detailed simulation of portions of the freeway network using the microscopic simulation model INTRAS. The hybridized simulation co-simulates portions of the network with both models, using INTRAS to simulate vehicle dynamics, and DYNASMART to control vehicle routing through the network. The simulation module provides sensor data to a number of incident detection algorithms, an expert system for incident management (which also receives information from the incident detection algorithms), an origin/destination estimation module, a static traffic assignment module, an adaptive traffic signal control algorithm, and an adaptive ramp meter control algorithm. These modules process the sensor data along with data passed between them to produce new control settings and ATIS information which is fed back to the simulation module which then continues to simulate. Communication with FFTCoS is handled through a separate module allowing data to be passed between the sections.

Dynamic Assignment Algorithm
A radically new approach for Dynamic Traffic Assignment (DTA) was developed as part of Testbed research. Network assignment is accomplished for time-dependent O-D demands based on a bi-level optimization program with an analytically embedded traffic simulation. This is the first such model where simulation equations are incorporated as constraints in an optimization framework for network assignment. The assignment is performed based on link densities as opposed to link flows used in static assignments, which is a significant conceptual advance. The bi-level framework is required to estimate the nodal equilibrium arrival time estimates which fix the incident matrix between time-dependent paths and their constituent links at one level, and then to assign traffic similar to a static assignment at the other level. The time-steps of the assignment are small (on the order of 15 seconds) and thus congestion dynamics and shockwaves are captured. The algorithm has been applied to the Testbed network with considerable success, however, further work is needed to make it computationally more efficient in real-time.

Distributed Network Assignment with Decomposition
The distributed assignment framework uses ELUCIDATE capabilities for concurrent processing on multiple machines. The decomposition algorithm developed in the Testbed is hierarchical, based on physical decomposition of the network into subnetworks and creating an abstract (upper-level) network which includes only the origin-destination nodes and the gateways between the subnets. The algorithm operates iteratively. Gradient Projection is applied to the subnetworks, resulting in link characteristics for the abstract network whose assignment provides the gateway demands for the next iteration of subnetwork assignments. The subnetwork runs are carried out in a distributed fashion, which results in significant benefits in computational time. Computational tests so far indicate that the benefits depend on the number of subnetworks as well as the size of each subnetwork. This capability, though not as significant for smaller traffic networks, is crucial for the application of the ATMS concepts to larger regional networks.

Image Processing for Vehicle Tracking
The image processing algorithms developed in the testbed are based on recent advances in cognitive vision. This module has been developed as a stand-alone part of the testbed at this point and utilizes off-line stored images (videos). The reason why this has not been integrated into the Testbed is that real-time video data downloading has not been possible thus far and also because video data could not be generated using our simulation framework in a similar fashion to traffic loop data. Algorithms have been successfully developed to track vehicle movements using video and to capture lane-changing and stopping behavior for use in video-based incident detection. Real-time operation is yet to be tested.