M&l.
Pergamon
Cornput. Modelling Vol. 22, No. 4-7, pp. 377-395, 1995 Copyright@1995 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0895-7177/95-$9.50 +- 0.00
0895-7177(95)00145-x
A Dynamic Route Assignment Model for Guided and Unguided Vehicles with a Massively Parallel Computing Architecture G.-L.
CHANG AND T. JUNCHAYA Department of Civil Engineering University of Maryland College Park, MD 20742, U.S.A. gangQeng.umd.edu
Albstract-This article presents a dynamic assignment model which has been developed primarily for ATMS/ATIS real-time applications. In order to satisfy the real-time computational requirements, the proposed model has been developed on the Connection Machine, CM-2, a massively parallel computer. Its implementation on the Connection Machine is aimed to exploit the parallel nature of the problem and to take advantage of the underlying computing architecture. The model follows an integrated assignment-simulation framework which assigns both guided and unguided vehicles to the network dynamically in both spatial and temporal dimensions. It uses a learning process in which the new route assignment uses information gained from previous iteration in computation of new paths. During each iteration, guided vehicles follow routes according to time-dependent shortest paths while unguided vehicles follow static shortest paths. Several numerical examples have been carried out to illustrate the potential applications in analyzing the network performance under various ATh,lS scenarios. Keywords-Dynamic
assignment, Parallel processing,
Route guidance.
1. INTRODUCTION With Advanced Traveler Information Systems (ATIS) and Advanced Traffic Management Systems (ATMS), the traffic control center has the ability to provide real-time traffic and offer routt: guidance instructions to ATIS-equipped vehicles. The anticipated benefits of these route guidances depend on the complex interactions arnong various factors, including the market penetration rate of ATIS-equipped vehicles, driver behavior and compliance rate, level of congestion, dynamic nafure of traffic patterns, and signal control strategies. Any dynamic assignment model destined for successful applications in ATMS/ATIS must be able to address all those critical issues. Furthermore, since such a dynamic assignment model will be used m real-time environment involving thousands of streets, intersections and vehicles, its computational speed poses another critical research issue. Although advanced parallel computing architectures can offer the required computational power with an economical price/performance factor, it will require a fundamental change in the modeling methodologies so as to achieve the full benefit of parallel processing. This article describes our efforts in exploring parallel computing architectures for the de\~elopmpnt of a real-time
dynamic
route assignment
model.
This article is organized as follows: the next section briefly reviews the existing dynamic assignment models along both mathematical formulations and assignment-simulation directions. Section 3 describes our modeling approach to the proposed dynamic route assignment model. Typeset
by A&T-QX
378
G.-L. CHANG
AND T. JUNCHAYA
Section 4 reports the parallel description of the parallel variables and their relationships. Section 5 explores the potential model applications with several experiments and their preliminary results. The conclusion and future research are outlined in the last section.
2. LITERATURE
REVIEW
Even without considering factors such as the market penetration
rate of ATIS-equipped
vehi-
cles, or driver compliance rate; the presence of time-varying travel demand and time-dependent link flow have already made the dynamic assignment model much more difficult to solve than the static one. There are relatively few optimization models developed to date which examine dynamic traffic assignment on networks and most of them emerged in the recent year. Each of such dynamic models is based on a different set of decision variables and behavioral basis; it offers different capabilities in formulating dynamic traffic systems or determining control actions. The first mathematical
programming approach to the dynamic assignment problem is due to
Merchant and Nemhauser [l]. Their model was formulated as a discrete-time,
nonconvex, and
nonlinear program, in which congestion is represented explicitly in the constraints. An efficient solution procedure for a stepwise linear version of this model was given by Ho [2]. Carey [3] reformulated the Merchant-Nemhauser problem as a convex nonlinear program, which offers the analytical and computational advantages over the original non-convex formulation. However, in extension to multiple destinations or traffic types scenarios, the need to satisfy the first-in, first-out (FIFO) traffic relation has resulted in a nonconvex constraint set [4]. Apart from the mathematical programming formulation of the discrete time problem, Friesz et al. [5] proposed dynamic extensions for both static system and static user optimization models, using the continuous time optimal control formulation which assume both the O-D demands and link flows to be continuous time functions. Their dynamic system optimization model can be considered as a continuous time version of the Merchant and Nemhauser model. Furthermore, their dynamic user optimization model is the first dynamic generalization of Beckmann’s equivalent optimization problem for static user equilibrium. Wie [6] extended the above models to include elastic time-varying travel demand which implicitly considers departure time choices. Ran and Shimazaki [7] also developed a dynamic system optimal assignment model for a simplified network with multiple origins and multiple destinations using optimal control theory. Along the same line, Ran and Shimazaki [8] and Ran, Boyce and LeBlanc [g-11] proposed several extensions of their dynamic user optimal traffic assignment models, including choices for both departure and route preference. Another research direction, introduced by Papageorgiou et al. [12], was derived from classical automatic control methods. They used a macroscopic modeling framework for time-varying traffic demands with a feedback regulation to establish dynamic traffic assignment conditions. This is carried out by using independent splitting rates of each traffic subhow with a given destination. In terms of the ATMS applications, the optimal control approach remains to be improved. For instance, it lacks an explicit first-in, first-out (FIFO) requirement, and thus models traffic congestion unrealistically. Most importantly, it does not have an efficient solution procedure for a network of realistic size. The feedback regulation approach also does not meet FIFO requirement. Even though the feedback regulation concept is attractive from the control standpoint, the formulation does not establish the underlying mathematical basis in order to estimate some key model parameters for system operations. A detailed review of these models can be found in [13]. Since it is often difficult to take all flow interactions into account with theoretical formulations, assignment-simulation framework emerged as one of the main tools to overcome those limitations. An example assignment-simulation model in the ATIS context was proposed recently by Mahmassani and Jayakrishnan [14]. Their model explicitly represents path selection decisions of individual motorists (both guided and unguided) along their path in response to supplied information. At every node where an alternate path is available, the user will switch routes if
A Dynamic Route Assignment Model the benefit modeled
from switching
with
approach
a traffic
proposed
assignment-simulation by Kaufman
t,raffic simulation computed
following
The
[17] to create
with the in-vehicle
these paths,
Otherwise
the iteration
condit,ions
until a termination
creating
condition
particularly
bilities
and route choice behavior,
mine the traffic control
estimating
trip time.
fully integrates
The next
assignment
above models
link trip time
access to an external camlot
assignment
offers several vehicles
be projected, assignment
section
an external
DRAM queues,
(Dynamic and paths
unguided varyhlg
Assignment
traffic
performance
criterion
performance
until the criterion
mathematical In DRAM,
route guidance
the above
model to clct,er-
model is down or not availthle,
to have a rudimentary a dynamic
route
in an integrated
simulation
assignment
model to determine
model
that
the anticipated standpoint,
then time-varying
It is thus desirable
for estimating
for
model.
Prom the controller
strategy.
capability.
mechanism
ii the
link trip t,ime for the dynnmic,
trip time.
Model)
is an assignment route
guidance
a road network.
[18] which considers
performance
that
predicts
control
It is an extension DRAM
reiterates
predicts
assignment
the predictions
is met.
Note that the iteration
solution
due to the difficulty
Given
paths
and .work t,ime-
for vehicles
mechanism. of paths
process is employed in obtaining
flows,
center)
of the earlier
only one type of vehicles.
using event-driven
the model
model
from traffic
IF the
and traffic
only to search
an elegant
and efficient
solution. unguided
and static
shortest
both the signal queuing
In each iteration, the initial
vehicles
are assigned
link travel time obtained
bot,h time-dependent
for actual
However,
traffic simulation
guided vehicles are assigned to follow time-dependent
the instantaneous
While
efficiency.
APPROACH
can receive
is not satisfied,
link t.ravel time, whereas
mcluding
and the new set of traffic pro-
for both guided and unguided vehicles,
network’s
for the optirnal
(e.g.,
and Zhuang
O-D demands
and estimates
are
capa-
decisions.
mechanism
as they travel through
Junchaya
paths
Concept
of guided
vehicles
by Chang,
Route
vehicles
shortest
and thus lacking the “look-ahead!’
of route guidance
Modeling
it simulates
over the mathematical
mechanisms traffic
model to have a rudimentary
Fundamental
advantages
model
3. MODELING 3.1.
based on the shortest
then the loop is terminated.
guidance
model is down or not available,
for evaluation
for
with various classes of route guidance
will introduce
from the assignment
simulation
be projected
policy
occurs.
and simulation
require
resulting
paths,
was recently
routing
The time-dependent
if the access to a simulation
for the dynamic
are
(MPSM)
INTEGRATION
The
Using INTEGRATION,
models still depend on an external
then all future link trip times cannot It is t,hus desirable
data.
and also in terms of computing
perspective,
procedure
system is then determined
the new route
in terms of representing
two assignment-simulation
system
Model
by running
set of forecast
trip time.
framework
grams,
begins
agree with the previous
continues,
assignment-simulation
process
new link trip time.
If the new paths
of the traffic
Simulation
based on an iterative
iterative
information
and produces
The dynamics
logic.
an initial
under time-dependent
then recomputed.
The
threshold.
from the Macroparticle
model framework
et al. [16].
model
vehicles equipped
The
a certain adapted
[15], and follows a fixed time-step
Another
paths
exceeds
simulator
379
shortest
to follow static
at the time of their departure.
path takes
into account
delay and the oversaturated
paths using predicted
shortest
the vehicle
spillback
paths
based on
The computation intersection
deployment,
and later departed used to progressively
may yield paths
it provides
traffic
an insight
along different
reduce
and network
uncertainties
performance
into the complex
paths.
that
interaction
The new time-varying
in a subsequent
prediction
delay
delay.
guided and unguided vehicles are assigned on both spatial and temporal iterations
of
basis.
are not satisfactory between
the current
link information of paths
for both
is then guided
380
G.-L.
CHANG AND T. JUNCHAYA
and unguided vehicles. Ultimately, near user-optimal paths for guided vehicles are produced at the end of the iterative process. More specifically, the proposed model differs from other assignment- simulation in the following respects: 1. DRAM
has a route loading mechanism to estimate link trip time rather than depending
on an external simulation model. 2. It uses speed-concentration
function to estimate links trip time and realistically consider
the link capacity constraints. 3. It incorporates a learning process, where information from previous iterations are used to obtain better path selections and assignment. 4. Intersection control and queue buildup are explicitly represented in the model. Thus, the signal delay and queuing delay can be realistically estimated. 5. The model assigns ATE-equipped vehicles to routes with the “look ahead” feature. Nonequipped vehicles are assigned to other predetermined routes.
PROPOSED METHODOLOGY
I
4
PHASE 1 PATH PREDICTION
+
PHASE 2
+
INTERNAL ASSIGNMENT
Figure 1. Outline of DRAM
3.2.
overall structure
Model Structure
The overall structure of DRAM is outlined in Figure 1. The inputs are time-varying traffic demand for both guided and unguided vehicles. Predefined route plans for unguided vehicles can also be used instead of assuming static shortest paths. DRAM consists of four phases: path prediction, internal assignment, evaluation and learning, and implementation. These phases are described as follows: PHASE
1:
PATH
PREDICTION
The current version of DRAM consists of two classes of vehicles: guided and unguided vehicles, depending on whether drivers have direct access to the real time information. In this study, guided vehicles are assigned paths that are fastest in a time-dependent network using predicted information. Unguided vehicles are assigned static shortest path computed from instantaneous link information at their departure time. For example, the travel time of time-dependent shortest path through link 1, 2, and 3 at time t is computed as follows:
*DTTl-3(t)
=
c1[tl + cz [t + cl(t)]
+ c3 [t +
cl(t) + c2 (t + q(t))]
.
A Dynamic
Route
.4ssignment
delay nodeexit time by At
Model
381
Assign trafiic to next link on path
3 compute signal b
updatenode exit the
queue delay
ib Figure
2. Internal
Assignment
Logic
Flowchart.
Similarly, the travel time of static shortest path through link 1, 2, and 3 using instantaneous travel time information at time t is represented as follows: STTl_3(t)
=
c1[t] + ca[t] + cg[t],
where TDTTr_a(t)
= Time-dependent travel time from 1 to 3 at time t,
SSTl-3(t)
=
G(t)
= Travel time on link i at time t.
Static travel time from 1 to 3 at time t,
From this phase, an initial route plan will be est,ablished for each O-D pair at each time period. Each O-D pair will also have a pointer that keeps track of its current position on its route plan that will be updated during the course of internal loading. 2: INTERNAL ASSIGNMENT(AN EVENT-DRIVENsIh4uLmIoN MECHANISM) The internal assignment phase assigns time-varying travel demands along the predicted paths for both classes of vehicles. The assignment procedure is similar to event-based simulation where the assignment takes place along both spatial and temporal dimensions. After each loading and assignment step, DRAM predicts time-varying flows, queues, and travel time in the network according to the time-varying distributions and evolution of vehicles. These results will be used internally for evaluation in the next phase. A logic flowchart of the internal loading process is shown in Figure 2. The process starts by finding all those vehicles with their earliest link exit time. Each vehicle has a pointer to its PHASE
382
G.-L. CHANG
AND T. JUNCHAYA
vehicle enterlink
2 Figure 3. A graphical illustration of link travel time and intersection components in overall link trip time.
delay time
current link along the path and the exit time from this link. When its exit time is the earliest, one, this vehicle is loaded to the next link in the route plan, provided that there is enough available capacity left on that link. The process continues by loading all vehicles with the same link exit time. If some of the vehicles cannot be loaded because of link concentration constraint,, their exit, times from the current link is delayed by one time interval. When all vehicles with the same link exit time have been loaded, we would like to estimate the time these vehicles will stay in this link and their new link exit time. Assume that subsequent vehicles have no effect on the speed of current group of vehicles in the link, we model link trip time as the sum of two components: a travel time along the link and the time delayed at the downstream intersection (Figure 3). The travel time of vehicles along a link depends on the average speed given by the macroscopic speed-concentration relation [15]. The function used in DRAM is given by
t
?I; = (Vf -v,)
2
P.
( >+ 1-
w,,
(1)
0
where Vf
=
the mean speed in Section i during the tth time step;
vf, v, = the mean free-flow speed and minimum speed on the facility, respectively; k,
= the maximum or jam concentration;
k;
= the concentration at the entry time; and
Pi
= a location dependent parameter.
The intersection delay is further separated into two components: signal delay and queue delay. The delays are estimated at the time when vehicles arrive at the end of the link. The total signal
A Dynamic Route Assignment
Model
383
delay varies with their arrival time at the intersections. For instance, if vehicles arrive during the red phase, they will be subjected to signal delay and queuing delay, depending on the number of vehicle waiting in queue at that time. Finally, DRAM can update the new departure time of these vehicles from their current links after computing link travel time, signal delay and queuing delay. These vehicles are subjected to another assignment when their departure time become the earliest one among all time-varying demands. DRAM will repeat the internal loading process for another group of vehicles with the same link exit time. If the link exit time is still within the projected time horizon, these vehicles are assigned to links. Otherwise, the internal assignment stops and we continue to the next phase. The procedures for internal loading phase is summarized in Table 1. Table 1. Internal Assignment Step
1
Select an event with the earliest link entry time from the event list. If the link entry time is greater than the projected time period, then go to Step 7, else go to Step 2.
Step
2
Check the link’s capacity for the selected event (i.e., O-D travel demand) to enter. If there is enough capacity, assign this event to the downstream link, otherwise increment the link entry time by one time interval.
Step
3
Check the event list for another event with the same link entry time. If there is one, repeat Step 2, otherwise go to Step 4.
Step
4
Compute the link speed and travel time subjected on each link.
Step
5
Compute
Step
6
Compute the intersection delays (i.e., signal, queue, and queue spillback delays), depending on the projected intersection arrival time computed in Step 5.
Step
7
This step updates various information down into the following processes:
.
PHASE
Procedure.
to the assigned number of vehicles
the arrival time of vehicles at the downstream
intersections.
for the next assignment of events. It is broken
Update the new link entry time of events selected in Step time and intersection delays;
.
Reinsert the events selected in Step
.
Update the number of vehicles in the corresponding ing to the duration that those vehicles will stay.
2 based on the link travel
2 into the event list; and links at each time period accord-
Step
8
Go to Step
Step
9
Finish the assignment of all travel demands within the projected time period, compute network system performance data during one iteration of the iterative route assignment model.
3:
EXALUATION
1.
AND LEARNING
At the beginning of the previous phase, the paths for guided vehicles are based on projected time-dependent link travel times. However, after the completion of the internal loading phase, all the projected time-varying O-D demands have been loaded to the network and the time-dependent travel times are now based on “actual” traffic loads. The anticipated travel time can then be compared with the “actual” travel time. If the difference is above the predefined threshold, then the assigned. routes for guided and unguided vehicles have not reached stable traffic patterns, and the iterative process continues in repeating the path prediction module. However, if the difference is below the threshold, the predicted routes for guided vehicles have stabilized and can be broadcast to guided vehicles as the real route guidance instructions. The motivation of an iterative assignment process is due to the significance of information gamed from each path confirmation, namely the “actual” time-dependent travel times. By assigning the time-varying O-D demands to the network along anticipated paths, some traffic patterns will be developing and will not change in the next iteration. Thus, this information can be used to improve the prediction of paths for guided and unguided vehicles in the next iteration. Thus, the uncertainties in the prediction of time-dependent travel time are progressively reducing as more tral% patterns are developed.
384
G.-L.
CHANG AND T.
JUNCHAYA
Currently, traffic signal operation in DRAM is modeled as an isolated intersection with simple two-phase operation with the same cycle length for all intersections. The phase duration is adjusted at the end of each iteration based on the corresponding traffic flow volume. Using the following algorithm that relates the duration of a phase split to the projected volume of incoming flow, we can adjust the phase splits after each iteration. At each intersection,
there are four approaches: north-south, south-north,
east-west, and west-
east. The total flows in each approach are tabulated and the maximum flow between the total flows of north-south and south-north direction, (fmax~_s), and between east-west and westeast direction,
(f max~_W), are computed. The green time for the north-south direction can be
computed as follows:
= f
QN-S
m=N-S
f m=N-S
+f
m=E-W
rN_s = Cycle length - gN_s.
1
* Cycle length.
The green and red time in the east-west direction are simply the reverse for the north-south direction. Note that, the above algorithm is rather simple and will be used only to illustrate the capability of DRAM in adjusting signal splits based on predicted flow. A more elaborated algorithm that make better use of the information from preceding iterations to adjust signal splits and/or cycle length can certainly be incorporated
as required.
PHASE 4: IMPLEMENTATION When the differences between anticipated and actual travel times are minimized through successive learning and iteration, the traffic control center will be able to predict the traffic patterns and network performance based on the current predicted time-dependent shortest paths. These final time-dependent shortest paths can then be recommended to ATIS-equipped vehicles.
4. DEVELOPMENT OF A DATA PARALLEL ASSIGNMENT MODEL One of the most critical components in designing a parallel model for SIMD (Single Instruction/Multiple Data) computers is the specification of parallel variables. These parallel variables will determine the effectiveness of the parallel model since different selections of parallel variable will lead to different communication patterns, and/or algorithms. 4.1.
Parallel
Variables
There are four main parallel variables used in the proposed model: network link performance, intersection, dynamic demand, and time-varying path parallel variables. Their data structure and relationships are discussed below. 4.1.1.
Network
Link Performance
Parallel
Variable
(NLP-PV)
The NLP-PV is shaped as a two-dimensional parallel variable as illustrated in Figure 4. The first dimension is equal to the number of maximum links in the network, and the second dimension is equal to the number of predetermined projected time periods. The NLP-PV is used to hold time-varying link information during the projected time period. The information for each link includes length, capacity, maximum and minimum speed, beginning node and ending node of each link. 4.1.2.
Intersection
Parallel
Variable
(I-PV)
The intersection parallel variable is also shaped as a two-dimensional parallel variable. The first dimension is equal to the number of maximum nodes in the network, and the second dimension is
A Dynamic Route Assignment
. . . . . .
385
Model
llnk h
..I...
T+AT
I
T+2At
I
I
I
T+SAT if 3
:
if II
T+ PAT
EACH CELL CONTAINS:
tink ID Be&nw-&~~M~e Maximum & Minimum Speed Link Speed Equation Parameters Current Number of Vehkles Current Speed & Travel Time Queueing & Signal Delay
Figure 4. A graphical illustration of network link performance
NodeNcde2
. . . . . .
parallel variable.
Nodeh
T+AT 3
4
T+2AT T+3AT
ii
:
%
:.
a
T+ pAT
. . . . . .
EACH CELL CONTAINS:
Link Indices for all links feeding this node all links exiting this node Signal Control
Figure 5. A graphical illustration of intersection parallel variable.
equal to the number of projected time periods. Each element in the parallel variable corresponds to an intersection. The information in each element includes signal control information, indices for links enter and exit a node, and the time-varying number of vehicles of incoming links. Figure 5 shows the general layout of I-PV. 4.1.3.
Dynamic Demand Parallel Variable (DD-PV)
The dynamic demand is structured as a three-dimensional parallel variable as shown in Fig;ure 6. The first and second dimensions are equals to the number of maximum nodes in the network. The third dimension is equal to the projected time periods. Each element of DD-PV holds information about origin-destination (O-D) trips in each time period. This information consists of the number of vehicle trips for each O-D, current link in the path, and exit time from current
386
G.-L. CHANG
AND
T. JUNCHAYA
T+PAT
DYNAMIC DEMAND PAFIALLEL VARIABLE ynb&ofv&k&w~m~and CurrentLhk Current and Next Node Spillbach Delay Trip Statistics TIME-VARYING PATH PARALLEL VAFIIABLE Link ID Pati Travel Time Next Node on Shortest Path Figure 6. A graphical illustration of dynamic demand and time-varying path parallel variable.
link. Each type of vehicle (e.g., guided and unguided) will require separate type of vehicle will have different demand and will follow different path. 4.1.4.
Time-Varying
DD-PV,
since each
Path Parallel Variable (TP-PV)
The time-varying path parallel variable is also structured as a three-dimensional parallel variable and share the same dimensions as the DD-PV. Each element holds path information for each vehicle type. The information includes path costs, and next node to follow. Thus, for each vehicle type, there is a corresponding pair of DD-PV and TP-PV. In this study, we have two pairs of DD-PV and TP-PV, one pair for guided vehicles and the other for unguided vehicles. 4.2. Parallel Processing In this section, we will discuss the main operations of Phase 1 and Phase 2 that has been parallelized. The main task in Phase 1 is to compute path for each type of vehicle, while Phase 2 assigns vehicles according to path predicted in Phase 1. 4.2.1.
Path prediction
There are two types of path computation in this study: static shortest path and time-dependent shortest path. Habbal et al. [19] have parallelized the all-pairs static shortest path algorithm the SIMD computing architecture using a two-dimensional data structure and network decomposition with the following logic for the parallel model: fork=
l,... processor
, N do in parallel Pi do the following:
Chang et al. [18] extended Habbal’s algorithm to compute all-pairs time-dependent path with the FIFO requirement using a three-dimensional data structure. The parallel all-pairs, time-dependent shortest path algorithm is shown as follows:
shortest logic for
A Dynamic Route Assignment
for t := T + .NPAt, . . , T (End of projection
Model
387
period ---> Beginning of projection
period)
fork = l,...
, N do in parallel processor Pi do the following: &(t)
= min (c$‘(t),
d:;‘(t)
+ d&l (t + d:;‘(t)))
‘d’ifkfj
where
travel cost from node i to node j through node k dfj (t): N: Number of nodes T: starting time of the projection period NP: number of projection periods At: length of one projection period In the beginning, all-pairs, time-dependent shortest paths for time period T + NPAt computed first, followed by the paths computation for the time period T + (NP - l)At, on until all paths in the first projection time period T have been computed. 4.2.2.
are
and so
Parallel internal loading process
In the sequential process, each demand with the same link exit time is loaded in succession. However, with this parallel implementation, all O-D demands with the same link exit time can be loaded in parallel. Each element of the DD-PV has the information about the number of vehicles in each O-D demand and the next nodes on the time-dependent shortest path. Figure 7 illustrates the parallel loading of demands from DDA-PV to links in NLP-PV using this information as indices in the interprocessor communication.
AU O-D DEMANDS WITH DEPARTURE TIME -T ARE AMGNED NEWORK
SIMULATNEOUSLY
UNK PERFORMANCE IS UPDATED FROM ASSIGNMENT IN PARALLEL
Dynamic Demand Assignment Parallel Variable
Network Link Performance Parallel Variable Figure 7. Parallel operations
between DD-PV
and NLP-PV.
After all demands with the same link exit time have been loaded on the network, the the prevailing ispeed and travel time for all links in NLP-PV can be computed simultaneously for the current time period. Furthermore, signal delay, queuing delay, and the links trip time for these demands can be estimated in parallel. Finally, the number of vehicles loaded in this period can be used to update the links information in the subsequent time periods for the duration of their links trip Cme.
G.-L. CHANG AND T. JUNCHAYA
388
4.2.3.
Parallel signal control and adjustment of phase splits
In this paper, signal control at all intersections are assumed to operate individually. With this assumption, the signal control process is inherently parallel and thus the signals can be structured as a parallel variable. With the use of intersection parallel variable, two main operations involving signals can be parallelized.
The first operation involves the internal assignment process where
each O-D demand has reached the end of the link, and need to check the signal control status to determine whether or not it is possible to leave. This checking process is done in parallel where each O-D demand in DDA-PV communicate with I-PVV to check the condition of all signals. The other operation involves the adjustment of phase splits. This operation is parallelized by letting the I-PV communicate with NLP-PV to obtain link flows information from all approaching links. Each element of the I-PV simultaneously determines the maximum flow and computes the flow ratios among all approaches before adjusting the phase splits for the next iteration.
5. ILLUSTRATIVE
EXAMPLES
The proposed dynamic route assignment model can be used to analyze the traffic network performance under various ATMS traffic scenarios. Each scenario can incorporate factors such as (a) the market penetration rate of guided vehicles, (b) levels of traffic congestion, and (c) traffic arrival patterns. The route guidance or signal control strategies can also be chosen as control variables though they are fixed in the current study. The proposed model can generated various measure of effectiveness (MOE) that can be used for policy analyses. For instance, the overall benefits of ATMS/ATIS under different levels of market penetration rate of guided vehicles can be estimated with the total system travel time. While this indicator gives some basic insight about the potential benefits of ATMS-ATIS, there are still several unanswered questions concerning the relative performance of guided vehicles and unguided vehicles. Hence, the following outputs, aggregated over all O-D pairs and the projected time period can be used as the basis for exploratory analyses. (a) total O-D distance travel from all O-D pairs for equipped and unequipped vehicles; TDTe = TDT” =
c allO-D
&LLt pairs
%ant
c
all O-D pairs
(b) total O-D travel time from all O-D pairs for equipped and unequipped vehicles (including signal and queue delay); TTTe =
en*t
c
all O-D pairs
TTT” =
c allO-D
pairs
ttMnt
c allO-D
pairs
(c) total spillback delay from all O-D pairs; TSDe = TSD” =
sG,t
c all O-D pairs
SGmt
A Dynamic Route Assignment
Model
389
(d) average speed (a/(b + c)); AvgSpe = AvgSp” =
(e) number of trips that reach destinations
TDT” TTTe + TSDe TDT” TTT”
+ TSD”
for equipped and unequipped vehicles
(f) total vehicle travel time TVTTe
=
c
4nnt
. cmt
C
n4knt . %A
all O-D pairs
TVTT”
=
all O-D pairs
where e
ZZ
equipped vehicles; unequipped vehicles; dt&,, dt;,,, = distance traveled by a equipped/unequipped vehicle between origin m and destination n starting at time period t; travel time of a equipped/unequipped vehicle between origin m tth,,,, tt;,,, = and destination n starting at time period t; spillback delay of a equipped/unequipped vehicle between sd&tr ad;,,, = origin m and destination n starting at time period t; number of equipped/unequipped vehicles between origin m and nuKnt, m&t = destination n starting at time period t; In addition, DRAM can produce various MOEs output at an individual O-D level. The output at this level may be necessary for some special needs. =
U
Thus, the following questions can be answered with DRAM: What is the optimal market penetration rate of guided vehicles for a particular network with certain O-D trip rate? (b) Who receives the potential benefits under an ATMS/ATIS environment (e.g., guided or unguided vehicles)? Cc)Is there any tradeoff in the network performance (e.g., shorter travel time but longer distance)? (4 What is the distribution of performance gain among O-D pairs?
(4
The dynamic route assignment model have been implemented on the Connection Machine CM-2 configured with 16,384 processors running at the clock speed of 7MHz with code written in C* (Thinking Machine, 1990), an ANSI C standard with parallel extension. 5.1.
Experiment
Designs
To illustrate the potential applications of DRAM, the following experiments have been carried out with the proposed model, and implemented on the Connection Machine CM-2 with the code written in C* (Thinking Machine, 1990), a data parallel C extension of an ANSI C standard. The CM-2 is configured with 16,384 processors running at the clock speed of 7MHz. Note that the model properties have already been investigated by Chang et al. [18]. The model was shown to have a desirable property which converges to a near optimal loading pattern. Furthermore, the running time of a data parallel model shows a dramatic improvement over the sequential counterpart.
G.-L. CHANG AND T. JIJNCHAYA
390
5.1.1.
Network
The test network (Figure 8) is a hypothetical grid network with 30 nodes (5x6) and 98 links. There is a two-way street connecting each adjacent pair of nodes. This two-way street is represented by two uni-directional links. The free-flow speed on all links is specified as 30mph with a minimum speed of 5 mph. Each link has three lanes, and its length is arbitrary assigned to be between 2000-2500ft long. Each intersection in a network is controlled by a two-phase signals. The cycle length is fixed at 90 seconds, and initial phase duration is evenly distributed between green and red at 45 seconds. ing vehicles
for each 5 minutes
phase duration
by 5 seconds
DRAM will adjust time interval.
signal phasing
The adjustment
with the minimum
according
to number
of approach-
will increment/decrement
previous
green set at 20 seconds.
a a a a
ORIQIN
& DESnNATlON
NODE
DEsmNATlON
Figure 8. Test network and O-D trip patterns boundary nodes.
5.1.2.
Origin-destination
NODE
0
from north boundary
THROUGH
NODE
nodes to south
demand
The O-D trip patterns used in this study are fixed between nodes along the boundary of the network and along four directions (north-south, south-north, east-west, and west-east). Figure 8 shows the O-D trips for these nodes and the direction from north to south. In each time period, there will be trips between 78 O-D pairs. There
are two types
of vehicle
currently
supported
in DRAM
(guided
and unguided).
While
we assume the same ratio of guided/unguided vehicles for all O-D pairs. We vary the ratio of guided/unguided vehicles into six levels of market penetration (0, 20, 40, 60, 80, and 100%). Other factors involve in setting the O-D demand is the demand volume and its distribution over the projected time period. The projected time period is 60 minutes long and is divided into two half hour periods. For evaluation purpose, we assume that all O-D trips will arrive during the first half hour and none will arrive during the second half hour. Figure 9 shows the arrival pattern used in this study. This pattern assumes uniform arrival pattern during the first half hour. There are five levels of demand at 10, 20, 30, 40, and 50% of maximum flow within this pattern (abbreviated as U-10, U-20, U-30, U-40 and U-50, respectively. We assume the maximum flow to be lBOOvph/lane. In a real-world scenario, the demand volume and distribution for each O-D pair may vary substantially.
A Dynamic
IO
Route Assignment
20
Model
391
40
50
60
Time pZod (min) Figure 9. Uniform arrival pattern
5.2.
of vehicles at different demand level
Results
Note that the results and interpretations of the potential ment,
applications
and not intended
Table
2 presents
ket penetration. almost
for U-20.
the total system
Otherwise,
penetration
the network
At medium
vehrcles,
the performance
significantly
pattern
hicles for U40,
it improves
degradation
the total system Tables
demand
p.attern
Tables similar
is divided
performance
vehicles.
While
from 12-14%
pattern
of trips finished.
improves
vehicles
is that
all cases
benefits
for a total following
guided
of IVHS,
among guided and unguided
for six demand
and number
distance,
for
at 20% guided ve-
into the potential
5), the guided vehicles
travel similar
slightly
of 100% guided vehicles.
penetration
speed,
de-
(U-40 and U-50’), there
deteriorates
are distributed
vehicles
in
At 100% guided
for all arrival patterns
At low demand,
(distance.
improves
for U-10 and from 7-10%
performance
consists
and unguided
(Table
environ-
levels of mar-
performance
has no improvement.
will give some insight
into six levels of market
both type of vehicles
speed and number
trend
the guided and unguided
in all categories
at different
While the performance
flow, respectively.
demand
pattern
For high levels of demand
when the demand
the guided
under an ATMS
the system
the system
One common
performance
3 and 4 compare
At 30% uniform
improves
to study how these benefits
3-8 compare
a.nd 20% of maximum
degrades.
for U-50.
system
at 100% guided vehicles where the performance
(U-30),
of improvements.
it may also be of interest vehicles.
except
and U-20),
then at SO-SO%, the performance
is no consistent
show the performance
(U-10,
performance
level of demand
guided vehicles,
a network
policy analysis.
travel time for six demand
At low level of demand,
20-40%
While
given in the following section are mainly for illustration in analyzing
to be an indepth
all llevels of market
teriorates.
of DRAM
patterns
of 36 cases.
uniform
demand
and unguided of finished start
Each
have
trips).
to outperform
guided vehicles
at 10%
vehicles
unguided
have slight
edge in
G.-L. CHANG
392
AND T. JUNCHAYA
Table 2. Total system travel time (hr) and relative performance.
Table 3. Guided vs. Unguided vehicles with uniform arrival pattern at 10% demand level. Market penetration
MOE 0
20
40
of guided vehicles (%) 60
80
100
Table 4. Guided vs. Unguided vehicles with uniform arrival pattern at 20% demand level. Market penetration
MOE
Guided
Unguided
of guided vehicles (%)
0
20
40
60
80
100
Distance
0
6072
5931
5936
5992
5992
Speed
0
26.08
26.33
26.87
26.40
26.26
Trips
0
2340
2340
2340
2340
2340
Distance
6040
6066
5926
5932
5974
0
Speed
26.45
26.30
26.38
26.43
26.28
0
Trips
2340
2340
2340
2340
2340
0
Table 6 presents the results at 40% level of maximum flow. Guided vehicles clearly outperform unguided vehicles, covering equal amount of distance in less time thus yielding better average speed. The guided vehicles also finish more trips than unguided vehicles except at 60% market share where the unguided vehicles have a very slight edge. However, total system travel time does not improve with the introduction of guided vehicles and became worse as the network consists of 100% guided vehicles. At 0% level of guided vehicles, the unguided vehicles achieve very good
A Dynamic Route Assignment Model Table 5. Guided vs. Unguided vehicles with uniform level.
393
arrivalpatternat 30% demand
Table 6. Guided vs. Unguided vehicles with uniform arrival pattern at 40% demand level. Market penetration
MOE
of guided vehicles (%)
0
20
40
60
80
100
Distance
0
5962
5735
5949
5495
5704
Speed
0
24.12
23.13
26.08
22.21
21.63
Trips
0
2179
2035
2256
1933
2005
Distance
5992
5711.
5676
5990
5545
0
Speed
25.68
20.86
20.90
24.50
17.15
0
Rips
2302
1980
1984
2265
1898
0
1
Guided
Unguided
Table 7. Guided vs. Unguided vehicles with uniform arrival pattern at 50% demand level. Market penetration
MOE
Guided
Unguided
of guided vehicles (%)
0
20
40
60
80
100
Distance
0
5428
5728
5124
5026
4571
Speed
0
21.23
22.10
19.09
17.65
13.08
Trips
0
1784
1896
1505
1388
1066
Distance
5338
5060
5625
5052
5092
0
Speed
18.12
16.75
18.60
13.80
11.15
0
Trips
1694
1549
1815
1390
1335
0
performance which deteriorate as market share of guided vehicles increase. This implies that guided vehicles improve their performances at the expense of unguided vehicles. The results for 50% maximum flow in Table 7 illustrates similar patterns as Table 6. Guided vehicles outperform unguided vehicles at all level of market share except for 100% level. The guided vehicles travel faster than unguided vehicles while reaching more destinations. The total system travel time improves by 4% at 20% market share, remain stable at other level except at 100% level where the performance worsens. In addition to the simplified arrival pattern and common ratio of guided/unguided vehicles for each O-D pair, this study focuses mainly on the path assignment with minimal attention being focused on the setting of signal control. Signal and queue delays are significant parts of
G.-L. CHANG AND T. JUNCHAYA
394
travel time, and with proper signal control strategies, the network performance m$y improve tremendously. It is anticipated that route assignment when properly coordinated with adaptive signal control will show improved performance over the current results. However, adaptive signal control
is beyond
the scope of this study
and should
be addressed
in future
ATMS/ATIS
studies.
6. CONCLUSIONS This article puting
presents
architecture.
a dynamic
route assignment
Its implementation
model with a massively
on the Connection
Machine,
parallel
in particular
SIMD comof the design
of parallel variables, fully exploits the parallel nature of the problem and takes advantage of the underlying computing architecture. The proposed model also addresses several critical issues that are needed to determine the potential benefits of ATMS/ATIS under various traffic scenarios. In particular, various
market
the model has the capabilities penetration
rates
to analyze
of ATIS-equipped
the network
vehicles,
performance
different
path
with regard
selection
to
strategies
for multiple classes of vehicles, and various levels of congestion. Furthermore, the output can be analyzed at different levels of detail, ranging from an individual O-D pair, vehicle class, to an aggregated network level. The proposed preliminary model does not provide in-depth study in some areas, in particular, the treatment of signal control. However, the model platform is flexible enough so that more complicated models can be utilized in future work to provide more realistic results. Furthermore, the results indicate that most of the improvement received by guided vehicles is at the expense of unguided vehicles. While definitive conclusion cannot be drawn from these results, they provide some guidelines in carrying out future research. We are currently pursuing two lines of research in order to improve the model. The first research direction addresses signal control and route guidance aspects. Signal and queue delays constitute the major part of travel times in an urban street network. Thus, by coordinating route guidance strategies with the adaptive signal control and freeway control strategies, the network performance could be significantly improved. The results also show that a single route guidance strategy may not be adequate for guiding vehicles from all O-D pairs. Hence, a network’s performance may be improved when guided vehicles received route guidance instructions that depend on the distance to destination, levels of congestion, etc. The second line of research explores the development of dynamic route assignment model with MIMD parallel computing architecture. The main computation issues for the MIMD programming model are synchronization and load balancing among processors. An exploration of using the MIMD machines is currently conducted in parallel to the development of a SIMD model. These extensions and further optimization of the parallel models are required before the actual implementation of dynamic route assignment or real-time traffic simulation models can be successfully implemented in an ATMS/ATIS environment.
REFERENCES 1.
2. 3. 4. 5. 6. 7.
D.K. Merchant and G.L. Nemhauser, A model and an algorithm for the dynamic traffic assignment model, Transportation Science 12 (3), 183-199 (1978). J.K. Ho, A successive linear optimization approach to the dynamic traffic assignment problem, Ransportation Science 14, 295-305 (1980). M. Carey, Optimal time-varying flows on congested networks, Operations Research 35 (l), 58-69 (1987). M. Carey, Nonconvexity of the dynamic assignment problem, ZVunsportation Research 26B (2), 127-133 (1992). T.L. Priesz, F.J. Luque. R.L. Tobin and B.W. Wie, Dynamic network traffic assignment considered as a continuous time optimal control problem, Operations Research 37 (6) (1989). B.W. Wie, Dynamic analysis of user optimized network flows with elastic travel demand, Presented at the 7oth Annual ‘Z’RB Meeting, Washington, DC, (1990). B. Ran and T. Shimazaki, A general model and algorithm for the dynamic traffic assignment problems, In Pmceedings of the Fifth World Conference on tinsport Resemrch, Yokohama, Japan, (1989).
A Dynamic Route Assignment 8. 9. 10. 11. 12. 13.
14. 15.
16.
17. 18. 19.
Model
395
B. Ran and T. Shimazaki, Dynamic user equilibrium traffic assignment for congested transportation networks, In Proceedings of the Fifth World Conference on Tkansport Reserarch, Yokohama, Japan, (1989). B. Ran, D.E. Boyce and L.J. LeBlanc, Dynamic user-optima1 departure time and route choice model: A bilevel, optimal-control formulation, Annals of Operations Research (submitted). B. Ran, D.E. Boyce and L.J. LeBlanc, (1992). B. Ran, D.E. Boyce and L.J. LeBlanc, An new class of instantaneous dynamic user-optimal traffic assignment models, Operation Research 41 (l), 192-202 (1993). M. Papa.georgiou, A. Messmer and H. Senninger, Dynamic Network Traffic Assignment Via Feedback Regulation ( 1990). H.S. Mahmassani, S. Peeta, G.L. Chang and T. Junchaya, A review of dynamic assignment and traffic simulation models for ATIS/ATMS applications, Technical Report DTFH61-90-R-0074-1, University of Texas at Austin, (1992). H.S. Mahmassani and Jayakrishnan, System performance and user response under real-time information in a congested traffic corridor, l%ansportation Research (1990). G.L. Chang, H.S. Mahmassani and R. Herman, A macroparticle traffic simulation mode1 to investigate peak-period commuter decision dynamics, Transportation Research Record 1005, TRB, pp. 107-121, National Research Council, Washington, DC, (1985). D.E. Kaufman, R.L. Smith and K.E. Wunderlich, An iterative routing/assignment method for anticipatory real-time route guidance, In Vehicle Navigation & Information Systems Conference Proceedings: Part 2. pp. 701-708, (1991). M. Van Aerde and S. Yagar, Dynamic integrated freeway/traffic signal networks: A routing-based modeling approach, Transportation Research 22A, 445-453 (1988). G.L. Chang, T. Junchaya and L.M. Zhuang, A user-optimum route navigation model with a massively parallel computing architecture, Zbansportation Planning and Technology 17 (to appear). M. Habbal, H. Koutsopoulos and S.R. Lerman, A decomposition algorithm for the All-Pairs Shortest Path Problem on massively parallel computer architecture, Intelligent Engineering System Laboratory Report lESL91-09, Massachusetts Institute of Technology, (1991).