D. E. IHiTlEY
du Pont Associate Professor
in Mechanical Engineering,
Massachusetts Institute of
Technology, Cambridge, Mass.
Assoc. Mem. ASME
I Finite State Approach to Vehicle lirgiig 1
A stati space approach is taken to vehicle merging on high speed highways.
The vehicles ate assumed to be traveling in equal sized "slots" which move at the group velocity.
At points where two or more lanes merge, some vehicles must be moved forward or backward to other slots to accomplish the merge. The state of a group of vehicles to be merged
is defined in terms of the slots occupied at any time. A finite set of admissible
terminal
stales, representing possible merged configurations, is easily determined. The sequence
of moves required to obtain a merge is found as a shortest path in the space of all states,
running from the initial state to the terminal manifold.
Various costs may be applied
to moves in this space, such as time consumed, or number of vehicles being moved
simultaneously.
Costs may also be assigned to the terminal arrangements,
reflecting,
for example, the size of platoons in the resulting merge. Estimates are made of required
computing load and the method is compared with other approaches.
Introduction
H T THE merge points in future automated highways, some method of rearranging the vehicles to allow merging
will be necessary. The approach one takes to this problem is influenced substantially by the operational characteristics assumed to be in effect away from the merge.
If one assumes that vehicles nearing a merge could have a
variety of positions and velocities, one concludes that velocities
as well as positions must be controlled by a. central algorithm
during merging. This led Athans [1| 2 to describe merging in
terms of a first order vector differential equation and to extract
merging action as an optimal control law based on a quadratic
penalty. In his treatment, a state vector of dimension 2n is
required, where n is the number of cars being merged. Real time
communication of position and velocity information between
ail the merging cars is also required.
If one assumes, by contrast, that the vehicles in normal
cruising are constrained by on-board autonomous control loops
to travel at or near group velocity V in imaginary "slots" of
length A,3 then one tends to ignore velocities and think of
merging as pure rearrangement.
Vehicles may have to be
•Work supported in part by DOT, OHicc of High-Speed Ground Transportation, contract C-85-65.
^Numbers in brackets designate Reierenees at end of paper.
"The advantages of this approach for maintenance of longitudinal control
are well known [8, 9],
Contributed by the Automatic Control Division of T H E AMERICAN SOCIETY OF MECHANICAL ENGINEEHS and presented at the Joint ASCE-AS±\IK
Transportation Engineering Conference, Seattle, Wash., July 26-30, 1971.
Manuscript received at ASME Headquarters, Marcii 13, 1972.
shifted from slot to slot in order to accomplish the merge. The
slots are synchronized by a signal traveling along the roadway
[2, 6], A predetermined velocity program in each car is called
into action to direct any shift called for by a central control
station, which also monitors progress. These assumptions led
Godfrey [3| to investigate sequential merging strategies for resolving conflicts as they came up at a merge. He proposed and
tested several decision rules for granting right of way to one or
the other of two cars arriving simultaneously and seeking to
enter the same downstream slot. Inter-vehicle communication
is not required.
Examining these two approaches, we find that neither gives
the designer direct control over how the vehicles shall move
during merging and how they shall end up. Athans explicitly
penalizes the possibility of collisions, wide variations in velocity,
and control effort. Godfrey judges his strategies by how they
avoid shifting any car too many slots. This latter cost will be
shown in the following to be expressible as a terminal cost only,
leaving one free to pick such path costs as the number of vehicles being shifted simultaneously. Athans' approach is continuous, in the sense that positions and velocities of vehicles are
monitored continuously in time, and central control is being
exerted continuously during the merge. Godfrey's approach is
discrete in that the central control calls only for cars to move an
integer number of slots, and makes such calls only at isolated
times. On-board controllers in each car exert controls to accomplish the moves.
The merging method described in the following is discrete. A
finite group of vehicles to be merged is identified and described
by a state vector. Terminal "merged" states are enumerated
and a merging control sequence is obtained as a shortest path
in the space of all possible slates, the path running from the
initial state to the terminal manifold. The motivation for this
approach is to be found in reference [7],
Journal of Dynamic Systems, Measurement, and Control
JUNE
1 9 7 2 / 147
Copyright © 1972 by ASME
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 11/02/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use
MERGE
POINT
JSLLJSLL
U-A ~4~ A -4— A -4«- A —j
Fig. 3 Transition diagram for either lane in Fig. 1
Fig. 1 Two lanes of vehicles approaching merge
STATE OF LANE 2
(decimal equivalent!
• -POSIBLE
MERGE STATE
Fig. 4 Transition diagram for either lane, showing cost equal to
number of car-slots moved
-PRESENT STATE
<B
©
®
£>
3
5
6
9
INITIAL STATE
10
STATE OF LANE I
(decimal equivalenr)
12
POSSIBLE
MERGE STATE
Fig. 2 State space corresponding to Fig. 1
Discrete State Model for Merging
Consider two lanes of vehicles approaching a merge, as in
Fig. 1. There is a conflict in the second slot upstream of the
merge. Define the state of a lane as follows. Construct a binary
number whose highest order bit represents the slot nearest, the
merge point. This bit is unity if the slot is occupied and zero if
not. Proceed upstream similarly. The resulting binary number
uniquely describes the locations of vehicles and empty slots and
is convenient for computer representation. For the vehicles in
Fig. 1, we have
state of lane 1 = 1100 = 12 w
(1J
state of lane 2 = 0110 = 6I0
(Subscript 10 denotes decimal notation.)
Assuming that we consider only the four slots nearest the merge
point in each lane, the set of allowed states for lane 1 is the same
as for lane 2: (in decimal notation)
3, 5, 6, 9, 10, 12
(2)
In Fig. 2 we show the state space for these two example lanes,
with the present state (12, 6) indicated. The terminal states
which allow merging are on the antidiagonal as shown. These
states satisfy the following relation:
state (si, s2) is a merge state if and only if
Si + s, = 15io
(3)
Here, 15 is the largest value Si or s2 could have if all four cars
were in a platoon in one lane, as they will be after being merged.
This result obviously generalizes to any such space: For a space
which represents merging N cars into a single solid platoon, a
state (si, Sa) is a possible merge state if and only if
s, + s , = 2-v
1
(4)
Of course, it is not necessary to merge vehicles into a, solid
platoon and there may indeed be good reasons not to. Platoons
will be assumed, however, in what follows to simplify the discussion.
An allowed transition in such a space is defined to be one in
which any number of cars move one slot each, all in the same
direction. Other definitions are possible, but this one will do
for illustration. Then we may draw a transition diagram for
each of the lanes in Fig. 1 (see Fig. 3). The numbers inside the
circles correspond to states. Using this diagram we can draw
148 / J U N E 19 7 2
Fig. 5 Portion of Fig. 2 showing allowed transitions and costs based
on number of car-slots moved
transition lines on the state space. Aside from cluttering the
figure, this would be incomplete until we added lines representing
simultaneous moves in both lanes, which are allowed in our
definition. Complete examples will be shown in the following.
Path Costs and Termina! Costs
We can do a lot of things with the foregoing transition diagrams. Assume first that we assign a cost to each transition
equal to the total number of slots traversed by all the cars moving
in that transition. These costs are shown in Fig. 4. If we apply these costs to Fig. 2 and allow no diagonal moves in the state
space,*1 then it is clear that, all shortest paths from initial state
(12, 6) to the terminal set are of cost = 2, and that there are
three terminal states so reached: (9, 6), (10, 5), and (12, 3). If
we allow diagonal transitions and are concerned solely with the
total number of slots moved by all the cars involved in each
transition as a cost,3 then each diagonal move must cost the
same as the shortest nondiagonal path which accomplishes the
same maneuver. The relevant section of Fig. 2 is shown in Fig.
5 to represent these facts. Clearly all shortest paths to the
terminal set have the same length, illustrating the fact proved
by Godfrey [3], that "total delay" in merging is independent of
merging strategy. Thus we could make this cost a terminal cost,
rather than a path cost if we wished.
A more interesting terminal cost is found by squaring the net
number of slots moved by each car during the entire merge, and
summing over all cars moved. For the foregoing example, this
cost is 4 for terminal state (9, 6) since one car moves 2 slots in
the transition from (12, 6) to (9, 6). Similarly, this cost is 2 for
terminal states (12, 3) and (10, 5) since here two cars move one
slot each. This cost criterion tends to favor merges in which
no vehicle is moved a long distance compared to distances
4
A diagonal move in a space such as that in Fig. 2 is one which is not parallel to either axis. A move parallel to one axis represents motion of cars in
one lane only, while a diagonal move represents simultaneous motion of cars
in both lanes.
s
Godfrey calls this the total dtlay [3].
Transactions of the ASME
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 11/02/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use
LANE a
X—X
—s<-^<
INITIAL STATE
X
X
s
x
Fig. 7 Sequence of vehicle positions corresponding to shortest
path from (12, 6) to (10, 5) in Fig. 6
[2l-TERMINAL
COST
Fig. 6 State space w i t h costs assigned to penalize m o v i n g p l a t o o n s
and favor diagonal moves
moved by the others. This is one of the criteria Godfrey uses to
judge his merging strategies.
Another useful terminal cost comes into play if we allow more
than enough room in which to place the cars. Then we may
distribute empty slots as we wish. Usually it is desirable to
avoid merging vehicles into long platoons, so we could penalize
terminal states according to the iength of platoons in them.
Terminal costs may be used exclusively to pick a terminal
state, as Godfrey does, or they may be combined with path
costs to pick both a path and a terminal state, as Athans does,
or terminal costs may be used to pick a terminal state while
path costs are used to find a desirable path to that slate. In the
first instance one does not care what maneuvers the cars go
through while coming to a merge state. In the second instance
one must choose a weight to combine the two kinds of costs
before a p a t h and terminal state can be chosen. In the third
instance one may pick the terminal state for purely configurational reasons and pick a path to that state for reasons based
only on maneuvering. Since these two considerations, maneuvering and final configuration, do not bear on each other, the third
method is a natural one. It is available to us using the methods
discussed here.
Possible maneuvering considerations include:
1 Avoid moving platoons—instead move one car at a time.
This might improve reliability. Less monitoring is needed and
slight differences in the vehicles' responses will not matter very
much;
2 Achieve the merge state in the fewest number of transitions—this means move along diagonals in state space which
tends to minimize the time needed to merge;
•3 Remove the conflict nearest the merge first. This is also
a reliability consideration.
To achieve the first of these, one method is to make the cost
of each state transition equal to the square of the total number of
slots moved by all the cars during that transition. To achieve
the third, we note that assigning the high order bit to the slot
nearest the merge point guarantees that conflict states near the
merge point will be concentrated in one easily identified corner
of state space. For the space of Fig. 2, these head-end conflict
states satisfy
,S'i and s-> > 9 simultaneously
0'')
All transitions in this region can then be assigned extra, high
costs. The result is that all shortest paths will leave this region
in as few transitions as possible, with subsequent maneuvers
being determined largely by other costs.
If should be noted that resolution of head-end conflicts is the
only consideration discussed which requires simultaneous examination of maneuvers in both lanes. In other cases, the lanes
may be considered separately. To do this, we use a condition
like (4) to enumerate the possible merge states. Then for each
pair (initial state, terminal state) we consider paths in two
Fig. 8 Transition diagram with costs to favor moving platoons
transition diagrams, one diagram for each lane. For example,
take initial state (12, 0) and terminal state (10, .">). Then we
find the minimum cost path from 12 to 10 in one transition diagram and from 6 to o in the other, add these costs to the terminal
cost on (10, 5) and compare with similar results using the other
terminal states. The lowest: overall cost; wins and designates
both the terminal state and each lane's move sequence. Since the
lanes are being dealt with simultaneously, maneuvering consideration (2) in the foregoing is achieved automatically. This
sequential solution procedure uses much less computer memory
than the overall optimal path approach employing the full
state space and usually gives optimal or near-optimal results.
As an example, assume that the cars are arranged as in Fig. 1,
and that we wish to avoid moving platoons and avoid moving
any cars excessively long distances. To counteract the time lost,
by not moving platoons, let us favor diagonal moves. Then we
could assign the costs as in Fig. 6. Note in particular the high
cost of moving, for example, from (!), 6) to (9, 3) in one transition (by moving a two car platoon) compared with the cost of
accomplishing the same net transition in two steps in which one
ear is moved at each step. It is clear that (10, 5) wins with total
cost 2 + \ / 2 - The sequence of vehicle positions is shown in
Fig. 7.
By contrast, suppose that we wished, in the interest of short,
merging time, to favor moving platoons but, in the interest of
computer memory requirements, decided to work only with
transition diagrams, and used terminal penalties as before.
Then we could assign costs as in Fig. S. Note in particular the
low cost of moving from state 6 to state 3 in one transition (via
a two car platoon) compared with the cost of doing so in two
transitions which move one car each. Now we find that (12, 3)
wins with a total cost of 2 + V 2 . Had we used the full space
we would have discovered that (10, 5) gives the same cost. The
example shows that good results, though not necessarily the
best, can be obtained by considering the much more compact
transition diagrams.
Implementation Strategy
The proposed merging method operates by looking upstream
and planning ahead the required moves. This can be implemented in the following way.
Look upstream from the last set of merged cars until a conflict
is detected. Count upstream from there, totaling cars in both
lanes and distance in slots. When the number of cars equals the
number of slots (call the number A") then these K cars can be
conceptually isolated and merged into a solid platoon merely
by rearranging them within their region of highway, the region
being A' slots long. If there are U cars in this stretch of lane 1
and L, cars in the opposite stretch of lane 2 (/,, + L% = K) then
the transition diagram for lane 1 contains K\/((Li)\(K
- U)\)
states and that for lane 2 has K\/((U)KK
— £•>)!) states. These
Journal of Dynamic Systems, Measurement, and Control
J u N E I 9 7 2 / 149
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 11/02/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use
(START)
Fig. 9
group
THIS PATTERN REPEATS
INDEFINITELY Ifi THIS ' J ^ C T I O N
M a r k o v process r e p r e s e n t i n g counting back to find a m e r g e
numbers are maximum when L, = L« = K/2, so that K is a
measure of the computing load required to merge. That is, in
states, more
that case the space would have \K\/{(K/'i)\)'!-f
than for other values of Li and LL. The number of states indicates the amount of computer memory required to store the
space and the amount of time required to find the shortest path.
See [4] for an efficient pathfinding algorithm.
To obtain an estimate of this computing load, let us assume
that cars are placed in each lane by independent Bernoulli
processes with occurrence probability ]>i for lane I and p-> for
lane 2. That is, there is a -car in a slot in lane i with probability
Pi independent of all other slots in either lane, and so on. Clearly
we require f>\ + p-. < 1 in order for merging to be possible on the
average.
Define
= Pi'l-
-
p -o,2
A
p -0.4
o
P .0.6
x
p -o.8 J
ALL FOR p.-p.
_1 1 5_
F i g . 10(a)
Pi + P-i
«i
D
P r o b a l i t y m a s s f u n c t i o n f o r K f o r v a r i o u s v a l u e s of p
P'i'l
= Ih'fh
22
u
S
u' a
lie •
e
a
a
•\f
'
a ,
a
(1
<a
w
L Js
s
V a
S
e
V,
P
- a e e
el a s c
) 1
I
,
e
|
e e
j
k
,0
s
a
e h,.
20
1 8
6
• i
e
•H
»e e
e
e e is-, i Aa s
a
<
14
s s a a j 'e n e we s ai a j
i "
—
e n e
>eg
g
e
- i we a M S e i g> •- a e
12
it si i
i s t n !1
a
a
i --s e esp
a
e
T
]
Wf
go
es a e / — > s % a
s oe '
a
s
e
ess w c
—
e we
i
i
i"
ue i a
I we t i t i
s>a e
=
e i' s
c
e
ai k
a -.
e
a'e
je
a i
i id A —
i
-. ne i a
co i J i e
e n ea i A" a
-IJ
u
e, a
i /
A w
»
a j ' as pa an e e s s i g fl w gij h e i
e-.
I <• esi ii ' i A" e x*i
i
/ w e
uc 1 e ie-.ui s
depend s i< g
'
a i l
wejk
i
a '>
0 1 0.2 C3 04 0.5 OS 0.7 0.8 09
1.0
P, +P 2
sepiua p
1 s; '
-awa
s • i j I J I ge
•><
ig
s \\s A" a
i :i o /v as
iF i g . 10{o) M e a n K a n d s t a n d a r d d e w i a t i o n <rK v e r s u s Pi + PFhus i a ' o e a i s
a
ie-s he a
- u e u< se MI , I —
p, ^ ' i s A" v\ i 1 p ai I
e ess JI
i • js
-ft
e ' >t
in sc A
e n n hi ge
e
i g
in e no e '
s i< i ' e a.a-'a e
n >e
s w e A is •, a (
I e a
previously proposed methods but has some advantages ver
oe bo i we-'
' s wa
e a se
i e i< k auea<
i" e i i u
each. As with the others, it will work well if traffic deiisoy is
procedure.
not too high.
It is easy to vary this procedure to include interspersed empty
Some advantages of this approach are:
solts in the merged group. If we wish in empty slots (in > 0),
1 A group of vehicles to be merged may be identified as far
we keep counting umil « = in. This obviously will increase A_\
in advance of the merge point as desired. This provides the
The terminal states may be identified in a similar manner ami
system with valuable look-ahead characteristics which can use
those with the spaces not desirably distributed may be eliminated
computer time advantageously and coordinate with overall
by giving them very large terminal cost.
network scheduling;
2 Regardless of the number of cars to be merged, the diSummary
mension of the state space is two;
'•> Merging strategies may be chosen to satisfy a varieU of
requirements.
Desirable features of the final merged conallows a designer considerable freedom in choosing how the vefiguration can be sought independently of desirable features of
hicles are to be maneuvered and in what configurations they
the merging maneuvers;
will emerge. This method shares some of the features of twro
1
i r<
150 / J U N E 19 7 2
Transactions of the ASME
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 11/02/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use
4 The state representation is chosen in such a way that
vehicles nearest the merge point may bo maneuvered into
merge positions before the others, an important feature if time
is short.
References
1 Athans, M., "A Unified Approach to the Vehicle Merging
problem," Transportation Research, Vol. 3, No. 1, Apr. 1909,
pp. 123-134.
2 Baumann, D.M.B., et. al., Project Melran, M.I.T. Press.
3 Godfrey, M. B., "Merging in Automated Transportation
Systems," >SeD thesis, M.I.T., Department of Mechanical
Engineering, 1908.
4 Hart, P., el al., "A Formal Basis for the Heuristic Determination of Minimum Cost P a t h s , " IEEE Trans. Si/sl. Sci. and
Ci/h., VSSC-4, July, 190S, pp. 100-107.
5 Howard, U. A., Dynamic Programming and Markov
Processes, Technology Press and Wiley, New York, 1960.
0 "Study of Synchronous Longitudinal Guidance as Applied
to Intercity Automated Highway Networks," T R W Systems
Group Report, No. 06818-W666-RO-00, Sept, 1969.
7 Whitney, I). E., "State Space Models of Remote Manipulation Tasks," 'IEEE Trans. Automatic Control, Vol. AC-14, No.
0, Dec. 1969, pp. 617-623.
8 Wilkie, 1). F., "A Moving Cell Control Scheme for Automated Transportation Systems," Transportation Science, Vol.
4, No. 4, Nov. 1970, pp. 347-364.
it Whitney, I). E., and M. Tomizuka, "Normal and Emergency Control of a String of Vehicles by Fixed-Reference Sampled-! )at a Control," in press.
Journal of Dynamic Systems, Measurement, and Control
JUNE
1972 /
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 11/02/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use
151