Game theoretic approach for predictive lane-changing and car-following control

Transportation Research Part C 58 (2015) 7392
Contents lists available at ScienceDirect
Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc
Game theoretic approach for predictive lane-changing

and car-following control
Meng Wang a,b,, Serge P. Hoogendoorn a, Winnie Daamen a, Bart van Arem a, Riender Happee b
a
Department of Transport and Planning, Delft University of Technology, Stevinweg 1, 2628 CN, Delft, The Netherlands
b
Department of BioMechanical Engineering, Delft University of Technology, Mekelweg 2, 2628 CD, Delft, The Netherlands
a r t i c l e i n f o a b s t r a c t
Article history: This contribution puts forward a receding horizon control approach for automated driving
Received 17 March 2015 systems, where tactical-level lane change decisions and control-level accelerations are
Received in revised form 16 June 2015 jointly evaluated under a central mathematical framework. The key idea is that controlled
Accepted 12 July 2015
vehicles predictively determine discrete desired lane sequences and continuous accelera-
Available online 23 July 2015
tions to minimise a cost function reecting undesirable future situations. The interactions
between controlled vehicles and surrounding vehicles are captured in the cost function.
Keywords:
The approach is exible in terms of application to controller design for both
Receding horizon control
Lane-changing
non-cooperative control systems where controlled vehicles only optimise their own cost
Car-following and cooperative control systems where controlled vehicles coordinate their decisions to
Dynamic game theory optimise the collective cost. To determine the controller behaviour, the problem is formu-
Connected vehicles lated as a differential game where controlled vehicles make decisions based on the
expected behaviour of other vehicles. The control decisions are updated at regular fre-
quency, using the newest information regarding the state of controlled vehicles and sur-
rounding vehicles available. A problem decomposition technique is employed to reduce
the dimensionality of the original problem by introducing a nite number of
sub-problems and an iterative algorithm based on Pontryagins Principle is used to solve
sub-problems efciently. The proposed controller performance is demonstrated via numer-
ical examples. The results show that the proposed approach can produce efcient
lane-changing manoeuvres while obeying safety and comfort requirements. Particularly,
the approach generates optimal lane change decisions in the predicted future, including
strategic overtaking, cooperative merging and selecting a safe gap.
2015 Elsevier Ltd. All rights reserved.
1. Introduction
Intelligent vehicle (IV) systems support or automate driving tasks in a safe and efcient way (Varaiya and Shladover,
1991). Different classes are categorised in literature with respect to sensing and control characteristics. Based on the differ-
ences in sensing the driving environment, we distinguish between autonomous vehicles and connected vehicles. Autonomous
vehicles rely solely on on-board sensors based on cyber-physical sensing technologies, such as radar, lidar, machine vision
(VanderWerf et al., 2002; Wang et al., 2014a). Connected vehicles exchange (state and control) information with each other
Corresponding author at: Department of Transport and Planning, Delft University of Technology, Stevinweg 1, 2628 CN, Delft, The Netherlands. Tel.: +31
152784030; fax: +31 152783179.
E-mail addresses: m.wang@tudelft.nl (M. Wang), s.p.hoogendoorn@tudelft.nl (S.P. Hoogendoorn), w.daamen@tudelft.nl (W. Daamen), b.vanarem@
tudelft.nl (B. van Arem), r.happee@tudelft.nl (R. Happee).
http://dx.doi.org/10.1016/j.trc.2015.07.009
0968-090X/ 2015 Elsevier Ltd. All rights reserved.
74 M. Wang et al. / Transportation Research Part C 58 (2015) 7392
via Vehicle-to-Vehicle (V2V) communication or with road infrastructure via Vehicle-to-Infrastructure (V2I) communication
to improve situation awareness (Van Arem et al., 2006; Shladover et al., 2012; Monteil et al., 2013; Ge and Orosz, 2014;
Wang et al., 2014b). With respect to control, the controllers or decision-making systems of IVs can be either
non-cooperative or cooperative. IVs with non-cooperative control strategies make control decisions for their own sake,
i.e. they do not consider the responses of surrounding vehicles to the control actions and there is no negotiation nor consen-
sus in the decision-making process (Wang et al., 2014a). IVs with cooperative control strategies coordinate their behaviour
and take into account the expected response of other vehicles when making decisions (Wang et al., 2014b). Note that with
this denition, an autonomous vehicle employing a cooperative control strategy can exhibit cooperative behaviour without
the need of V2V communication (Wang et al., 2014b).
Considerable efforts have been devoted to IV systems that automate longitudinal driving tasks in both autonomous vehi-
cle systems, e.g. Adaptive Cruise Control (ACC) systems (VanderWerf et al., 2002; Kesting et al., 2008; Wang et al., 2014a),
and connected vehicle systems, e.g. Cooperative Adaptive Cruise Control (CACC) systems (Van Arem et al., 2006; Shladover
et al., 2012; Monteil et al., 2013; Wang et al., 2014b). Simulation and analytical studies show that IVs change macroscopic
trafc operations and the resulting trafc ow characteristics of ACC and CACC systems depend on the control algorithms
and parameter settings. The increase of capacity is mainly a result of shorter time headways compared to human drivers
(Kesting et al., 2008; Shladover et al., 2012), while choosing a larger time headway could have negative impacts on capacity
(VanderWerf et al., 2002). Regarding stability, some authors provide evidence that ACC/CACC systems improve ow stability
(Hasebe et al., 2003; Van Arem et al., 2006), while others are more conservative on the stabilisation effects of ACC systems
(Marsden et al., 2001; Wang et al., 2013).
Automotive studies focus on steering control, given that a desired lane change trajectory based on constant vehicle speed
has been decided by the high-level controller (Falcone et al., 2007; Yoshida et al., 2008). Trajectory planning methods for
autonomous vehicles have also been proposed (Hatipoglu et al., 2003; Xu et al., 2012; Soudbakhsh et al., 2013). But the
expected response of surrounding vehicles to the controlled actions of the subject vehicle is not considered and the current
approaches are not scalable to connected vehicles. Trajectory planning for robots have been extensively studied, with sample
based methods (Luders et al., 2013), potential eld based methods (Shimoda et al., 2007), meta-heuristic based methods
(Hussein et al., 2012). However, vehicles operate at much higher speed than robots and the corresponding safety require-
ments are much higher. Road geometries, lane markings and trafc rules pose many hard constraints on the problem and
lead to many local minima when travelling at centre of each lane. It is difcult to deal with these constraints and
non-convexity with current approaches.
Although many attempts have been made, the issue of when and where it is optimal to change lane remains largely unre-
solved. The discrete lane change events relative to the continuous vehicle positions and the coupled nature between the lat-
eral and longitudinal vehicle dynamics render the problem much more complex compared to the longitudinal control (Kita
et al., 2002; Sarvi et al., 2004; Toledo et al., 2007; Kesting et al., 2007; Liu et al., 2007; Rajamani, 2011; Katzourakis et al.,
2012; Marczak et al., 2013; Zheng, 2014). In addition, the increasing demand for cooperative systems calls for a scalable
approach for designing both autonomous and connected vehicle systems controllers. The aforementioned makes the inte-
grated lateral and longitudinal control of IVs a very challenging topic.
In this article, we generalise previous work on acceleration control (Wang et al., 2014a,b) to a exible mathematical con-
trol approach for fully automated Lane-changing and Car-following Control Systems (LCCS), where discrete and continuous
control variables, i.e. lane change decisions and accelerations, are jointly evaluated. The approach entails that controlled
vehicles make decisions to minimise predicted cost that reects undesirable situations. The interaction between controlled
vehicles and surrounding vehicles is captured in the cost function by including proximity costs. The approach is applied to
controller design for both non-cooperative control systems where controlled vehicles only optimise their own cost and coop-
erative control systems where controlled vehicles coordinate their decisions to optimise the joint cost. To determine the con-
troller behaviour, the problem is formulated as a dynamic game where controlled vehicles make decisions based on the
expected behaviour of other vehicles. A problem decomposition technique is employed to reduce the dimensionality of
the original problem by introducing a nite number of continuous sub-problems. An iterative numerical solution algorithm
based on Pontryagins Principle is used to solve the formulated problem. The proposed control framework is applied to derive
optimal lane change decisions and accelerations for both autonomous and cooperative controllers. The proposed controller
properties and their performance are veried at the microscopic level via numerical examples.
Note that we do not aiming at developing a new driver behaviour model but propose a new predictive approach that gen-
erates lane change sequences and accelerations in the future and is applicable for both autonomous and connected vehicle
systems. The optimal control decisions generated by the approach determine a unique and continuous path that can be used
by the automated vehicle actuators to follow. The approach can anticipate the lane changes in the future and can deal with
the interaction and cooperation of multiple vehicles in conicting situations. The current work focuses on the mathematical
formulation of the complex control problem and derivation of control algorithms, showing the generalisability and applica-
bility of the mathematical framework. In the discussion and conclusion sections, we highlight some of the possible directions
in which the approach could be extended, such as inclusion of anticipation and relaxation in lane changes, verication of the
scalability of the approach to many vehicles and impact assessment of the proposed controller on dynamic trafc ow.
In the remainder, we rst present the necessary assumptions regarding the LCCS controller. Then we formulate the con-
trol problem under a mathematical framework. After that, we decompose the original problem into a nite number of
M. Wang et al. / Transportation Research Part C 58 (2015) 7392 75
sub-problems and derive the optimal control strategies. Numerical examples are shown to verify the performance of the pro-
posed approach, followed by a discussion on some properties of the approach. Concluding remarks are presented in the end.
2. Control design assumptions
Fig. 1 shows an abstract representation of a model predictive LCCS controller (the grey rectangle). Let us consider the sys-
tem from the perspective of an LCCS controlled vehicle i, of which the state z can be described by the positions and speeds of
the controlled vehicle and the surrounding vehicles. At time instant t0 , the estimator receives the positions and speeds of
surrounding vehicles from observations either made by its on-board sensors, i.e. autonomous vehicle systems (Wang
et al., 2014a), or transmitted from other sensors through V2V and/or V2I communications (Wang et al., 2014b), i.e. connected
vehicle systems. Based on this information, the estimator makes an estimate of the current state of the system and passes the
estimate to the controller. The optimisation-based controller uses a system dynamics model to predict the future state in a
time horizon T p , with the current system state as the initial conditions. The optimal control input/variable u, e.g. lane change
sequences and accelerations, is determined to optimise some criterion or cost reecting undesirable situations. The optimal
control variable determines a unique path, and the low-level vehicle steering, propulsion and braking systems are automat-
ically controlled to follow the path. In a receding horizon implementation, only the rst sample of the control input is exe-
cuted by the actuators (Morari and Lee, 1999; Mayne et al., 2000). As the vehicle manoeuvres, the system state changes, and
the optimal control signal u is recalculated with the newest information regarding the system state at regular time intervals,
e.g. Dt.
The control objectives of LCCS include multiple criteria of maximising (travel) efciency, minimising risk of collision, and
maximising comfort, while obeying trafc rules. The importance of each of these objectives may vary according to prefer-
ences of drivers, trafc conditions or vehicle types. The control objectives will be recast into a scalar cost function under
the dynamic optimisation framework described in the next section. Here we introduce the main assumptions regarding
the LCCS design and operations, upon which the mathematical control framework is built:
LCCS scan the driving environment via on-board sensors, e.g. radar or camera, identifying positions and speeds of other
vehicles in the vicinity. The problem of state estimation is not covered in this paper.
Fig. 1. Schematic representation of the LCCS controller.

LCCS predict the future evolution of the system state, including the expected behaviour of surrounding vehicles, using a
system dynamics model. We emphasise that the prediction of behaviours of surrounding vehicles needs not to be perfect.
LCCS make control decisions, i.e. lane choices and accelerations, to full control objectives and the vehicle actuators exe-
cute the control decisions automatically.
We consider fully automated systems with vehicle actuators operating at a much higher accuracy and sampling rate com-
pared to the lane change and acceleration decisions. Thus the vehicles are assumed to be able to track the desired path accu-
rately. Since the main objective is to formulate and solve the integrated lane-changing and car-following control problem
under the mathematical framework and demonstrate workings of the approach, we exclude time delay in the current study.
3. Mathematical control framework
In this section, we describe the mathematical control framework that is used to formulate the control problem for LCCS
controllers.
3.1. State prediction model
For controlled vehicle i, let us dene the continuous state vector z as z xi ; v i ; yi T , where xi and v i denotes the longitu-
dinal position and speed of vehicle i and yi denotes its lateral position. Lateral velocity and heading angle are omitted since
lateral vehicle dynamics can be ignored in non-emergency lane change conditions. The continuous lateral position yi is pre-
dicted and implemented based on representative lane change trajectories from driving experiments (Katzourakis et al.,
2012), cf. Eq. (5). This enables us to obtain feasible and efcient lane change trajectories under linear representation of sys-
tem dynamics.
The control vector u is dened as u ai ; ni ; wi T , where ai denotes the longitudinal acceleration. The vector
ni ni;1 ; ni;2 ; . . . ; ni;n T with ni;k 2 0; T p includes the time instants when the controlled vehicle changes its desired lane ri ,
i.e. the time when initiating a lane change. wi wi;1 ; wi;2 ; . . . ; wi;n T with wi;k 2 f1; 0; 1g denoting the direction of lane change,
i.e. f1; 0; 1g : {change left, no lane change, change right}. Since the discrete lane changes take place at a much lower fre-
quency compared to the innite continuous accelerations, we assume that the pair ni ; wi is evaluated at p Hz, while the
accelerations are evaluated at a frequency 510 times higher than the lane change decisions. The design choice will reduce
the computational complexity of the control problem as we will show in Section 4.1. Note the desired lane sequence ri can
be uniquely determined by the vector pair ni ; wi T with
8
>
>
ri t0 wi;1 if t 0 6 t 6 t 0 ni;1
>
>
>
> X
k
>
< ri t 0 wi;q if t 0 ni;k1 < t 6 t 0 ni;k ; 8k 2 f2; . . . ; ng
ri t q1 1
>
>
>
> Xn
>
>
: ri t 0
> wi;q if t 0 ni;k < t < t 0 T p
q1
and hence the vector u ai ; ni ; wi T is equivalent to ai ; ri T . Eq. (1) implies that the desired lane ri may change (instanta-
neously) at time ni;k .
The controller uses the following kinematic vehicle model to predict its state dynamics:
0 1 0 1
xi vi
d dB C B C
z @ v i A @ ai A fz; u 2
dt dt
yi gni ; wi
with initial condition of zt0 x0 ; v 0 ; y0 T .

The lateral position of the controlled vehicle is determined by the control input pair ni ; wi under heuristic rules. It is
assumed that the controlled vehicle nishes a lane change in a xed time window t lc , which is underpinned by the average
lane change duration from empirical observations (Toledo and Zohar, 2007; Aghabayk et al., 2011) In addition, the controlled
vehicle maintains at least a short interval tinter on the new desired lane after a lane change before initiating a subsequent lane
change. This adds a constraint to the choice of subsequent lane change time instants with:
ni;k1 P ni;k t lc tinter ; 8k 2 f1; 2; . . . ; n 1g 3

To ensure closed-loop stability on lane choices, the controller predicts and plans a complete lane change trajectory within
the prediction horizon, which adds another constraint on the choice of lane change time instant as:
ni;k 6 T p t lc ; 8k 2 f1; 2; . . . ; ng 4
The lateral trajectory is assumed to follow a trigonometric (cosine) function of time (Katzourakis et al., 2012) and its
dynamics can be described by:
(w
i;k w p tt0 ni;k
d 2t lc
sin tlc
p if t 0 ni;k 6 t 6 t 0 ni;k tlc ; 8k 2 f1; 2; . . . ; ng
y gni ; wi 5
dt i 0 if otherwise
where w is the lane width and y0 is the initial lateral position of the vehicle in the middle of the current lane.
The controlled vehicle uses the same model and the decision-making algorithm described in the following section to pre-
dict the behaviour of other vehicles j j i in its vicinity. Note that this prediction needs not to be perfect, i.e. there can be
errors in the predicted behaviour in the open loop and the actual vehicle behaviour in the closed loop. The feedback nature of
the receding horizon framework corrects the mismatch in the prediction (Wang et al., 2014a).
As we will show in the ensuing, this model allows the efcient derivation of optimal lane-changing strategies and
car-following laws that can anticipate the changes when travelling in different lanes in the future.
3.2. Cost optimisation
The optimal control problem for LCCS at time t0 is to seek the optimal control u over a prediction horizon T p that drives
the system along a trajectory such that the predicted cost J is minimised, i.e.:
ut0 ;tf arg minJz; u; tjzt 0 6
u
Z tf
Jzt; ut; tjzt0 Lzt; ut; tdt 7
t0
subject to the state dynamic equation of (2) and the initial condition zt0 ^ z0 , and other constraints on state and control
variables. In Eq. (7), tf t 0 T p denotes the terminal time. L denotes the running cost, describing the costs incurred during
an innitesimal period s; s ds. Other constraints on the state and control vectors include:
Speed constraint: the controlled vehicle drives under a maximum speed of v max .
v i 2 0; v max 8
Acceleration constraint: the accelerations are limited to some admissible bounds.
ai 2 amin ; amax 9
3.3. Cost specication for non-cooperative controller
For the non-cooperative controller, the running cost function for longitudinal movement on a certain lane is specied as:
bsafe 2 2 0 end
Lzt; ut; t Dv 2i;ri HDv i;ri beq v e si;ri v i bctrl a2i beff v d v ari broute expd =di;ri bpref hri
si;ri |{z} |{z} |{z} |{z} |{z}
|{z} equilibrium control travel efficiency route lane preference
safety
bswit Lswit ri 10
|{z}
lane switch
where
si;ri xlri xi l 11
Dv i;ri v lri v i 12
xlri and v lri are the position and speed of the lead vehicle on lane ri . si;ri and Dv lri are the longitudinal gap and relative speed
with respect to the lead vehicle on lane ri . l is the vehicle length.
v d is the free/desired speed and s0 is the minimum gap at standstill conditions. ve denotes the local equilibrium speed
determined by local density, which is a non-decreasing function of gap:
8
< vd if si;ri > sf
v e si;r : s r s
i i; i
0 13
td
if si;ri 6 sf
td denotes the desired time gap and sf v d t d s0 is a gap threshold dening the spatial range of interaction with predecessor.
bsafe ; beff ; beq ; broute ; bpref ; bswit ; bctrl are positive weight factors. v ari is the anticipated speed on lane ri , which is the mini-
mum of predecessor speed v lri and the speed limit v lim
ri on lane ri :

v ar i
min v lim
r ; vr
l
i i
14
The rst three cost terms in Eq. (10) are only related to the longitudinal motion of the controlled vehicles. The running
cost formulation (10) brings the free driving (cruising) and constrained car-following modes into a unied function, as
opposed to the two-regime paradigms in previous studies (VanderWerf et al., 2002; Wang et al., 2013). The last four cost
terms are related to the incentives of lane change decisions. Without lane change cost terms, the controller with only the
rst three cost terms denes an ACC controller that has been tested in representative scenarios (Wang et al., 2014a). The
controller makes some trade-off among those costs:
The safety cost stems from approaching the predecessor, i.e. when Dv i;ri < 0, and it ensures a large penalty when the con-
trolled vehicle approaches its preceding vehicle at small gaps.
The equilibrium cost implies that the vehicle tends to relax to the local equilibrium speed determined by the distance to
the predecessor. The equilibrium and safety costs are the proximity cost terms reecting the interaction with the preced-
ing vehicle.
The control cost is represented by penalising large accelerations or decelerations. It also reects driving comfort to some
extent.
The travel efciency cost stems from deviation of the preceding vehicle speed from the free/desired speed v 0 . Note that
when the preceding vehicle is driving at a speed higher than the desired speed, the efciency cost equals zero.
The route cost stems from driving on a lane that does not lead to the destination, i.e. approaching a dead end to the des-
end
tination. di;ri is the distance between the location of the controlled vehicle to the dead end position xend :
end
di;ri xend xi 15
0
This cost increases exponentially as distance to the dead end decreases to below d . The route cost only occurs when the
end
di;ri is smaller than a certain distance dr , which is around 1 km as indicated by trafc signs or navigation systems.
The lane preference cost can be described as a function of the lane number. We start with a simple formulation that cap-
tures the keep-right directive in most European countries:
hri ntotal ri 16
where ntotal denotes the total lane number. ri increases from left to right with ri 1 for the leftmost lane. Eq. (16) reects
the keep-right directive, since the cost increases when moving from right to left.
The lane switch cost stems from the change in the desired lane. It reects some discomfort caused by lateral accelerations
or the drivers unwillingness to change lane. It also guarantees that the controlled vehicle will not change lane due to
small incentives such as marginal speed gains.
8
< 1 if limri t s limri t s
s!0 s!0
Lswit t 17
:0 if limri t s limri t s
s!0 s!0
3.4. Cooperative control of connected vehicles
The aforementioned basic formulation of the controller pertains to non-cooperative control systems for autonomous
vehicles. It can be easily extended to cooperative control of connected vehicles where controlled vehicles exchange informa-
tion and cooperate with each other using V2V communication. In the remaining of this section, we show how to formulate
the cooperative control problem with two vehicles, though the extension to more vehicles is straightforward (Wang et al.,
2014b). It is assumed that connected vehicles exchange their state and control information and make control decisions
together to minimise a joint cost function representing the collective situations of all vehicles.
3.4.1. Expansion of state and control space

Consider two connected vehicles i and j interacting with each other with the possibility of changing lanes. We consider a
joint state vector of z xi ; v i ; yi ; xj ; v j ; yj T and a joint control vector of u ai ; ni ; wi ; aj ; nj ; wj T . The state and control space are
expanded. The system dynamics model becomes:
0 1 0 1
xi vi
B v i C B ai C
B C B C
B C B C
d dB yi C B gni ; wi C
z B CB C fz; u 18
dt dt B C B
B xj C B v j C
C
B C B C
@ v j A @ aj A
yj gnj ; wj
The desired lane sequence is determined by Eq. (1) for both vehicles and the heuristics lateral position dynamics (5) and con-
straints on the lane change time instants (3) and (4) are the same for both vehicles.
3.4.2. Joint cost function

The optimal control problem for the cooperative controller remains the same as in Eqs. (6) and (7), except that the dimen-
sion of the problem is expanded compared to the non-cooperative controller. The joint running cost function is formulated
as:
0 1
bsafe 2
B sr Dv rq HDv rq b eq v srq v q bctrl aq
2 e 2
|{z} |{z} C
B |{z}
q C
XB B equilibrium control C
C
Lzt; ut; t safety 19
B C
B
qi;jB 2 end C
b v d
v a
b expd =d b hr b L r C
@ eff rq route 0 q;r pref q swit swit q A
} |{zq} |{z} |{z}
|{z
travel efficiency route lane preference lane switch
Eq. (19) shows that the connected vehicles aim at minimising a joint cost function consisting of the safety, equilibrium, con-
trol, efciency, route, lane preference, and lane switch costs of all vehicles.
3.5. Cooperative behaviour of autonomous vehicle
As we have shown in the previous subsection, cooperative control of connected vehicles entails exchanging information
via V2V communication and making decisions together. The prerequisite of inter-vehicle communication limits the applica-
tion of this strategy to scenarios with high penetration rates where the probability of connected vehicles neighbouring each
other is high. However, the cooperative game concept is not restricted to control of connected vehicles. Even an autonomous
vehicle can exhibit cooperative behaviour by taking into account the expected responses of non-controlled vehicles to its
own control decisions (Wang et al., 2014b). In the remainder of this section, we present the problem formulation for an
autonomous vehicle employing a cooperative control strategy.
We assume that the current positions and speeds of neighbouring vehicles are known to i from on-board observations. Let
us denote ui as the set of neighbours whose behaviours are inuenced by the control actions of an autonomous vehicle i.
These vehicles can be the follower of vehicle i on the current lane ri and the potential followers on adjacent lanes ri 1.
The key is to use a driver model M, e.g. lane-changing model and/or car-following model, to predict the (lane change
and acceleration) decisions of surrounding vehicles. For any control decision sequence ai ; ni ; wi T of vehicle i, we can predict
the evolution of lane change and acceleration decisions with the driver model M, and (lateral and longitudinal) positions and
speeds of all neighbouring vehicles in ui with the system dynamics model of Eqs. (1)(5) in the nite horizon T p , using the
future evolution of position and speed of vehicle i as known variables. Hence we can calculate the predicted costs of for all
neighbours of vehicle i and introduce the following cost function for autonomous vehicle under cooperative control strategy:
Xj2ui
Lzt; ut; t Lai ; ni ; wi j ; nj ; w
La 20
j
|{z} |{z}
non-cooperative cooperative
The non-cooperative cost term in Eq. (20) is the running cost function of non-cooperative controlled vehicle, i.e. Eq. (10). The
last cooperative cost term reects the extra cost that the control action of autonomous vehicle i brings to its neighbouring
vehicles.
We remark that the specication for the cooperative cost is not unique and is dependent on the control objectives. In the
numerical example (Experiment 6) we will show in Section 5, the cooperative cost is specied as:
j ; nj ; w bsafe Dv 2 HDv j;r beq v e sj;r v j 2 Hv e sj;r < v j

La 21
j j;rj
sj;rj j j j
The rst term on the right hand side of Eq. (21) gives high penalty when manoeuvres of controlled vehicle i lead to unsafe
situations for the neighbouring vehicle j. The second term penalises manoeuvres of controlled vehicle i that result in a lower
equilibrium speed than the current speed of neighbour j. For the sake of trafc efciency, Eq. (21) is chosen to ensure that
vehicle i does not gets extra cost if its control decisions lead to positive acceleration to its neighbours. As we will show in the
ensuing, this formulation (20) and (21) can guarantee that the optimal lane change manoeuvre of controlled vehicle i does
not result in hard braking of other vehicles, which is similar to human driver behaviour.
4. Derivation of optimal accelerations and lane change sequences
In the previous section, we formulated lateral and longitudinal control problem into an optimal control framework. The
inclusion of the discrete changes in the desired lane sequence in the framework makes it challenging to nd the feasible
solution efciently. This section presents an efcient solution approach to the optimal control problem of LCCS. The main
idea is to decompose the problem into a nite number of sub-problems, each of which is continuous with linear system
dynamics. The sub-problems are differentiated by the time instant of each lane change decision and continuities in longitu-
dinal and lateral positions, speeds and accelerations of controlled vehicles are guaranteed at boundaries of adjacent
sub-problems. This allows us to use the efcient solution approach based on Pontryagins Principle for sub-problems to nd
optimal accelerations at the lower level and to calculate the predicted cost for each lane change strategy. The predicted costs
are then compared to optimise lane change decisions at the higher level. In the remainder of this section, we present details
of the problem decomposition scheme and the optimisation algorithms.
4.1. Problem decomposition
In general there is an innite number of possible ni ; wi pairs of lane change strategies for controlled vehicle i during the
prediction horizon, which makes the problem difcult to solve. The dimensionality of the problem is reduced by differenti-
ating the update frequency of lane change decisions and accelerations. Although the lane change decisions and longitudinal
accelerations are predicted jointly, the lane change decisions take place at a lower frequency compared to accelerations. The
pair ni ; wi is evaluated at p Hz, while the accelerations are evaluated at a frequency 510 times higher than the lane change
n o
decisions. Hence the choice of ni;k 2 ni is limited to a nite set of N 0; 1p ; 2p ; . . . ; mp , with m the largest integer less than or
equal to pT p tlc . The choice of wi;k is constrained to the set of W f1; 0; 1g. With the constraints of (3) and (4), this sig-
nicantly reduces the dimensionality of the considered problem. In addition, for each chosen feasible pair ni ; wi with n lane
change decision points, a lane sequence rt can be determined with Eq. (1). Hence, the original problem of Eqs. (6), (7), (10)
is translated into several equivalent sub-problems of
ai;t0 ;tf arg minJxi ; v i ; ai jzt0 ; ni ; wi 22

ai
Z tf
Jxi ; v i ; ai jzt 0 ; ni ; wi Lxi ; v i ; ai jzt0 ; ni ; wi dt
t0
Z ni;1 Z ni;2
Lxi ; v i ; ai jzt 0 ; ni;1 ; wi;1 dt Lxi ; v i ; ai jzni;1 ; ni;2 ; wi;2 dt
t0 ni;1
Z ni;n Z tf
Lxi ; v i ; ai jzni;n1 ; ni;n ; wi;n dt Lxi ; v i ; ai jzni;n dt 23
ni;n1 ni;n
where z xi ; v i T denotes the sub-state variable and u

ai denote the sub-control variable of the controlled vehicle. The
sub-state dynamic equation becomes:

d d xi vi
z fz; u
24
dt dt vi ai
Eq. (23) implies that the original problem under a certain lane change strategy (ni ; wi ) with n lane change decisions is tran-
scribed into n 1 sub-problems. The sequential sub-problems are coupled via the terminal and initial conditions for
sub-state and sub-control variables, i.e. the terminal conditions of the sub-state and sub-control variables of one
sub-problem is the initial conditions of the sub-state and sub-control variables of the following one. The transcribed
sub-problems are continuous in the sub-state and sub-control variables and hence are nite horizon optimal control prob-
lems with known parameters and exogenous variables, including lane sequence and surrounding vehicles state.
With this problem decomposition technique, we can compute the optimal acceleration for each sub-problem and inte-
grate the running cost function over the complete prediction horizon t 0 ; t f to calculate the predicted cost for each lane
change strategy. After that, we can search the optimal lane change strategy or equivalently desired lane sequence among
a nite number of possibilities that gives the minimum cost. The evaluation frequency p determines the number of possible
lane change trajectories within the prediction horizon from current time and consequently the total number of
sub-problems proportionally. A lower p leads to fewer feasible lane change strategies but may result in sub-optimal solu-
tions to the original problem, while increasing the evaluation frequency of lane change decisions increases the number of
sub-problems nitely.
In the following, we present an efcient numerical solution to the sub-problems.
4.2. Solution approach to sub-problems based on Pontryagins Principle
Without going too much into details, the solution approach based on Pontryagins Principle entails dening the
Hamiltonian H as follows:
Hz; u kT fz; u
; k Lz; u 25
where k denotes the so-called co-state or marginal cost of the state

z, which reects the relative extra cost incurred due to a
small change d
z on the state.
Using the Hamiltonian, we can derive the necessary condition for optimality with u arg minu H ; k, s:t: u
z; u 2 U, where
U denotes the admissible control space. This requirement will enable expressing the optimal control input u as a function of
the state
z and the co-state k. Furthermore, the co-state has to satisfy the following dynamic equation:
d @H @L @f
k kT 26
dt @ z @ z @ z
subject to the terminal conditions of kt f (Pontryagin et al., 1962; Hoogendoorn et al., 2012).
The optimal control problem now is relaxed to solve the set of ordinary differential equations of (2) and (26). An iterative
algorithm has been proposed to solve the two-point boundary value problem efciently (Hoogendoorn et al., 2012).
4.3. Optimal acceleration control law and non-cooperative lane change decision
Applying the aforementioned solution approach based on Pontryagins Principle yields the following optimal acceleration
control law for vehicle i employing non-cooperative control strategy:
kv i
ai 27
2bctrl
where kv i denotes the marginal cost of speed v i . The optimal acceleration shows the direction in which the cost decreases
fastest. The co-state dynamics of the controller satises the following differential equation:
d xi @H broute d0 b 2beq
end safe
2 exp d0 =di;ri 2 Dv i;ri HDv i;ri v e si;ri v i
2
k 28
dt @xi end si;ri td
di;ri
d v i @H 2b
k safe Dv i;ri HDv i;ri 2beq v i v e si;ri kxi 29
dt @v i si;ri
Using the iterative numerical schemes can solve the state dynamic equation and co-state dynamic equation efciently
(Hoogendoorn et al., 2012).
Solving the state and co-state dynamic equations yields the optimal acceleration ai and the state trajectory
z for the
sub-problem. Hence the predicted (optimal) cost for a given ni ; wi pair can be calculated with:
Z tf
Jai tjni ; wi Lz t; ai t; tjzt 0 ; ni ; wi dt 30
t0
and the optimal lane sequence is determined by:

ni ; wi arg min J ai tjzt 0 ; ni ; wi 31
ni;k 2N; wi;k 2W
Eq. (31) entails that the LCCS controller evaluates the predicted costs on all possible lane change scenarios (including
remaining on the same lane) and chooses the optimal lane change decision points or equivalently the optimal desired lane
sequence with the minimum predicted cost. Due to the limited number of lane change possibilities, exhaustive enumeration
is used to nd the optimal lane change strategy that minimises the predicted cost (30).
4.4. Cooperative lane change strategies
Using the same problem decomposition technique as described in Section 4.1, we arrive at the problem reformulation for
the cooperative controller under the lane change strategy of ni ; wi ; nj ; wj as:
ai ; aj t arg min Jxi ; v i ; ai ; xj ; v j ; aj jzt 0 ; ni ; wi ; nj ; wj 32

0 ;t f ai ;aj
Z tf
Jxi ; v i ; ai ; xj ; v j ; aj jzt 0 ; ni ; wi ; nj ; wj Lxi ; v i ; ai ; xj ; v j ; aj jzt 0 ; ni ; wi ; nj ; wj dt 33
t0
z xi ; v i ; xj ; v j T and the sub-control variable of u

where the expanded sub-state variable ai ; aj T satisfy the following
dynamic equation:
0 1 0
xi vi 1
d d B vi C B C
C B ai C
z B B C B C fz; u
34
dt dt @ xj A @ v j A
vj aj
and the same initial conditions and constraints on state and control variables. The transcribed sub-problems are continuous
z xi ; v i ; xj ; v j T and sub-control vector of u
in the sub-state ai ; aj T of the controlled vehicle, and hence are nite horizon
optimal control problems with known parameters and exogenous variables.
The sub-problem can be solved using Pontryagins Principle. Taking the necessary conditions, we arrive at the optimal
acceleration laws for both vehicles as:
kv i kv j
ai and aj 35
2bctrl 2bctrl
Eq. (35) shows the optimal accelerations of the two connected vehicles are negatively proportional to the co-state or mar-
ginal costs of their speeds respectively, and the co-state dynamic equations can be determined in the same way as the
non-cooperative case with Eqs. (26) and (19).
The following mathematical programme shows how to compute the optimal lane change strategy for connected vehicles.
Solving the state and co-state dynamic in each sub-problem gives us the optimal trajectory driven by the optimal acceler-
ations ai ; aj under the corresponding lane change strategy ni ; wi ; nj ; wj . This allows us to calculate the predicted cost under
this lane change strategy with
Z tf
Jai ; aj jzt0 ; ni ; wi ; nj ; wj Lai ; aj jzt0 ; ni ; wi ; nj ; wj dt 36
t0
and the optimal lane change sequences for connected vehicles can be determined by:
ni ; wi ; nj ; wj arg min J ai tjzt 0 ; ni ; wi ; nj ; wj 37
ni;k ;nj;k 2N wi;k ;wj;k 2W
Note that the problem of (36) and (37) forms a dynamic cooperative game, where connected vehicles, also known as players,
coordinate their behaviour and seek the optimal strategy to minimise the joint cost (Basar and Olsder, 1995). The cooperative
formulation also shows the interdependency of lane change decisions and accelerations, since the predicted cost is deter-
mined by both the desired lane sequence and the acceleration trajectory. The desired lane sequence determines whether,
when, and how the controlled vehicles interact with each other, e.g. whether they will travel on the same lane.
For cooperative control of an autonomous vehicle, the derivation of acceleration control law is the same as the
non-cooperative case and the costate dynamics can be determined in the same way as the non-cooperative case with
Eqs. (26) and (20). The only difference is the calculation of the predicted cost J, where an extra cooperative cost of Eq.
(21) should be included when optimising lane change decisions. The extra cooperative cost for neighbouring vehicle j can
be calculated with the driver model M and system dynamics model (1)(5), using control decisions of the autonomous vehi-
cle i as known variables, cf. Section 3.5.
5. Verication of the controller performance
To verify the performance of the proposed approach, we carry out several numerical experiments for representative sce-
narios in this section. The main objective is to show that the proposed approach leads to working algorithms and generates
plausible lane change decisions and accelerations. While acknowledging the chosen scenarios are a subset of reality, they
demonstrate the fundamental workings of the approach and show face validity in the controller behaviour. Verication of
the approach in more scenarios is left for future work.
5.1. Numerical experimental design
In Experiment 1 and 2, we demonstrate the controller performance in making single lane change decisions of one con-
trolled vehicle on a two-lane highway section. There is no follower behind the controlled vehicle in these two scenarios.
In Experiment 1, the leader travels with a constant speed of 25 m/s, the controlled vehicle starts at a location of x1 0 m,
50 m behind the leader and with an initial speed of 25 m/s. The controlled follower has a desired speed of 30 m/s. It is
expected that the controlled vehicle should change lane to gain speed. The scenario is shown in Fig. 2(a).
In Experiment 2, we examine whether the proposed approach can correctly perform lane changes that are preferred in the
future. The scenario is shown in Fig. 2(b). The controlled vehicle starts at the same location of x1 0 m on the right lane, with
the same initial speed of 25 m/s. The leader on the right lane travels with a constant speed of 20 m/s and with an initial dis-
tance of 50 m in front of the controlled vehicle. On the left lane, there is another uncontrolled surrounding vehicle travelling
with a constant speed of 25 m/s but with a longitudinal gap of 5 m in front of the controlled vehicle. It is ideal for the con-
trolled vehicle to travel on the left lane to gain speed, but an immediate lane change will result in a very close distance to the
leader on the left lane. Hence it is expected that a delayed lane change leads to better situation for the controlled vehicle.
In Experiment 3, we verify whether the approach can predict multiple lane change decisions of one controlled vehicle in a
longer prediction horizon. We increase the prediction horizon to 15 s to allow two lane change decisions in the prediction
horizon under the constraints (3) and (4). The scenario is shown in Fig. 2(c). The controlled vehicle starts at the same location
of x1 0 m on the right lane, with an initial speed of 30 m/s. The leader on the right lane travels with a constant speed of
Fig. 2. Experimental scenarios.
20 m/s. No surrounding vehicles prevail on the left lane. The controlled vehicle is expected to change left to reduce the risk of
collision and after it passes the right lane vehicle it should change right under the keep-right trafc rule.
In Experiment 4 and 5, we examine the cooperation between two controlled vehicles under merging scenarios, where
interaction and cooperation between merging vehicles and mainline vehicles are observed in real trafc (Sarvi et al.,
2004; Liu et al., 2007; Marczak et al., 2013). The network is a two-lane highway with an on-ramp. In Experiment 4, the merg-
ing vehicle (vehicle 1) starts at the on-ramp at x1 0 m with an initial speed of 25 m/s, as shown in Fig. 2(d). The end of the
on-ramp is at xend 350 m. The mainline vehicle (vehicle 2) travels on the right lane of the main highway, at the position of
x2 30 m and initial speed of 30 m/s. We examine how different lane change strategies inuence the individual and total
costs of the two controlled vehicles. In Experiment 5, we make the situation more demanding for both controlled vehicles to
explore the benets of cooperation. The scenario is shown in Fig. 2(e). All other conditions are the same as in Experiment 4,
only the initial gap between the two controlled vehicles is reduced to 10 m.
As the last example, we demonstrate how an autonomous vehicle generates cooperative lane-changing behaviour as for-
mulated in Section 3.5 and selects a safe gap in the target lane. The scenario is shown in Fig. 2(f). We consider a two-lane
highway where vehicle 1 and 2 are travelling on lane 1 and lane 2 respectively, both starting 20 m in front of the controlled
vehicle 3. Vehicle 1 and 2 travel with constant speed of 30 m/s and 25 m/s respectively and the initial speed for the con-
trolled vehicle is 25 m/s. The controlled vehicle has a desired speed of 33 m/s and has desires to change to the left lane
for higher speed. However, there is another vehicle 4 approaching from behind on the left lane, with a speed of 30 m/s
and gap of 10 m. Vehicle 4 follows the autonomous vehicle as long as vehicle 3 starts to change lane and cuts in front of
it. Vehicle 4 does not change lane and its longitudinal motion is modelled with the Intelligent Driver Model (IDM)
(Treiber et al., 2000). Vehicle 3 uses the same IDM model to predict the response of vehicle 4 to its decision. Autonomous
vehicle 3 is expected to reject the current gap and wait a larger gap after vehicle 4 passes it when considering extra cost
for vehicle 4 it may bring. We increase the prediction horizon to 15 s to anticipate more lane change possibilities in this
demanding scenario.
The temporal evolution of running costs, speeds, desired lane sequences, and accelerations will be used to examine the
behaviour of the controlled vehicles. The predicted running costs and total cost are used to compare the performance of the
controlled vehicles. The predicted (total) cost is the optimal cost-to-go, i.e. J . The predicted running costs are the integral of
different running cost terms in the running cost function (10) from the current time t0 to the terminal time t f . The default
controller parameters in Table 1 are used throughout the experiments unless stated otherwise.
5.2. Immediate lane change (Experiment 1)
In the rst experiment, we consider a two-lane section where no vehicle is on the left lane. Hence the anticipated speed
on the left lane is 30 m/s while that on the right lane is 25 m/s, which is the speed of the leader. Within the prediction
Table 1
Controller parameter.
Notation Parameter Default value

vd Desired speed 30 m/s
td Desired time gap 1.2 s
s0 Minimum gap at standstill conditions 2m
tlc Lane change duration 5s
tinter Minimum interval between successive lane changes 2s
p Frequency of lane change decisions 1 Hz
T Prediction horizon 8s
p
bsafe ; beq ; beff ; bpref ; bswit ; bctrl Weight matrix (2, 0.02, 0.1, 1, 1, 0.5)
l Vehicle length 4m
0
d Distance parameter of route cost 350 m
v max Maximum vehicle speed 40 m/s
amax Maximum acceleration value 2 m/s2
amin Minimum acceleration value 8 m/s2
Table 2
Predicted running cost terms and total cost under different lane change sequences for Experiments 1 and 2.
n1;1 0; w1;1 1 n1;1 1; w1;1 1 n1;1 2; w1;1 1 n1;1 3; w1;1 1 No LCa

Experiment 1
Safety cost 0.00 0.02 0.18 0.52 1.47
Equilibrium cost 0.83 0.82 0.82 0.83 1.47
Efciency cost 0.00 2.50 5.00 7.50 20.00
Lane preference cost 8.00 7.00 6.00 5.00 0.00
Lane switch cost 0.20 0.20 0.20 0.20 0.00
Control cost 2.02 2.03 2.04 2.00 1.00
Total cost J 11.05 12.58 14.24 16.04 23.95
Experiment 2
Safety cost 0.00 0.58 0.82 1.17 4.27
Equilibrium cost 18.99 11.12 7.98 7.02 2.53
Efciency cost 20.00 27.50 35.00 42.50 80.00
Lane preference cost 8.00 7.00 6.00 5.00 0.00
Lane switch cost 0.20 0.20 0.20 0.20 0.20
Control cost 5.98 5.98 4.71 3.21 1.03
Total cost J 53.17 52.39 54.72 59.12 87.84
The bold values indicate the optimal solutions.

a
No lane change.
horizon of 8 s, the controlled vehicle has options to change lane immediately, to change lane after 1 s, to change lane after 2 s,
to change lane after 3 s, or to remain on the right lane. The predicted running cost terms and total cost of different decisions
are shown in Table 2, while the temporal evolution of running cost L, desired lane sequence, speed and acceleration are
shown in Fig. 3. The abrupt change in the running cost is caused by the discrete change in the desired lane and the corre-
sponding time instant identies the boundary between two subsequent sub-problems.
It is clear from Table 2 that the longer the controlled vehicle stays in the right lane, the higher efciency costs and the
lower lane preference costs it incurs. Changing lane also increases the control cost, but leads to higher speed, as shown in
Fig. 3(c). The optimal lane change strategy is to change lane immediately, reducing the travel efciency cost at the expenses
of lane switch cost and higher lane preference cost. The optimal strategy leads to the lowest cost of 11.05.
5.3. Delayed lane change (Experiment 2)
In the second experiment, the changed initial conditions still provide incentives for the controlled vehicle to change lane
but make it difcult to change lane immediately. The predicted running cost terms and total cost of different decisions are
shown in Table 2, while the temporal evolution of running cost L, desired lane sequence, speed and acceleration are shown
in Fig. 4.
The leader on the right lane has a lower speed compared to that in Experiment 2, hence the controlled vehicle has more
desire to change lane to reduce efciency cost as well as safety cost. However, the initial gap to the leader on the left lane is
only 5 m. This corresponds to an equilibrium speed of 2.5 m/s according to Eq. (13), much lower than the initial speed of
25 m/s. Changing lane immediately will result in a higher equilibrium cost and the controlled vehicle would have to lower
its speed substantially to match the available gap to the leader on the left lane. Table 2 clearly shows that staying longer in
the right lane results in higher safety cost and efciency cost, while changing lane immediately leads to higher equilibrium
cost and lane preference cost. The trade-off among the cost terms points out that the optimal strategy is to wait one second
and then change left rather than to change lane immediately. Note that most utility-based lane change decision models fail
to predict this delayed lane change.
4.5 0.5
Desired lane number

1
3.5
Running cost 3
1.5
2.5
2
2
1.5
1 2.5
0 2 4 6 8 0 2 4 6 8
time (s) time
(a) (b)
29 1.5
no lane change
1,1 = 0, 1,1 = 1
28
1 1,1 = 1, 1,1 = 1
a (m/s2)
1,1 = 2, 1,1 = 1
v (m/s)
27
1,1 = 3, 1,1 = 1
0.5
26
0
25
0 2 4 6 8 0 2 4 6 8
time time
(c) (d)
Fig. 3. Experiment 1: (a) running cost; (b) desired lane number; (c) speed and (d) acceleration.
5.4. Multiple lane change decisions (Experiment 3)
Now we increase the prediction horizon to 15 s so that multiple lane change decisions within the prediction horizon are
possible and test whether the controller generates the expected behaviour in a simple overtaking scenario. The leader on the
right lane travels with a speed of 20 m/s, while the controlled vehicle starts with the desired speed of 30 m/s, 50 m behind
the leader. No vehicle is travelling on the left lane. We evaluate different lane change strategies (staying on the same lane,
only changing left, and changing left rst and then right) at different decision time instants. The resulting cost of different
strategies are summarised in Table 3. The infeasible combinations are situations that violate constraints of (3) and (4), i.e.
either the controller cannot predict a full lane change process within the prediction horizon or a subsequent lane change
occurs before the initial one is completed.
When the controlled vehicle stays in the right lane, it incurs safety cost, equilibrium cost and efciency cost. These terms
demand the vehicle to decelerate to match the speed of the leader on the right lane. The total cost of this strategy is 166.95.
The vehicle path in the x; y plane, desired lane sequence, speed and acceleration proles of the follow-the-leader or
no-lane-change strategy are shown in red lines in Fig. 5.
It is clear from Table 3 that the longer the controlled vehicle stays behind the leader on the right lane, the higher cost it
gets. Changing lane leads to lower cost compared to the keep-the-same-lane strategy. In addition, overtaking the leader
results in lower cost compared to change-to-left-lane strategies. This is mainly due to the higher lane preference cost for
staying in the left lane.
The optimal strategy is to change to the left lane immediately to avoid high safety and equilibrium cost and then change
to the right lane after 7 s, i.e. immediately after satisfying constraint (3), to avoid high lane preference cost. This results in a
cost of 7.4, signicantly lower than the follow-the-leader strategy. The resulting optimal path, desired lane number
sequence, speed and the acceleration are depicted with blue lines in Fig. 5. As depicted in Fig. 5(a), the optimal lane change
strategy results in a much higher speed and longer distance travelled compared to those of the follow-the-leader strategy,
which is a clear benet for trafc efciency.
5.5. Non-cooperative and cooperative merging (Experiments 4 and 5)
In Experiment 4, we consider the interaction and cooperation of two controlled vehicles in a merging scenario at a high-
way on-ramp. The possible strategies and corresponding predicted cost are shown in Table 4.
25 0.5
Desired lane number

20
1
Running cost
15
1.5
10
2
5
0 2.5
0 2 4 6 8 0 2 4 6 8
time (s) time
(a) (b)
25 1
24 0
a (m/s2)
23 1
v (m/s)
no lane change
1,1 = 0, 1,1 = 1
22 2
1,1 = 1, 1,1 = 1
= 2, = 1
21 3 1,1 1,1
= 3, = 1
1,1 1,1
20 4
0 2 4 6 8 0 2 4 6 8
time time
(c) (d)
Fig. 4. Experiment 2: (a) running cost; (b) desired lane number; (c) speed and (d) acceleration.
Table 3
Predicted (total) cost under different lane change sequences for Experiment 3.
b
Experiment 3 n1;2 7; w1;2 1 n1;2 8; w1;2 1 n1;2 9; w1;2 1 n1;2 10; w1;2 1 w1;2 0
n1;1 0; w1;1 1 7.4 8.4 9.4 10.4 15.2
n1;1 1; w1;1 1 Infeasiblec 21.76 22.76 23.76 28.56
n1;1 2; w1;1 1 Infeasible Infeasible 36.55 37.54 42.34
n1;1 3; w1;1 1 Infeasible Infeasible Infeasible 51.45 56.24
n1;1 4; w1;1 1 Infeasible Infeasible Infeasible Infeasible 69.51
w1;1 0 a 166.95

a
No lane change.
b
No secondary lane change.
c
All infeasible combinations are due to violation of constraint (3).
If vehicle 1 and 2 are non-cooperative controllers, it is evident that the best situation for vehicle 1 is to change left imme-
diately. Given this situation, the best for vehicle 2 is to change left immediately. The running cost, speed and acceleration of
the two vehicles in this situation are shown in Fig. 6. In this case, vehicle 2 incurs a lane switch cost and higher lane pref-
erence cost, while vehicle 1 incurs a cost due to equilibrium speed, which is higher than the current speed. Thus vehicle 1
accelerates to match the equilibrium speed. This leads to a cost of 10.92 for vehicle 1 and 16.20 for vehicle 2, resulting in a
total cost of 27.12. This denes a Nash equilibrium solution, i.e. neither vehicle can improve its individual situation
unilaterally.
The best situation for non-cooperative vehicle 2 is keeping the same lane while vehicle 1 stays in the on-ramp. In this case,
vehicle 2 only incurs a cost of 8.03 due to the lane preference cost, but vehicle 1 incurs a much higher cost of 34.76 due to the
equilibrium cost, route cost and travel efciency cost. The total cost of this strategy amounts to 42.79, signicantly higher
than the best situation for vehicle 1. This situation is worse for the collective cost of two vehicles. This is not a Nash equi-
librium situation, since vehicle 1 can improve its situation by changing lane earlier.
0.8
0 1
Desired lane number

1.2
1
1.4
y (m)
2 1.6
1.8
3
2
4 2.2
0 100 200 300 400 500 0 5 10 15
x (m) time (s)
(a) (b)
30 1
28 0
26
a (m/s2)
v (m/s2)
1
24
2
22
20 3 Optimal strategy
No lane change strategy
18 4
0 5 10 15 0 5 10 15
time (s) time (s)
(c) (d)
Fig. 5. Experiment 3: (a) vehicle path in x; y plane, (b) desired lane sequence, (c) speed and (d) acceleration of optimal strategy with n1;1 0; w1;1 1;
n1;2 7; w1;2 1 and no lane changes strategy.
Table 4
Predicted total cost J (cost of vehicle 1, cost of vehicle 2) for different strategies in Experiments 4 and 5.
n2;1 0; w2;1 1 n2;1 1; w2;1 1 n2;1 2; w2;1 1 n2;1 3; w2;1 1 w2;1 0d

Experiment 4
n1;1 0, w1;1 1 27.12a,b (10.92, 16.20) 32.35 (11.00, 21.36) 36.24 (11.03, 25.20) 38.59 (11.04, 27.54) 34.47 (11.04, 23.42)
n1;1 1, w1;1 1 28.73 (12.46, 16.26) 27.73 (12.46, 15.26) 32.07 (12.46, 19.60) 34.74 (12.46, 22.28) 30.92 (12.46, 18.46)
n1;1 2, w1;1 1 30.33 (14.07, 16.26) 29.33 (14.07, 15.26) 28.33 (14.07, 14.26) 32.42 (14.07, 18.35) 29.68 (14.07, 15.61)
n1;1 3, w1;1 1 32.18 (15.92, 16.26) 31.17 (15.92, 15.25) 30.17 (15.92, 14.25) 29.17 (15.92, 13.25) 29.61 (15.92, 13.69)
w1;1 0c 51.00 (34.76, 16.23) 50.00 (34.76, 15.23) 48.99 (34.76, 14.23) 47.99 (34.76, 13.23) 42.79 (34.76, 8.03)
Experiment 5
n1;1 0, w1;1 1 34.36b (18.16, 16.20) 106.63 (18.18, 88.45) 130.03 (18.18, 111.85) 143.14 (18.18, 124.96) 145.62 (18.18, 127.44)
n1;1 1, w1;1 1 35.83 (19.58, 16.25) 34.83 (19.58, 15.25) 83.95 (19.58, 64.36) 97.06 (19.58, 77.48) 99.83 (19.58, 80.24)
n1;1 2, w1;1 1 37.39 (21.14, 16.26) 36.39 (21.14, 15.26) 35.39 (21.14, 14.25) 42.04 (28.84, 13.20) 32.62 (24.62, 8.00)
n1;1 3, w1;1 1 39.07 (22.87, 16.20) 38.07 (22.87, 15.20) 37.07 (22.87, 14.20) 36.07 (22.87, 13.20) 31.64a (23.64, 8.00)
w1;1 0c 54.57 (38.37, 16.20) 53.57 (38.37, 15.20) 52.57 (38.37, 14.20) 51.57 (38.37, 13.20) 46.37 (38.37, 8.00)

a
Collective optimum solution.
b
Nash equilibrium solution.
c
No lane change for vehicle 1.
d
No lane change for vehicle 2.
When both vehicles are cooperative, they seek a coordinated strategy that minimises the joint cost, i.e. collective optimum
solution. Hence, the mainline vehicle will change left immediately to create space for the merging vehicle, although this
implies a higher cost for himself compared to his best situation when vehicle 1 does not change lane. The merging vehicle
will also change lane immediately. In doing so, the joint cost reaches the minimum. Note that in this experiment, the collec-
tive optimum solution coincides with the Nash equilibrium solution.
The gaming facet becomes more pronounced when the situation is more demanding for the merging and mainline vehicle
in Experiment 5. The merging vehicle starts at the on-ramp at the same location, but with a lower initial speed of 20 m/s. The
mainline vehicle (vehicle 2) travels on the right lane of the main highway, at the position of x2 = 10 m and initial speed of
30 m/s. The predicted costs of different lane change strategies of vehicle 1 and 2 are also summarised in Table 4.
4
30
3.5 1
running cost
3 0
a (m/s2)
25
v (m/s)
2.5
1
2 20
2
1.5 Vehicle 1 (Merging vehicle)
Vehicle 2 (Mainline vehicle)
1 15 3
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
time (s) time (s) time (s)
(a) (b) (c)
Fig. 6. Experiment 4: (a) running cost, (b) speed and (c) acceleration of the merging vehicle and the mainline vehicle under coinciding Nash equilibrium and
collective optimum solutions: vehicle 1 changes left immediately and vehicle 2 changes left immediately (n1;1 0; w1;1 1; n2;1 0; w2;1 1).
6
30
1
5
running cost
a (m/s2)
25
v (m/s)
3 1
20
2 2
Vehicle 1 (Merging vehicle)
1 15 3
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
(a) (b) (c)
Fig. 7. Experiment 5: (a) running cost, (b) speed and (c) acceleration of the merging vehicle and the mainline vehicle under Nash equilibrium solution:
vehicle 1 changes left immediately and vehicle 2 changes left immediately (n1;1 0; w1;1 1; n2;1 0; w2;1 1).
If both vehicles are controlled by non-cooperative or autonomous controllers, the best strategy for vehicle 1 is still to
change lane immediately and given this situation the best for vehicle 2 is to change lane immediately. The situation is shown
in Fig. 7. This leads to a cost of 18.61 for vehicle 1. Although the immediate lane change will not lead to route cost for vehicle
1, the lower initial speed leads to higher equilibrium cost for vehicle 1 and hence it has to accelerate harder to match the
equilibrium speed on the main highway compared to Experiment 4. Given this situation, the best for vehicle 2 is to change
lane immediately, resulting in a cost of 16.20 for vehicle 2. However, the total cost for this situation is 34.36, not the opti-
mum from the collective perspective. This solution is a Nash equilibrium, but not the collective optimum.
The best for non-cooperative vehicle 2 is keeping the same lane while vehicle 1 does not change lane or delay its lane
change manoeuvres to 2 s later. In this case, vehicle 2 is able to pass vehicle 1 to be the leader of the pair, resulting in a min-
imum cost of 8.00 for vehicle 2. However, none of them is a Nash equilibrium solution since the non-cooperative vehicle 1
would be better-off by changing lane earlier.
When both vehicles are cooperative, they seek the collective optimum. Vehicle 1 will delay its lane change after 3 s while
vehicle 2 does not change lane. This will allow vehicle 2 to pass vehicle 1 before vehicle 1 merges and gives sufcient time for
the gap between the new leader (vehicle 2) and the new follower (vehicle 1) on the main highway to grow so that the equi-
librium cost of the merging vehicle will not increase substantially due to the lane change. This results in the minimum joint
cost of 31.64, which is lower than that in the Nash equilibrium solution. The situation is shown in Fig. 8.
Experiments 4 and 5 demonstrate the applicability of the proposed approach in computing optimal strategies for con-
nected vehicle systems. The last experiment shows the clear benet of cooperative controller compared to their
non-cooperative counterpart. Unlike other work using game theoretical approaches in modelling human merging behaviour
(Kita et al., 2002; Liu et al., 2007), the cost of different strategies in our approach is based on the predicted path of the con-
trolled vehicle in the future, taking into account the anticipated behaviour of surrounding vehicles. The optimal lane change
strategy does not necessarily occur at the current time, but may happen in the future. In addition, each strategy corresponds
to an optimal solution for the sub-problem and denes a unique and continuous vehicle path that can be sent to the
lower-level vehicle actuators to track.
5.6. Cooperative lane-changing manoeuvre of an autonomous vehicle (Experiment 6)
In the last experiment, we demonstrate that the proposed cooperative control strategy for autonomous vehicles as for-
mulated in Section 3.5 results in human-like cooperative gap acceptance behaviour. The predicted costs for the autonomous
6
30
1
5
running cost
a (m/s2)
25
v (m/s)
4
3 1
20
2 2
Vehicle 1 (Merging vehicle)
1 15 3
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
(a) (b) (c)
Fig. 8. Experiment 5: (a) running cost, (b) speed and (c) acceleration of the merging vehicle and the mainline vehicle under collective optimum solution:
vehicle 1 changes lane after 3 s and vehicle 2 does not change lane (n1;1 3; w1;1 1; w2;1 0).
Table 5
Predicted running cost terms and total cost under different lane change sequences for Experiment 6.
Experiment 6 n3;1 0 n3;1 1 n3;1 2 n3;1 3 n3;1 4 n3;1 5 n3;1 6 n3;1 7 n3;1 8 n3;1 9 n3;1 10 No LCa
Non-cooperative cost 18.83 20.92 24.30 28.20 31.27 32.61 32.61 33.42 34.84 36.66 38.16 42.69
Cooperative cost 37.64 52.68 1.0e+11 7.6e+10 6.9e+10 0 0 0 0 0 0 0
Total cost J 56.47 73.60 1.0e+11 7.6e+10 6.9e+10 32.61 32.60 33.42 34.84 36.66 38.16 42.69

a
No lane change.
vehicle (vehicle 3) under different lane change strategies are shown in Table 5, where the non-cooperative cost is calculated
as the integral of the running cost function (10) and the cooperative cost is calculated as the integral of Eq. (21) in the pre-
diction horizon. It can be seen clearly that when vehicle 3 immediately cuts in front of vehicle 4, it gets lower
non-cooperative cost compared to other strategies. However, changing left immediately leads to a cooperative cost of
37.64, since vehicle 4 is approaching vehicle 3 with a small gap and the cutting in manoeuvre cause sudden change in
gap and consequently equilibrium speed of vehicle 4. When vehicle 3 changes lane after 1 s, the situation is more demanding
for vehicle 4 with a much shorter gap to vehicle 3. Consequently, vehicle 4 has to brake hard to avoid colliding with vehicle 3,
as shown in Fig. 9(a) and (b). This results in a cooperative cost of 52.68. When vehicle 3 changes lane after 2, 3 and 4 s, it
leads to collision with vehicle 4, which can also be reected by the extremely high cooperative cost it brings. When vehicle
3 changes lane after 5 s or later, vehicle 4 will pass vehicle 3 and thus no cooperative cost prevails.
When the autonomous vehicle makes non-cooperative decisions, i.e. when only optimising the non-cooperative cost, it
will change lane immediately. When employing the cooperative control strategy, it will delay the lane change after 6 s until
vehicle 3 passes it and then switches left to follow vehicle 3. Fig. 9 shows the evolution of gap and acceleration of vehicle 3
and vehicle 4 when vehicle 3 changes lane after 1 s and after 6 s respectively. We can clearly see the abrupt changes in gap
and acceleration for vehicle 4 due to the immediate lane change manoeuvre and how the optimal cooperative decision of
vehicle 3 avoids severe braking for vehicle 4 on the target lane. This experiment illustrates the workings of the cooperative
controller for autonomous vehicles and the fundamental differences of the non-cooperative and cooperative control
strategies.
Note that the cooperation during lane changes is one of the features of human driver behaviour and can be reproduced by
state-of-the-art lane-changing models such as MOBIL (Kesting et al., 2007). This example shows that when including coop-
erative cost term for an autonomous vehicle, the control system generates cooperative behaviour and rejects the decisions
that increase collision risks for surrounding vehicles.
6. Discussion
In the previous section, we illustrate the performance of the proposed control framework with numerical examples. This
section discusses some properties of the controller and possible extensions of the proposed approach. In particular, we will
discuss how the parameters inuence the controller performance, stability property, and the potential application of the pro-
posed approach in driver support systems.
6.1. Choice of controller parameters
Needless to say, the parameters inuence the controller performance. Here we highlight some of the controller param-
eters that have intuitive meanings and discuss how the choice of parameters inuences the controller performance.
50 1
acceleration (m/s2)
40
1
gap (m)
30
2
20
3
10 4
0 5
0 5 10 15 0 5 10 15
time time
(a) Gap evoluation with 3,1 = 1 (b) Acceleration evolution with 3,1 = 1
45 2
acceleration (m/s2)
40 1
gap (m)
35 0
30 1
25 2 Autonomous vehicle (Vehicle 3)

Human follower (Vehicle 4)
20 3
0 5 10 15 0 5 10 15
time time
(c) Gap evoluation with =6 (d) Gap evoluation with =6
3,1 3,1
Fig. 9. Experiment 6: (a) gap and (b) acceleration of autonomous vehicle and human follower when autonomous vehicle changes lane at n3;1 1 s; (c) gap
and (d) acceleration of autonomous vehicle and human follower when autonomous vehicle changes lane at n3;1 6 s.
One of the important parameters is the prediction horizon, which relates to the computational time and stability of the
controller (Rawlings and Muske, 1993; Morari and Lee, 1999; Mayne et al., 2000). Although in a nite horizon model pre-
dictive control setting there is no guarantee that the closed-loop behaviour matches the open-loop prediction, theory and
practices have shown that the closed-loop behaviour is relatively insensitive to changes in the prediction horizon over a wide
range of values (Morari and Lee, 1999; Wang et al., 2012). In general, a long prediction horizon is desirable if the prediction
model is accurate, but this adds computational burden and may render the controller impossible for online implementation.
The default prediction horizon is chosen based on a trade-off between the closed-loop performance and computational time
(Wang et al., 2012) and it allows anticipation of a full lane change process after three seconds from the current time accord-
ing to constraint (4). An adaptive strategy is to choose the prediction horizon based on the trafc conditions, i.e. in loose
trafc conditions where the motions of surrounding vehicles are less dynamic, the prediction horizon can be set longer so
that multiple lane change decisions such as overtaking can be predicted, while in dense trafc conditions where the driving
environment is more dynamic, the prediction horizon can be chosen shorter.
The evaluation frequency inuences the number of sub-problems and optimality. We set the frequency to 1 Hz for illus-
tration purpose, since this will yield a computationally tractable problem. Note that this setting will yield a more robust
number of lane change events when evaluating the impact of the proposed system on trafc ow characteristics in micro-
scopic trafc simulation with varying simulation time steps between 0.1 and 0.5 s, as opposed to most microscopic lane
change models (Laval and Leclercq, 2008). Tuning of the evaluation frequency can be carried out by checking the loss of opti-
mality (reected by predicted cost) in typical scenarios.
The xed time window t lc for a lane change process is chosen based on the average lane change duration of human-driven
vehicles from empirical observations (Toledo and Zohar, 2007; Aghabayk et al., 2011). As pointed out in the work of (Toledo
and Zohar, 2007), lane change duration of human-driven vehicles can be inuenced by actual trafc conditions such as den-
sity, the lane change direction of the change, and by positions and speeds of surrounding vehicles. To better mimic human
driver behaviour, an adaptive lane change duration can be employed by the LCCS controller, e.g. the change duration can be
prolonged for low trafc density and shortened in dense trafc and safety critical conditions.
The weight factors of different cost terms affect the controller performance. Tuning of these parameters is recommended
to generate desired performance. In the longitudinal dimension, we have systematically examined the inuence of the
weight factors of safety cost, equilibrium cost and control cost (Wang et al., 2014a) based on a heuristic approach and
the parameters bsafe ; beq ; bctrl are based on these tuning results. In general, increasing bsafe and beq relative to bctrl enhances
the stability properties of the controller (Wang et al., 2014a). beff ; broute ; bpref ; bswit affect the lane change decisions. beff
relates to the lane change incentive of gaining speed, and increasing it implies more opportunistic behaviour of the controller
in searching fast lanes to travel. Together with the parameter of d0 ; broute will inuence when and where to change to a lane
that leads to the destination. bpref denes a set of lanes which are more attractive to the controlled vehicle and decreasing it
will lead to more balanced distribution of trafc across lanes. bswit sets a threshold that the gain (cost reduction) due to
change lane should be greater than. Increasing the threshold will discourage lane change manoeuvres.
Note that the default parameters in Table 1 are for passenger cars. For trucks, the longitudinal and lateral dynamics are
different compared to cars and a different parameter setting should be used based on the characteristics of truck dynamics.
Nevertheless, the control concept and the proposed approach can be applied to multiple vehicle classes.
6.2. Controller stability
Amongst other controller properties, closed-loop stability is a basic requirement for control design. In previous work, we
have shown that the basic formulation (10) guarantees closed-loop stability or local stability in car-following in an innite
horizon setting (Wang et al., 2013) analytically and by simulation we have shown that the proposed approach in nite hori-
zon setting guarantees stability under step disturbance in representative car-following situations (Wang et al., 2014a).
String stability is of interest when examining the impact of the proposed controller on trafc ow characteristics.
Analytical study and simulation both show that the string stability depends on parameter settings. Increasing desired time
gap t d , weight factors of bsafe and beq relative to bctrl enhances string stability (Wang et al., 2013, 2014a).
6.3. Scalability of the approach to many vehicle scenarios
It is worth mentioning that there are four vehicles at most for the numerical examples we showed in the previous section.
In reality, it is possible to have more interacting vehicles during lane changes. This will increase the dimensionality of the
control problems and the computational load thereof. The numerical solution algorithm has been shown to increase linearly
with the increase of cardinality of the state and control variables (Wang, 2014). It is promising to use the current numerical
solution algorithm for controller design of autonomous vehicles in high density scenarios. For cooperative control of a large
group of connected vehicles, centralised communication and optimisation is infeasible in practice. Distributed algorithms
need be designed for real time implementation (Wang, 2014). Although relevant, the extension of the algorithm and system-
atic investigation for scalability of the algorithm at high density scenarios are beyond the scope of the article and are left for
future research.
6.4. Application in driver support systems
It is possible that the level of automation for intelligent vehicles increases gradually from manual driving to fully auto-
mated driving. Although this paper focus on illustrating the workings of the methodology in fully automated driving sys-
tems, it generates a unique and continuous lane-changing path, which can be used as guidance for drivers. The guidance
can be provided to drivers via visual or auditory interface or via haptic feedback. Although interesting, this is beyond the
scope of the presented study.
7. Conclusions
In this article, we have put forward a predictive approach for unied lane-changing and car-following control based on
receding horizon optimal control and dynamic game theory, where discrete lane change decisions and continuous acceler-
ations are evaluated jointly. We demonstrated the workings of the approach with numerical examples.
The proposed approach generates predicted lane change sequences and accelerations and is applicable for both autono-
mous and connected vehicle systems. The resulting lane change decisions and car-following control inputs determine a
unique and continuous path that can be used by the automated vehicle actuators to track. Numerical examples show that
the controller produces plausible lane-changing and car-following manoeuvres at the microscopic level in highway condi-
tions and cooperative controller can lead to lower joint cost in conicting situations such as merging scenarios.
Particularly, we show that the cooperative control strategy is not restricted to interactions of connected vehicles. By predict-
ing the response of human-driven vehicles with human driver models, an autonomous vehicle based on information from
on-board sensors can exhibit cooperative lane-changing behaviour without the need of inter-vehicle communication.
The presented work focuses on the mathematical framework and illustration of the fundamental workings of the pro-
posed approach. Extensions of the proposed framework in including time delay, adaptive controller parameters, anticipation
of trafc conditions on target lanes and relaxation after lane changes require scrutinised tests on the controller performance
and are interesting topics for future research. We assumed that the lower level controller and actuators can track the desired
trajectory determined by the proposed higher level controller exactly. The follow-up study will deal with mismatch between
the perfect assumption in the higher level controller design and the imperfect tracking capability of the lower level system.
Apart from that, systematic examination of the impact of the proposed controllers on dynamic trafc ow features is impor-
tant to understand future trafc ow characteristics and remains as a future study. Last but not least, in the eld of accident
prevention, automated emergency braking has recently been introduced and automated evasive systems are being devel-
oped. The current framework also has potential in generating optimal braking and/or evasive control strategies when taking
into account vehicle dynamics limitation.
Acknowledgements
The research presented in this paper is partially supported by the Dutch Technology Foundation STW under the project
Truck Merging Support a Step towards Autonomous Driving and partially funded by Shell under the project
Sustainability Perspectives of Cooperative Systems.
References
Aghabayk, K., Moridpour, S., Young, W., Sarvi, M., Wang, Y., 2011. Comparing heavy vehicle and passenger car lane-changing maneuvers on arterial roads
and freeways. Transp. Res. Rec.: J. Transp. Res. Board 2260, 94101.
Basar, T., Olsder, G.J., 1995. Dynamic Noncooperative Game Theory, vol. 200. Academic Press, London.
Falcone, P., Borrelli, F., Asgari, J., Tseng, H.E., Hrovat, D., 2007. Predictive active steering control for autonomous vehicle systems. IEEE Trans. Control Syst.
Technol. 15 (3), 566580.
Ge, J.I., Orosz, G., 2014. Dynamics of connected vehicle systems with delayed acceleration feedback. Transp. Res. Part C: Emerg. Technol. 46, 4664.
Hatipoglu, C., Ozguner, U., Redmill, K., 2003. Automated lane change controller design. IEEE Trans. Intell. Transp. Syst. 4, 1322.
Hasebe, K., Nakayama, A., Sugiyama, Y., 2003. Dynamical model of a cooperative driving system for freeway trafc. Phys. Rev. E 68 (2), 026102/1026102/6.
Hoogendoorn, S.P., Hoogendoorn, R., Wang, M., Daamen, W., 2012. Modeling driver, driver support, and cooperative systems with dynamic optimal control.
Transp. Res. Rec.: J. Transp. Res. Board 2316, 2030.
Hussein, A., Mostafa, H., Badrel-Din, M., Sultan, O., Khamis, A., 2012. Meta-heuristic optimization approach to mobile robot path planning. In: Proceedings of
the International Conference on Engineering and Technology (ICET).
Katzourakis, D., de Winter, J.C., de Groot, S., Happee, R., 2012. Driving simulator parameterization using double-lane change steering metrics as recorded on
ve modern cars. Simul. Model. Pract. Theory 26, 96112.
Kesting, A., Treiber, M., Schonhof, M., Helbing, D., 2008. Adaptive cruise control design for active congestion avoidance. Transp. Res. Part C: Emerg. Technol.
16 (6), 668683.
Kesting, A., Treiber, M., Helbing, D., 2007. General lane-changing model MOBIL for car-following models. Transp. Res. Rec.: J. Transp. Res. Board 1999, 8694.
Kita, H., Tanimoto, K., Fukuyama, K., 2002. A game theoretical analysis of merging-giveway interaction: a joint estimation model. In: Transportation and
Trafc Theory in the 21st Century, Adelaide, Australia. Springer Science, pp. 503518.
Laval, J.A., Leclercq, L., 2008. Microscopic modeling of the relaxation phenomenon using a macroscopic lane-changing model. Transp. Res. Part B: Methodol.
42 (6), 511522.
Liu, H.X., Wu, X., Adam, Z., Ban, J., 2007. A game theoretical approach for modelling merging and yielding behaviour at freeway on-ramp sections. In: 17th
International Symposium on Transportation and Trafc Theory, London, UK. pp. 197211.
Luders, B.D., Karaman, S., How, J.P., 2013. Robust sampling-based motion planning with asymptotic optimality guarantees. In: Proceedings of the AIAA
Guidance, Navigation, and Control Conference (GNC).
Marczak, F., Daamen, W., Buisson, C., 2013. Merging behaviour: empirical comparison between two sites and new theory development. Transp. Res. Part C:
Emerg. Technol. 36 (0), 530546.
Marsden, G., McDonald, M., Brackstone, M., 2001. Towards an understanding of adaptive cruise control. Transp. Res. Part C: Emerg. Technol. 9 (1), 3351.
Mayne, D.Q., Rawlings, J.B., Rao, C.V., Scokaert, P.O.M., 2000. Constrained model predictive control: stability and optimality. Automatica 36 (6), 789814.
Monteil, J., Billot, R., Sau, J., Armetta, F., Hassas, S., El Faouzi, N.-E., 2013. Cooperative highway trafc. Transp. Res. Rec.: J. Transp. Res. Board 2391, 110.
Morari, M., Lee, J.H., 1999. Model predictive control: past, present and future. Comput. Chem. Eng. 23 (45), 667682.
Pontryagin, L.S., Boltyanskii, V.G., Gamkrelidze, R.V., Mishchenko, E.F., 1962. The Mathematical Theory of Optimal Processes. CRC Press.
Rajamani, R., 2011. Vehicle Dynamics and Control. Springer.
Rawlings, J.B., Muske, K.R., 1993. The stability of constrained receding horizon control. IEEE Trans. Autom. Control 38 (10), 15121516.
Sarvi, M., Kuwahara, M., Ceder, A., 2004. Freeway ramp merging phenomena in congested trafc using simulation combined with a driving simulator.
Comput.-Aided Civ. Infrastruct. Eng. 19 (5), 351363.
Shimoda, S., Kuroda, Y., Iagnemma, K., 2007. High-speed navigation of unmanned ground vehicles on uneven terrain using potential elds. Robotica 25,
409424.
Shladover, S.E., Su, D., Lu, X.-Y., 2012. Impacts of cooperative adaptive cruise control on freeway trafc ow. Transp. Res. Rec.: J. Transp. Res. Board 2324, 63
70.
Soudbakhsh, D., Eskandarian, A., Chichka, D., 2013. Vehicle collision avoidance maneuvers with limited lateral acceleration using optimal trajectory control.
J. Dyn. Syst. Meas. Contr. 135, 041006.
Toledo, T., Zohar, D., 2007. Modeling duration of lane changes. Transp. Res. Rec.: J. Transp. Res. Board 1999, 7178. http://dx.doi.org/10.3141/1999-08.
Toledo, T., Koutsopoulos, H.N., Ben-Akiva, M., 2007. Integrated driving behavior modeling. Transp. Res. Part C: Emerg. Technol. 15 (2), 96112.
Treiber, M., Hennecke, A., Helbing, D., 2000. Congested trafc states in empirical observations and microscopic simulations. Phys. Rev. E 62 (2), 18051824.
Van Arem, B., van Driel, C.J.G., Visser, R., 2006. The impact of cooperative adaptive cruise control on trafc-ow characteristics. IEEE Trans. Intell. Transp.
Syst. 7 (4), 429436.
VanderWerf, J., Shladover, S.E., Miller, M., Kourjanskaia, N., 2002. Effects of adaptive cruise control systems on highway trafc ow capacity. Transp. Res.
Rec.: J. Transp. Res. Board 1800, 7884.
Varaiya, P., Shladover, S.E., 1991. Sketch of an IVHS systems architecture. In: Vehicle Navigation and Information Systems Conference, vol. 2, pp. 909922.
Wang, M., 2014. Generic Model Predictive Control Framework for Advanced Driver Assistance Systems. Ph.D. Thesis. Delft University of Technology, Delft,
The Netherlands.
Wang, M., Daamen, W., Hoogendoorn, S.P., van Arem, B., 2012. Driver assistance systems modeling by model predictive control. In: 15th IEEE Conference on
Intelligent Transportation Systems, pp. 15431548.
Wang, M., Daamen, W., Hoogendoorn, S.P., van Arem, B., 2014a. Rolling horizon control framework for driver assistance systems. Part I: mathematical
formulation and non-cooperative systems. Transp. Res. Part C: Emerg. Technol. 40 (0), 271289.
Wang, M., Daamen, W., Hoogendoorn, S.P., van Arem, B., 2014b. Rolling horizon control framework for driver assistance systems. Part II: cooperative sensing
and cooperative control. Transp. Res. Part C: Emerg. Technol. 40 (0), 290311.
Wang, M., Treiber, M., Daamen, W., Hoogendoorn, S.P., van Arem, B., 2013. Modelling supported driving as an optimal control cycle: framework and model
characteristics. Transp. Res. Part C: Emerg. Technol. 36 (0), 547563.
Xu, G., Liu, L., Ou, Y., Song, Z., 2012. Dynamic modeling of driver control strategy of lane-change behavior and trajectory planning for collision prediction.
IEEE Trans. Intell. Transp. Syst. 13, 11381155.
Yoshida, H., Shinohara, S., Nagai, M., 2008. Lane change steering manoeuvre using model predictive control theory. Vehicle Syst. Dyn. 46, 669681.
Zheng, Z., 2014. Recent developments and research needs in modeling lane changing. Transp. Res. Part B: Methodol. 60, 1632.

Game theoretic approach for predictive lane-changing and car-following control

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Game theoretic approach for predictive lane-changing and car-following control

Uploaded by

Copyright:

Available Formats

Transportation Research Part C 58 (2015) 7392

Contents lists available at ScienceDirect

Transportation Research Part C

Game theoretic approach for predictive lane-changing

2. Control design assumptions

Fig. 1. Schematic representation of the LCCS controller.

3. Mathematical control framework

3.1. State prediction model

with initial condition of zt0 x0 ; v 0 ; y0 T .

ni;k1 P ni;k t lc tinter ; 8k 2 f1; 2; . . . ; n 1g 3

3.2. Cost optimisation

3.3. Cost specication for non-cooperative controller

3.4. Cooperative control of connected vehicles

3.4.1. Expansion of state and control space

3.4.2. Joint cost function

3.5. Cooperative behaviour of autonomous vehicle

j ; nj ; w bsafe Dv 2 HDv j;r beq v e sj;r v j 2 Hv e sj;r < v j

4. Derivation of optimal accelerations and lane change sequences

4.1. Problem decomposition

ai;t0 ;tf arg minJxi ; v i ; ai jzt0 ; ni ; wi 22

where z xi ; v i T denotes the sub-state variable and u

4.2. Solution approach to sub-problems based on Pontryagins Principle

where k denotes the so-called co-state or marginal cost of the state

and the optimal lane sequence is determined by:

4.4. Cooperative lane change strategies

ai ; aj t arg min Jxi ; v i ; ai ; xj ; v j ; aj jzt 0 ; ni ; wi ; nj ; wj 32

z xi ; v i ; xj ; v j T and the sub-control variable of u

5. Verication of the controller performance

5.1. Numerical experimental design

Fig. 2. Experimental scenarios.

5.2. Immediate lane change (Experiment 1)

Notation Parameter Default value

n1;1 0; w1;1 1 n1;1 1; w1;1 1 n1;1 2; w1;1 1 n1;1 3; w1;1 1 No LCa

The bold values indicate the optimal solutions.

5.3. Delayed lane change (Experiment 2)

Desired lane number

5.4. Multiple lane change decisions (Experiment 3)

5.5. Non-cooperative and cooperative merging (Experiments 4 and 5)

Desired lane number

The bold values indicate the optimal solutions.

Desired lane number

n2;1 0; w2;1 1 n2;1 1; w2;1 1 n2;1 2; w2;1 1 n2;1 3; w2;1 1 w2;1 0d

The bold values indicate the optimal solutions.

5.6. Cooperative lane-changing manoeuvre of an autonomous vehicle (Experiment 6)

The bold values indicate the optimal solutions.

6.1. Choice of controller parameters

25 2 Autonomous vehicle (Vehicle 3)

6.2. Controller stability

6.3. Scalability of the approach to many vehicle scenarios

6.4. Application in driver support systems

You might also like

ai;t0 ;tf arg minJxi ; v i ; ai jzt0 ; ni ; wi 22

ai ; aj t arg min Jxi ; v i ; ai ; xj ; v j ; aj jzt 0 ; ni ; wi ; nj ; wj 32