Controllability of transportation networks

Transportation Research Part B 118 (2018) 381–406

Contents lists available at ScienceDirect

Transportation Research Part B journal homepage: www.elsevier.com/locate/trb

Controllability of transportation networks Marco Rinaldi University of Luxembourg, Faculty of Science, Technology and Communication. 6, Av. de la Fonte, L-4364 Esch-sur-Alzette (Luxembourg)

a r t i c l e

i n f o

Article history: Received 5 March 2018 Revised 9 October 2018 Accepted 10 November 2018

Keywords: Network-wide traffic control Controllability Optimization

a b s t r a c t Network-wide traffic control policies determine the optimal values for the different kinds of controllers equipped on a transportation network, with the objective of reducing delays and congestion, improving safety and reaching a target Level of Service. While models and algorithms for these problems have been extensively studied in literature, little attention has been devoted to investigating whether/how different locations, kinds (pricing controllers, traffic lights, …) and amounts of controllers in a network affect the overall performance of network-wide schemes. In this work, we adapt the control-theoretical approach of controllability of complex networks to the specific instance of transportation networks, considering both propagation/spillback dynamics and users’ behavior in terms of route choice. Thanks to the newly developed methodology, we then provide exact solutions to the Full Controllability Pricing Controller Location Problem for transportation networks. Comparing different pricing controller location policies through two artificial test cases, we empirically demonstrate how indeed the amount and kind of controllers in a network strongly affect the level of performance reachable by network-wide control policies, specifically in terms of Total Cost minimization. © 2018 Elsevier Ltd. All rights reserved.

1. Introduction Transportation networks require control. Ever since the advent of personal motorized transport in the early 1900 s, engineers and practitioners quickly realized that, as a system, the characteristic self-organisation of transport exhibits substandard performances in key areas. Equipping a high demand intersection with traffic lights is, for example, an effective approach to improve safety, by imposing constraints to the system’s natural behavior. Similarly, when facing congestion, external intervention in the form of control can be exploited to trigger a different desired behavior (better routing choices, controlled access to motorways, …), based on a different set of performance indicators (e.g. travel times). Similarly to what had been done previously to achieve a desired level of safety, control actions can be designed to mitigate the occurrence of congestion in transportation networks: these control actions can take the form of additional costs to the travellers, e.g. pricing, which can effectively be used to disincentivize the use of a specific route, or to reduce demand entering a specific area of a network (cordon pricing). Access to motorways can similarly be regulated by installing specific traffic lights, referred to as ramp meters, at on-ramps upstream of critical areas. It is important to point out, however, that the two objectives (safety on one hand, travel time reduction on the other) are strongly dissimilar from the point of view of control design: while safety concerns arise from highly localized dynamics, microscopic in scale, travel times are the result of emergent dynamics. At a local level, congestion is the result of transporta-

E-mail address: [email protected] https://doi.org/10.1016/j.trb.2018.11.005 0191-2615/© 2018 Elsevier Ltd. All rights reserved.

382

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

tion demand exceeding available supply, which causes progressive deterioration of the supply’s level of service (slowdowns, queueing). Excess demand itself is, though, a product of collective user behavior: departure time choice, route choice, mode choice are all determinants that might cause or exacerbate congestion in different portions of a given transportation network (Sheffi, 1985). Because of this, interventions aimed at locally reducing congestion might result in unexpected, aberrant dynamics: ramp metering effectively reduces the inflow on a busy motorway, but the spillback caused by reducing the outflow from the onramp might result in localized congestion in the underlying urban network, or in re-routing of vehicles towards the next downstream on-ramp, voiding any intended benefit (if not causing a net loss in network performance, (M. Rinaldi et al., 2016)). To address the latter issue, researchers in the past decades have been developing advanced control policies that i) are based on coordination of individual control points and ii) explicitly take into account the network-wide effects of control actions, thanks to predictive traffic propagation and route choice modelling. In works such as those of (Aboudolas et al., 2010; Geroliminis and Daganzo, 2008; M. Rinaldi et al., 2016; Rinaldi and Tampère, 2015; Taale, 2008; Taale and Hoogendoorn, 2012; Van De Weg et al., 2016) network-wide approaches combined with model-based control strategies have been object of considerable research effort. Effectiveness of the proposed solutions has been shown mainly on toy or artificial networks. Extending these results to large scale scenarios has however encountered considerable hardships, due to the fact that the higher the degree of behavior captured by the traffic model of choice, the more complex the resulting optimization problems (Rinaldi et al., 2017). A notable exception, where field testing was implemented, is that of (Hoogendoorn et al., 2014; Hoogendoorn et al., 2016), which evaluated the impact of (non model-based) coordinated control policies in and around the ring road of the city of Amsterdam, in The Netherlands. While the body of literature concerning traffic control strategies evolved considerably over the years, little attention has been devoted to strategic network control design: determining the location, kind and amount of controllers contributing to the aforementioned coordinated strategies is largely an ad-hoc process, in which either existing infrastructure is simply adapted and connected, or expert knowledge is applied in a case-by-case fashion, disregarding the role that control infrastructure plays in network-wide performances (Cantarella and Sforza, 1995). This can be considerably detrimental, as we detail later: the maximum extent to which any control strategy can steer traffic conditions towards more optimal alternatives is operationally bound to only those flow distributions that can be triggered by the existing Control Support Infrastructure (CSI). From a network-wide perspective, thus, the CSI should be optimally designed such that: - safety aspects are locally considered at the intersection level; - economic aspects, such as installation cost and maintenance are taken into account; - quantity, location and kind of controllers installed are sufficient to trigger a desired flow distribution. A few works in the field transportation economics have partly dealt with the latter objective, considering the effect of pricing controllers (de Palma et al., 2005; Evers and Proost, 2015; Verhoef et al., 1996; Verhoef, 2002a), although no conclusive methodology has been developed to solve the highly complex problem for generic transportation networks. In this work, our aim is to lay solid theoretical foundations for the problem of evaluating and designing a set of controllers on a network, both in terms of location and kind of controllers, such that the resulting Control Support Infrastructure is capable of reaching the desired levels of network-wide performance. To achieve said goal, we exploit the control theoretical concept of controllability. Dynamic systems are deemed “controllable” or “fully controllable” if, through a given control action, they can be steered from their current state, whichever it may be, towards any other state. In robotics applications, for example, a robotic arm will be considered fully controllable if its controller set (i.e. its joints) is sufficient in amount and complexity to enable its “hand” to reach any and all points in its range. In transportation networks, indeed, first-best pricing schemes are an instance of full controllability, where pricing on each and every link composing the network represents the CSI tools. Methodologies for determining the degree of controllability for dynamic systems have been historically introduced by (Kalman et al., 1963), and, over the years, this concept has been extended to more complex systems (Hahn et al., 2003). Controllability of systems whose dynamics can be described in terms of networks has recently received increased attention by researchers (Liu et al., 2011; Olshevsky, 2014; Summers et al., 2014; Yuan et al., 2013; Zhao et al., 2017). These works deal with many naturally or technologically occurring networked systems, such as power grids, social networks, electronic circuits etc. and assume interactions between different states to be limited at common, adjacent nodes. As introduced earlier, some behavioral aspects of transportation networks (demand distribution, routing effects due to congestion) violate said hypothesis, influencing topologically distant portions of the network. A key contribution of this work is therefore that of adapting these control theoretical concepts to the specific instance of transportation networks, while retaining the general shape and structure proper of classical dynamic system controllability. This in turn allows us to exploit existing solution approaches to the problem of complex network controllability, and assess whether the resulting CSI design is beneficial in terms of network-wide performance, as measured in terms of Total Cost. The validation is performed numerically, employing the latest state-of-the-art optimization formulations and strategies, on randomized instances of a non-trivial toy network. As will be shown in later Sections, our proposed approach is indeed capable of reaching significant levels of performance, close to System Optimal scenarios. Sensitivity to design parameters and assumptions, specifically in terms of available route set information, is also empirically investigated. The rest of this paper is structured as follows: in Section 2 a state-of-the-art review is presented, comprising of works both from the point of view of transportation engineering and control theory. In Section 3 the methodological contribution

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

383

of this work is detailed. In Section 4 the experimental setup employed for validation is discussed, in terms of simulation environment, selected metrics and scenarios. In Section 5 two test cases are presented: the first test case elaborates the formulation introduced in Section 3 on a small, analytically tractable network, while the second test case deals with the numerical validation of the proposed approach, its performances in terms of network control performance and its sensitivity to a-priori information. Finally, in Section 6 concluding remarks and pointers for future research are presented. 2. Literature review In transportation, traffic control literature can usually be sourced to the work of (Webster, 1958), who first introduced a methodology for devising optimal green splits for isolated intersections, based on their topological characteristics and the vehicular flows. Considering motorway control, the first ramp metering strategy was introduced by (Papageorgiou et al., 1991). Both these seminal works deal with isolated scenarios, and can be considered instances of local strategies. Coordinated traffic control strategies, aligning the configurations of multiple controllers in a given network, were introduced in the early ‘80 s (Little et al., 1981; Wallace et al., 1984), and gradually evolved from fixed time area based strategies towards online traffic responsiveness, as in the works of (Hunt et al., 1981; Lowrie, 1990; Mauro and Taranto, 1990). A further refinement of control strategies was marked by the introduction of explicit traffic modelling in coordinated strategies, as in the works of (Gartner, 1983; Henry et al., 1984; Sen and Head, 1997), yielding the first predictive control strategies applied in transportation. This trend has continued in the last decades, as computationally efficient numerical schemes arose for tackling the complexity of traffic models, yielding Model Predictive Control based approaches (Aboudolas et al., 2009; Dotoli et al., 2006; Hegyi et al., 2005). The inclusion of explicit modelling of user behavior, in terms of route choice, in the problem of optimal control of transportation networks was initially considered in the work of (Allsop, 1974), and subsequently researched extensively both in the time-static (Ghali and Smith, 1995; Yan and Lam, 1996; Yang and Bell, 1997; Yang and Yagar, 1995) and, most recently, time-dynamic modelling domains (Rinaldi et al., 2013; Rinaldi et al., 2016; Taale, 2008; Taale and Hoogendoorn, 2012; Ukkusuri et al., 2013). As the level refinement of the traffic models employed in predictive control strategies increases, a similar trend in the resulting complexity of the related optimization problem can also be observed. This became especially evident when we sought to tackle said complexity through advanced decomposition strategies (M. Rinaldi et al., 2016; Rinaldi and Tampère, 2015). Indeed, the inclusion of route choice in the prediction model introduced severe non-convexity and non-smoothness even when dealing with oversimplified toy networks and static User Equilibrium routing. In turn, this resulted in a high degree of locality of optimization solutions in both static and dynamic approaches, marked with a very strong dependence on the initial guess for the control signals. While investigating the complexity and regularity of traffic control optimization solution spaces, in (Rinaldi et al., 2017) we established a direct cause-and-effect relationship between route choice and solution space complexity, pinpointing how route activation-deactivation dynamics lie at the root of major discontinuities, and developed an optimization scheme that, explicitly taking these effects into account, exhibits improved convergence properties. When performing validation of the latter optimization scheme, we compared the impact of different sets of controllers on a small toy network. These preliminary results hinted at the fact that the amount and location of controllers equipped on a network strongly influence the extent to which optimization approaches can converge to a system optimal (SO) solution, in terms of Total Cost objective function performance. This example, which is later elaborated in Section 5, is at the base of this work’s research. Placement of controllers in the field of transportation has been dealt with mainly from two perspectives: on the one hand, traffic engineering manuals such as the Highway Capacity Manual (Transportation Research Board, 2010) and the German HBS (FGSV-Verlag, 2015) detail how, at a microscopic level, control support infrastructure can be installed on a given intersection, including design pointers for stage composition, cycle lengths, optimal lane widths etc. On the other hand, a few works in transportation economy have dealt with determining optimal network-wide control in terms of pricing. Firstbest pricing (Verhoef et al., 1996) represents in fact a corner solution of control placement, considering the effect of placing a pricing controller on each and every link of a given transportation network. It has been proven that first-best pricing is, by definition, capable of reaching System Optimal performances for general networks in static User Equilibrium. A more interesting problem is that of determining where to install controllers when facing budgetary constraints (as, indeed, control support infrastructure can be considerably expensive). The so dubbed second-best pricing approach has been researched in a few key works (de Palma et al., 2005; Shepherd and Sumalee, 20 04; Verhoef, 20 02a, 20 02b), where algorithms for computing the optimal placement and level of pricing under static User Equilibrium have been introduced. The transportation economical approaches listed above are indeed instances of full and partial controllability solutions to the problem of CSI design. As mentioned earlier, in control theory, the practice of selecting location, amount and kind of controllers such to achieve complete control over a given dynamical system is dubbed as “controllability” . The first work dealing with the concept of controllability for general dynamic systems (more specifically, linear time-invariant and linear time-variant systems) is the seminal paper of (Kalman et al., 1963), where the concept is introduced, and a clear connection between the state space representation of linear dynamical systems and the degree of controllability proper of a given set of controllers is drawn. Alongside defining controllability itself, researchers have also investigated algorithms and approaches for designing control systems such that the total amount or cost of equipping controllers is minimized (Olshevsky, 2014) or how a system would perform when full controllability cannot be reached (Bashirov et al., 2007).

384

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Recent works in control theory have been dealing with establishing controllability principles for complex network systems, starting with the work of (Liu et al., 2011). Specifically, the authors base their work on the concept of structural controllability (Lin, 1974), which defines how a complex network’s controllability characteristics depend mostly on the network’s connectivity (the existence of a dynamic connecting two adjacent nodes) rather than on the exact nature of the dynamic itself (in microelectronics networks, whether or not a signal exists, rather than the amplitude, frequency or magnitude of said signal). Algorithms and properties related to the computation of optimal full controllability solutions for the instance of complex networks have been researched in (Summers et al., 2016; Wang et al., 2012). In (Yuan et al., 2013), the authors develop an exact scheme for determining the minimal set of nodes necessary to completely control a complex network, based on the PBH controllability criterion. Networks whose dynamics are subject to flow-conservation criteria, as indeed is the specific instance of transportation networks, have been very recently researched in the work of (Zhao et al., 2017). In their paper, the authors remark that, when dealing with flow networks subject to conservation of flows, the modelling framework (and resulting algorithms) established earlier by (Yuan et al., 2013) can be successfully applied upon considering the network graph’s Laplacian matrix rather than its adjacency matrix. As discussed in Section 1, the main problematic affecting transportation networks (and warranting therefore considerable control efforts) is that of congestion. A considerable gap in the current literature regarding control of complex networks is indeed the lack of treatability of congested situations; this stems from the fact that congested flow dynamics are inherently nonlinear, and that therefore neither (Structural) Controllability Theory nor Popov-Belevitch-Hautus based controllability conditions can be of immediate applicability. A key contribution of the present work is that of bridging said literature gap, by drawing a connection between first order traffic flow theory, which can be formulated linearly under specific assumptions, and Structural Controllability Theory. As will be shown in Section 3, a key modelling difference lies in the choice of the state variable, being Cumulative Vehicle Number at nodes rather than flows. This allows to correctly represent formation, propagation and dissipation dynamics of congestion, thus yielding a solid modelling framework upon which existing controllability theory and algorithms can be applied. Interestingly, in control theory the problem of controllability is considered dual to that of observability, i.e. the design of locations, type and amount of sensors necessary to completely observe the state of a given dynamic system. In transportation networks, the problem of network flow observability has recently received considerable attention, starting with the work of (Castillo et al., 2008), with several subsequent works dealing with theoretical and algorithmic contributions to both full (Castillo et al., 2013; He, 2013; Ng, 2012) and partial observability scenarios (Castillo et al., 2014; Rinaldi and Viti, 2017; Viti et al., 2014). Network flow observability criteria and concepts have also been successfully applied recently in flow estimation techniques, e.g. in the works of (Bekiaris-Liberis et al., 2017; Contreras et al., 2016; Mora˘ rescu and Canudas-de-Wit, 2011). In the next Section we introduce the concept of controllability to transportation networks, extending both the concepts of controllability for complex networks and taking inspiration from the “dual” work already performed in the field of observability. Compared to the transportation economic domain, our approach tries to generalize beyond the sole control strategy of pricing (which might be unfeasible in some real-life applications) and, most importantly, tries to develop a methodology which can be applied to the time-dynamic domain. As a first stepping stone, we however limit ourselves to the time-static domain in the experimental section of this work, to maintain comparability with existing approaches (first-best and secondbest pricing approaches). 3. Methodology This Section outlines the methodological steps taken in order to adapt the theory of structural controllability of complex networks to the specific instance of transportation networks. For the sake of clarity, we begin this Section with a brief recap of what structural controllability of complex networks is, its main assumptions and its generic formulation. We then present our main contribution, i.e. the reformulation of transportation networks’ system dynamics, as well as that of the dynamics of the most standard traffic control support infrastructure tools, within the structural controllability framework. The assumptions under which our proposed methodology is considered valid are discussed in detail. Finally, we present and discuss the simple algorithms that we developed in order to perform the validation tests later shown in Section 5. 3.1. Structural control theory1 To introduce the concepts of Structural Control Theory and its application to complex networks, we begin by the simple Linear Time Invariant (LTI), Discrete-Time dynamic formulation of a generic linear dynamic system, shown in Eq. 1:

X (k ) = A · X (k − 1 ) + B · U (k − 1 )

(1) n

In the specific example of a complex network, represented by a directed graph D(N, L), X ∈ is the state vector, which captures the status of the network on its n nodes, dependent on the current (discretized) time k; A ∈ n × n is the so called 1

This section is loosely based upon Section III of the supplementary information provided by the authors of (Liu et al., 2011).

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

385

“state matrix”, which captures the influence that a given network node i has on its adjacent nodes {j, ...}, based upon the network topology adjacency matrix itself (i.e., based on which links {li } ∈ L connect which nodes). The matrix B ∈ n × m , defined “input matrix”, expresses the coupling between the M controllers equipped on a network and the controlled nodes, capturing indeed which controller’s action ug (·)⊂U(·) ∈ m affects which node(s) in the network. In classical LTI system control theory, a system is deemed controllable or fully controllable if one of its characteristic descriptive matrices, known as controllability gramian, can be shown to be of rank equal to the number of states in the system. Said matrix can be derived from the two system matrices (A, B) as follows, based on Lyapunov stability theory:

Wc =

∞ 

Am BBT ( AT )m

(2)

m=0

When dealing with LTI systems, Eq. 2 can be further simplified in the following form, obtained by successive horizontal tiling of the input matrix B left multiplied by growing powers of the state matrix A:

Wc = [B AB A2 B ... An−1 B]

(3)

As Kalman introduced in his early works, a sufficient condition for guaranteeing that the system described by (A, B) is indeed controllable is as follows:

rk(Wc ) = n

(4)

where rk(·) the matrix rank function and n the size of the (square) state matrix. The basic idea of structural controllability, as introduced in (Lin, 1974), can be summarized as follows: in networks, the exact values of dependencies between adjacent nodes might be unknown or uncertain in amplitude, but the zeroes, marking complete absence of connections (topologically and therefore dynamically), are fully known a priori. Thus, the elements {aij } ∈ A and {bij } ∈ B can either be considered as independent, free parameters or can be fixed zeroes, meaning that the two matrices (A, B) are in fact structured. Based on this consideration, Lin showed that a dynamic complex network system characterized by two matrices (A, B) can be defined to be structurally controllable if it is possible to fix the free parameters of said matrices to certain values, such that the obtained system is controllable in the classical sense. In short, such a system is either controllable or becomes controllable by slightly varying the weights of specific interconnections. Moreover, if a system can be shown to be symbolically structurally controllable, i.e. controllable for any value of the elements {aij } ∈ A and {bij } ∈ B, it is then guaranteed to be indeed controllable. In the next section, we derive the assumptions and conditions upon which controllability of transportation networks can be successfully modelled through Eqs. (1), (3) and (4). 3.2. Controllability of transportation networks 3.2.1. Modelling the dynamics of transportation networks In order to successfully fit the dynamics proper of transportation networks to the framework introduced in the previous Section, a set of key assumptions must be met. Firstly, we consider traffic dynamics as modelled through the first order traffic flow theoretical approach proper to Newell’s simplified kinematic wave theory (Newell, 1993), with the flow-density relationship expressed by the triangular shaped fundamental diagram. This latter assumption yielding considerable simplifications to the modelling framework, thanks to its linearity, while still allowing to correctly capture congestion dynamics. We choose as our state variable X(k) the Cumulative Vehicle Number (CVN) at each node {1, ..., n} ∈ N. To derive the equation(s) capturing vehicle propagation dynamics, we begin by considering the small network cutout of Fig. 1(a), and the CVN evolution instance of Fig. 1(b). We first focus on characterising the dynamics of nodes 1 and 3, and their interaction through link 2, ignoring for the moment any other incoming/outgoing link from either node. In this setting, following first order dynamics, the following equation holds:

x 3 ( k ) = a 1 , 3 x 1 ( k − τ1 , 3 ( k − 1 ) )

(5)

The CVN on node 1, x1 (k) acts as the upstream cumulative of link 2, while the CVN on node 3 x3 (k) as the link’s downstream cumulative. Fundamental traffic measures such as link in-outflow f in ,li (k ), fout ,li (k ), speed vli (k ) and density κli (k ) can be derived by performing opportune operations on these two values, as follows:

fin ,l2 (k ) = x1 (k ) − x1 (k − 1 ) fout ,l2 (k ) = x3 (k ) − x3 (k − 1 ) vl2 (k ) = τ 1(k)

κl2 (k ) =

(6)

1,3

x1 (k )−x3 (k ) Ll 2

Link travel time, τ 1,3 (k), can be computed as the time-difference between the two cumulatives x1 (k), x3 (k) when they exhibit equal values:

τ1,3 (k ) = (k − kt − ) : x1 (kt − ) = x3 (k )

(7)

386

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 1. Three links network cutout (a) and example node CVN evolution (b).

This value is of course time-dependent, and is connected to the system’s level of congestion: in congested situations this value will increase, and in extreme situations (complete breakdown, for example following an accident that entirely blocks the road) it becomes unbounded. Considering the three links highlighted in Fig. 1(a), the total system dynamics would thus be captured by the following set of equations:

x3 (k ) = a1,3 x1 (k − τ1,3 (k − 1 )) x4 (k ) = a3,4 x3 (k − τ3,4 (k − 1 )) x12 (k ) = a3,12 x3 (k − τ3,12 (k − 1 ))

(8)

Elements a3,4 and a3,12 capture, in this instance, the split fraction of vehicles traveling from node 3 towards nodes 4 and 12, respectively. As we discussed earlier in Section 1, transportation networks exhibit a degree of self-organisation: in this work we specifically consider the effects of route choice, by including an additional source of dynamics to our problem. Specifically, we assume that the system behaves according to deterministic User Equilibrium principles, i.e. each user is expected to minimize his/her own travel time by appropriately choosing the cheapest route alternative at the moment in which he/she enters the system. Stochastic User Equilibrium could also be considered, by introducing probability distributions for the travel time variables τ i,j (k). Explicit consideration of route choice dynamics in the problem allows us to extend the system dynamics (8), by including route flow conservation effects at nodes. Consider the small network example depicted in Fig. 2, where the two possible routes for reaching node 11 from node 1 are shown (in green and red). Applying conservation of vehicles at node 3, we can easily extract the following relationships between the state variables of nodes 1, 3, 4, 11 and 12:

x3 (k ) = a1,3 x1 (k − τ1,3 (k − 1 )) x4 (k ) = [x3 (k − τ3,4 (k − 1 )) − x12 (k − τ3,4 (k − 1 ) + τ3,12 (k − 1 ))] x12 (k ) = [x3 (k − τ3,12 (k − 1 )) − x4 (k − τ3,12 (k − 1 ) + τ3,4 (k − 1 ))] x11 (k ) = a4,11 x4 (k − τ4,11 (k − 1 )) + a12,11 x12 (k − τ12,11 (k − 1 ))

(9)

The two conservation equations relate the current values of x4 and x12 to that of x3 , simply assuming that the evolution of the two route flows (and thus of the corresponding cumulative vehicle numbers) obeys conservation of vehicles. In short, the sum of vehicles passing through nodes 12 and 4 at time k must be equal to that of the vehicles that passed through node 3 at appropriate previous times. Eq. (9) can only be considered as valid dynamics under deterministic User Equilibrium assumptions, if both routes are indeed utilized. If this condition is not met, e.g. when only the red route {3-12-11} is utilized, the equations collapse to simple node to node propagation dynamics. Origin centroids are considered connected to the corresponding nodes in the network via instantaneous travel time connectors. In the example of Fig. 1, considering a centroid placed on node 1, we can therefore assume that node 1 s CVN behavior follows exactly that of the demand flow entering the network at the corresponding centroid. Destination centroids are modelled in a similar fashion. Nodes acting both as origins/destinations and traversal nodes can be modelled by deriving appropriate values for the elements {aij }.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

387

Fig. 2. Network cutout with highlighted routes.

The total system dynamics thus far can be summarized by the following equation:

X (k ) = A · X (k − T(k − 1 ))

(10)

where X(k) = [x1 (k),..., xn is the CVN based state vector at nodes, A = {aij }∀(i,j) ∈ L is the state matrix, containing the node-to-node incidence matrix characterizing the link set L and static route split fractions, and T(k − 1 ) = {τi j (k − 1 )}∀(i, j ) ∈ L is the (time-dependent) link travel time vector. So far, the following assumptions have been considered: (k)]T

•

•

•

•

Linearity: the dynamics of traffic obey Newell’s first-order Kinematic Wave Theory, under the assumption of triangularshaped fundamental diagram. This yields piecewise affine dynamics on the chosen state variables (the node CVNs); Time Invariance: the relationships between adjacent nodes in the network (route splits) do not change with time, a condition which will be relaxed shortly; Deterministic User Equilibrium route choice: users entering the network have perfect information of the network travel times they will experience, and react accordingly in terms of chosen routes; Courant-Friedrichs-Lewy condition compliant time discretisation interval Ts .

In control theoretical terms, the explicit inclusion of travel time dependency in the dynamics of Eq. (10) represents a (time varying) pure delay component: while such a dynamic cannot be directly captured within the complex network controllability framework of Section 3.1, under the following specific reformulation this problem can be circumvented. Without introducing any loss of generality, we can rewrite Eq. (10) as follows:

X (k ) = A¯ (k − 1 ) · X (k − 1 )

(11)

This Linear, Time Variant (LTV) reformulation allows us to successfully discard the pure delay vector T, as well as to explicitly consider time-dependent route split fractions, by incorporating both sources of information in matrix A¯ (k − 1 ), whose elements are then defined as:

a¯ i, j (k − 1 ) = ai, j (k − 1 ) ·

xi (k − τi, j (k − 1 )) xi ( k − 1 )

(12)

that is, individual elements a¯ i, j (k ), other than time dynamic route split information ai,j (k), also capture the ratio between the preceding node cumulative at time (k − τ i,j (k − 1)) and that at time (k − 1). This ratio is an indicator of congestion: the lower this value, the more strongly congested the underlying network link, whereas a value of 1 indicates that the road is flowing freely and that the free flow speed is such that the distance between nodes i and j can be covered in a single time step Ts . As long as perfect, punctual knowledge of matrix A¯ (k ) is available at any time k, Eq. (11) is in fact a correct and complete representation of the given transportation networks’ dynamics, under the assumptions above, as long as the traffic system doesn’t experience complete breakdown (see Appendix B). Collecting exact knowledge of link travel times and turning fractions is however far from realistic, due both to prohibitive sensing requirements and generally unreliable, error-prone measurements. Luckily, our aim is that of capturing the structural controllability aspects of a transportation network, rather than the network itself. This shifts our focus to those conditions and situations in which members of the set {a¯ i j (k )} would become zero, rather than in their exact values over different transitional states. Thus, for the sake of establishing a controllability baseline, whether or not congestion is present on the

388

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 3. Network cutout with pricing controller installed on link 7.

network is substantially irrelevant, as long as the amount of vehicles flowing from one node towards the next does not collapse to zero over the discretisation period Ts . Based on these considerations, we can model the vehicular flow characteristics of a given transportation network, for the sake of assessing its controllability, through Eq. (13), assuming that all links in the network are operating in a steady-state condition, be it congested or uncongested:

X (k ) = A¯ · X (k − 1 )

(13)

returning therefore to the desired LTI form. Conditions and instances in which some of these steady-state elements a¯ i, j would collapse to zero would instead be studied separately, in a case by case fashion. It is however to be noted that, in real networks, conditions in which said elements would collapse to zero due to congestion are quite rare: in a given time horizon, average speeds would have to drop to exactly zero, causing infinite travel times.2 We believe that incidents causing complete breakdowns (gridlock), while possible, should not play a determinant role to the problem of optimally locating controllers in a network. In future works we will explore strategies to cope with loss of controllability in such extreme conditions.

3.2.2. Modelling the dynamics of controllers on transportation networks Besides modelling the dynamics of vehicle propagation and route choice, in order to build a framework for establishing controllability of transportation networks it is also necessary to correctly capture the dynamics of the different kinds of control support infrastructure (the actuators) that operate on said networks. In this Section, our objective is therefore that of deriving the shape and structure of the “input matrix” B when considering two common control tools on transportation networks, pricing controllers and traffic signals, our final objective being that of capturing the full controllability aspects of a controlled transportation network within the generic LTI dynamic system Eq. (1). To achieve this, we must relate the macroscopic effects of said controllers with the already established flow dynamics of Eq. (13). We begin by examining the dynamics of pricing controllers. These tools exert direct influence on the traversal cost of given links in the network, by levying a monetary price (toll) on users traversing said links. This, in turn, has an indirect effect on the amount of vehicles using the link at a given time, as users might choose less costly alternatives when they begin their trips. In terms of the system dynamics of Eq. (13), this would translate to a reduction in the corresponding link weight aij , the effect of which must be modelled by the input equation B · U(k − 1). We derive said model by considering the example of Fig. 3. The dynamics characterising this specific instance are described by the following set of equations:

x12 (k ) = (a3,12 − ξ12 (k − 1 ))x3 (k − 1 ) ξ12 (k ) = f (g7 (k )) ∈ [−a3,12 , 0]

(14)

2 Flows becoming zero due to lack of demand is a further instance of these elements collapsing to zero. However, a transportation network bearing zero demand certainly requires no control effort.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

389

Fig. 4. Simple example with traffic light controller installed on node 3.

that is, levying a toll g7 on link 7 causes a reduction effect ξ 12 on the amount of vehicles propagating from node 3 to node 12, function of the toll level itself. This effect can be straightforwardly reformulated in additive terms, as follows:

x12 (k ) = a3,12 x3 (k − 1 ) + b3,12 u12 (k − 1 ) b3,12 = 1 u12 (k − 1 ) = −ξ12 (k − 1 )x3 (k − 1 )

(15)

thus meeting the required shape and structure proper to Eq. (1). It’s important to consider that, for Eq. (15) to hold, the control signal must include a specific component of the system state at time (k − 1), which can only be assumed feasible under full observability conditions. However, exact knowledge of this information is again irrelevant in terms of establishing structural controllability, as is the specific toll level being applied and the nature of its relationship with the reduction effect, since the control signal vector U(·) is indeed not a component of either Eq. (3) or (4). We consider traffic light dynamics limiting ourselves to the instance of two competing flows accessing a common destination node, starting again from a very simple example, shown in Fig. 4. A traffic light equipped on node 3 distributes the node’s capacity to the two incoming flows, from node 1 and node 2 respectively, in an exclusive fashion. From a simplified, macroscopic perspective, this yields again reduction factors for both elements a13 and a23 , multiplicative in nature and joined by a simple summation constraint, yielding the following dynamics:

x3 (k ) = φ13 (k − 1 ) · a13 · x1 (k − 1 ) + φ23 (k − 1 ) · a23 · x2 (k − 1 ) φ13 (· ) + φ23 (· ) ≤ 1

(16)

assuming that all-red time could be present in the microscopic behavior of the chosen traffic light. As for the pricing controller, we reformulate (16) in an additive form, obtaining the following equations:

x3 (k ) = a13 x1 (k − 1 ) + a23 x2 (k − 1 ) + b13 u3 (k − 1 ) + b23 u3 (k − 1 ) where b13 = b23 = 1

(17)

u3 (k − 1 ) = a13 x1 (k − 1 ) · (φ13 (k − 1 ) − 1 )+ 

+ a23 x2 (k − 1 ) · (φ23 (k − 1 ) − 1 )

φi, j ≤ (1 − ε )

which can be shown to be exactly equivalent to (16). Notice that we essentially model the two separate movements controlled by the traffic light as disjoint (resulting indeed in two elements being simultaneously populated in the B matrix). The coupling of the two dynamics through their common cycle time is represented in the U(·) signal in terms of the summation constraint. It is important to consider that disjoint modelling can only be considered appropriate when assuming that the all-red portion of the cycle time is time-varying, meaning that the two competing flows can be both reduced at will. Assuming fixed all-red time (including a value of 0) would effectively reduce the state space of the controlled intersection by one dimension, thus yielding a non-controllable subsystem. For the purposes of controllability, the exact values of the previous system state(s) as well as those of the specific signal are irrelevant, as long as a minimum service rate ε is guaranteed. This latter condition is necessary to ensure that a nonzero quantity of vehicles always reaches the destination node from the two source nodes, in order to avoid loss of controllability.

390

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

By putting together the vehicle propagation / route choice modelling aspects of Eq. (13), the pricing dynamics modelling of Eq. (15) and the traffic light dynamics modelling of Eq. (17) we finally obtain the complete system dynamics:

X (k ) = A¯ · X (k − 1 ) + B · U (k − 1 ) A¯ = {a¯ i, j } ∈ n×n B = {bi, j } ∈ n×m m = | G| + || 1∀ j : g j > 0 bi j = 1∀(i, j ) : ∃h ∈ N : φi j + φh j ≤ 1 0⎧ otherwise ⎪−ξ j (k − 1 )xi (k − 1 ) ∀(i, j ) : g j > 0 ⎨ a¯ i j x j (k − 1 ) · (φi j (k − 1 ) − 1 )+ u j (k ) = ⎪ ⎩+a¯ h j x j (k − 1 ) · (φh j (k − 1 ) − 1 )∀ j : ∃(i, j, h ) ∈ N : φi j + φh j ≤ 1 0 otherwise

(18)

where n = |N| is the total amount of nodes in the network, G = {gj } is the set of pricing controllers equipped on the network and  = {φ ij } that of traffic light controllers. 3.2.3. Computing the degree of controllability for a given transportation network Finally, in order to determine the degree of controllability for the chosen transportation network, the controllability gramian matrix Wc can be computed from the two system matrices (A¯ , B ) derived in Eq. (18). Since the chosen system description is Linear Time Invariant, the matrix takes the following form:



Wc = B

A¯ B

A¯ 2 B

···



A¯ n−1 B

(19)

which can be further transformed as follows, by grouping multiplications:

⎡



Wc = I

A¯

A¯ 2

...

B ⎢0 ⎢ A¯ n−1 · ⎢0

⎢ ⎣0 0

0 B 0

0 0 B

0 0

0 0

0 0 0 .. . 0

⎤

0 0⎥ ⎥ 0⎥

⎥ 0⎦

(20)

B

This latter form allows to clearly highlight the presence of a constant, systematic component of the gramian matrix, whose shape and structure is solely due to the underlying network topology and to the equilibrium route set being chosen by the users. The only source of variability, depending on the quantity, location and kind of controllers being installed on the network, is the repeated diagonal tiling of matrix B, and correctly represents our desired design component. The degree of controllability for a given network D(N, L), route set R and controller set M = G∪, can therefore be computed as follows:

θ (B ) =

rk(Wc (B )) n

(21)

representing a normalized metric, whose value reaches 1 if and only if every single node in the network can be effectively controlled in terms of cumulative vehicle number. In the remainder of the methodology section we will introduce a formulation for the minimal controller location problem subject to full controllability and introduce the chosen solution scheme, which is adapted from the exact approach developed by (Yuan et al., 2013). 3.3. Controller location problem on transportation networks Thanks to the methodology developed thus far, we can finally formulate the optimal controller location problem for transportation networks, based on the network’s topology and route choice description. Specifically, we formulate an approach that places controllers on a network so to achieve full controllability by equipping the minimum possible amount of controllers, irrespective of their kind. Other considerations, such as CSI installation costs or reliability in case of loss of controllability or CSI failure, could also be modelled through an appropriate objective function. The resulting minimal controller location problem can be formulated as follows:

min ||B||F s.t.{θ (B ) = 1

(22)

where ||•||F is the Frobenius matrix norm and the elements {bij } ∈ B depend on the chosen controllers’ locations and types, following Eq. (18). In this work we derive exact solutions to a constrained version of problem (22), through an adaptation of the maximum geometric multiplicity approach introduced by (Yuan et al., 2013). Specifically, we constrain the acceptable locations in which

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

391

controllers might be installed, excluding all origin/destination nodes and their immediate successors. Since O/D centroids are not physical nodes, but rather coarse geographical representations of traveller’s behavior, exerting direct control over them is physically impossible, and thus undesirable from a network design perspective. By separating network nodes into candidate (C) and non-candidate (I) sets, we can refine problem ((22) as follows:

min ||B||F s.t.{θ (B ) ≥ 1 − ||NNI|| N = NC ∪ NI

(23)

The rank condition in (23) states that the degree of controllability should be at least enough to ensure that all nodes in the candidate set NC are indeed controlled, however this condition could be improved upon by solutions who manage to extend controllability on nodes beyond the candidate set, by indirectly affecting some nodes pertaining to the non-candidate set. The approach of Section 3.3.2 seeks solutions that guarantee full controllability on the subset of nodes NC , with no regard to whether or not extra controllability is gained on nodes pertaining to the set NI . We first introduce a simple algorithm which automatically extracts state matrix A from a given network topology and route set description, followed by the implemented controller location approach. 3.3.1. Extracting the state matrix a from network topology and route set In order to correctly populate the state matrix A with its appropriate non-zero elements we combine, through algebraic transformations and substitutions, the two adjacency matrices capturing the network topology and its relationship with the users’ route set. The algorithm’s inner working is sketched, in pseudo-code form, and dubbed Algorithm 3,1. Algorithm 3.1 Given a directed graph network description D(N, L) and a corresponding route set R = {ri }. 1. 2. 3. 4. 5. 6. 7.

compute the node to node incidence matrix Pnn compute the note to route incidence matrix Pnr set A ← 0.5 · Pnn compute the row reduced echelon form of matrix Pnr for each row s in rre f (Pnr ) set A(i, : ) ← s end(for)

We begin by forming the node to node incidence matrix Pnn = {0, 1} ∈ Rn×n , representing the physical topology of the network: an element pi,j will be equal to 1 if node i is a predecessor of node j. The node to route incidence matrix Pnr = {0, 1} ∈ Rn×|R| , representing the route choice behavior of individuals, is obtained similarly: an element pi,r will be set to 1 if node i is traversed by route r and 0 otherwise. Without additional information stemming from the route set, an acceptable guess for the state matrix A is none other than the node to node incidence matrix itself. As detailed in Section 3.2.2, we are not interested in the magnitude of the elements {aij }, we thus arbitrarily set all of them equal to 0.5. Note that, while any other value in the closed interval 0 < aij ≤ 1 would also be acceptable, values exceeding 1 would violate the assumptions that i) travel times in the network cannot be exactly zero and ii) vehicle conservation laws apply at nodes, following kinematic wave theory. If route information is available, further node to node relationships can be extracted from matrix Pnr by transforming it in its row reduced echelon form. This extracts dependent-independent node to node relationships, by algebraically deriving how nodes in the network are related to others through the given route set R. In steps 5 through 7, these informationenriched relationships replace the ones previously obtained from the physical topology alone, yielding the complete state matrix. It is important to consider the impact of route information on the overall problem: each captured route choice phenomenon represents gained knowledge on a system’s additional degree of freedom, which requires additional control efforts. Disregarding route choice entirely might result in a set of control locations that underestimates the overall system dynamics, yielding sub-optimal performances. Conversely, placing controllers considering a route set larger than the one being practically utilised by the users (but, critically, a superset of the latter), while resulting in a redundant controller configuration, will always achieve full controllability performances. This aspect is investigated, in terms of the proposed approach’s sensitivity to incomplete route information, in Section 5. 3.3.2. Exact solution approach to the pricing controller location problem In their recent work, (Yuan et al., 2013) developed an exact approach to the problem of determining the minimum set of controllers capable of guaranteeing full controllability for complex networks, based on the Popov-Belevitch-Hautus (PBH) controllability condition (Hautus, 1969). Specifically, the authors postulate that: i) the minimum amount of controllers necessary to steer a given complex network bearing connectivity matrix A is equal to the maximum geometric multiplicity μ(λM ) of its eigenvalue λM ;

392

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

ii) the locations on which to install said controllers can be derived by computing the column canonical form of the matrix (A − λM · I), and determining the resulting linearly dependent rows. Condition ii) specifically relates directly to the PBH controllability theory, which states that a dynamical system with state matrix A and control matrix B is controllable i.i.f. the following equation is satisfied:

rk(c · I − A, B ) = |N|∀c ∈ C

(24) c = λM .

and simply selects the complex number It is important to remark here that in the theoretical works of (Liu et al., 2011; Yuan et al., 2013; Zhao et al., 2017), controllers have been assumed as influencing one single node at a time. This is indeed the case for pricing controllers, which will be the focus of the remainder of this work. Proper handling of traffic light controller placement, including considerations of whether or not the existing theory can be directly applied to mixedcontroller-type instances, is instead left to future research. We implement the approach of Yuan et al. in Algorithm 3.2, explicitly considering the candidate nodes restrictions introduced in Eq. (23). Algorithm 3.2 Given a state matrix network description A, a candidate node set NC ⊆N. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.

compute the reduced connectivity matrix AC by eliminating rows and columns not pertaining to the candidate node set compute the eigenvalue set ← eig(AC ) for each λi ∈ : compute null(AC − λi · I) ∈ Cms × ns (compute the size of the null space of matrix (AC − λi · I)) set μ(λi ) ← ns (compute the geom. multiplicity of eigenvalue λi as the column size of the null space of matrix (AC − λi · I)) end(for) set λM ← arg maxi (μ(λi )) set AM ← (AC − λM · I) compute the column canonical form A M identify dependent rows in A M compute the node set pertaining to the rows identified in step 10

After reducing the adjacency matrix to the node set of interest and computing its corresponding eigenvalues, steps 3–6 identify the geometric multiplicity of each eigenvalue by computing the dimension of the eigenspace of matrix AC for each eigenvalue (which can be computed as the size of the null space of matrix (AC − λi · I)). Once this procedure is complete, an eigenvalue bearing maximum multiplicity is identified in step 8, and an exact set of minimal controller nodes can be identified by reordering the elements of matrix (AC − λM · I) in order to isolate linearly dependent-independent components. Finally, Step 11 relates the indices determined for the reduced connectivity matrix AC to the full network’s node set N. 4. Experimental setup To validate whether placing controllers according to our proposed network controllability approach proves indeed valuable in terms of network-wide control policies’ performance, in this Section we outline two sets of simulation experiments based on two artificially constructed networks. In order to have an objective term of comparison, we perform all tests and experiments assuming that vehicles behave according to static, deterministic User Equilibrium, that is, no time dynamics are included. While our model (18) is perfectly suited to consider time dynamics, simulations in the static domain enable us to correctly compute the System Optimal solution for all networks and examples, which represents the de-facto best possible flow distribution scenario on all routes, and a natural benchmark solution to compare to. To this end, we also limit the current comparisons to the instance of pricing controllers, as the introduction of traffic light controllers voids comparability to a System Optimal solution in terms of Total Cost, other than violating the assumptions leading to the utilisation of the exact algorithm of Section 3.2.2. Deterministic User Equilibrium is computed through the Method of Successive Averages (MSA) algorithm, specifically in its flow-averaging formulation, based upon deterministic all-or-nothing network loading (Sheffi, 1985). The available route set is enumerated and fixed a-priori, following the classical k-shortest path approach (Yen, 1971), based on free flow traversal costs, that is, considering link costs cl (fl ),fl = 0 ∀l ∈ L . For each network instance and controller location policy, we compare the results of minimizing the Total Cost objective function subject to equilibrium conditions:



fl (cl ( fl , g)) · cl ( fl , g)  l∈L (i ) f ∈ S f ∗ s.t. (ii )S ∗ : { f : f = arg min   fl c (y )dy , f ∈ S } f l l f 0 l ming

(25)

l∈L

where S f ∗ represents the feasible set of equilibrium flows and Sf the larger set of demand feasible flows. Constraint (ii) implies that all users behave according to User Equilibrium, distributing themselves on the network such that their individually perceived costs are minimized.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406 Table 1 BPR function Test Case 1. Link Link Link Link Link

parameters

393

for

Parameters 10 11 12 13

α 10 = 0.5, β 10 = 2 α 11 = 0.2, β 11 = 4 α 12 = 0.5, β 12 = 4 α 13 = 0.4, β 13 = 2

Specifically, we employ our recent reformulation of (25), devised to help overcoming some of the problems’ innate irregularities in terms of solution space shape, and thus achieving better convergence characteristics towards significant minima (Rinaldi et al., 2017). The corresponding objective function is as follows:

ming,λ



fl (cl ( fl , g) ) · cl ( f, g) +

 0 m⊆TSO

λm · |(Aeq [m, l ] · c )| +

⎧ l∈L (i ) f ∈ S f∗ ⎪  ⎨   fl s.t. (ii )S f ∗ : f : f = arg min c y d y , f ∈ S ( ) l l l f 0 ⎪ l∈L ⎩ (iii )λm , λv ≥ 0

 v

0 ⊆TSO

  λv · Aineq [v, l ] · c − ε

(26)

where the additional members of the objective function represent additional costs incurred by users utilizing routes different than those that are used in the System Optimal route configuration. For the first test case we perform extensive exploration of the solution space considering different sets of controllers, in order to showcase how the reachable minima change with the given CSI, and provide a proof-of-concept of how this is reflected in the value of the degree of controllability. For the second test case, we compare four different controller location approaches, and we measure their impact on the minimization of (26), compared to System Optimal solutions. Optimization is performed using MathworksTM MATLAB® Optimization Toolbox’s fmincon function, configured to utilise the quasi-newton BFGS approach and central differences approximation for gradient computation. The pricing controllers’ values are lower bound by g ≥ 0, while no upper bound is provided. For this network we also investigate the impact of incomplete route information, by introducing a mismatch between the amount of routes considered when solving the controller location problem and that considered during optimisation. 5. Experimental results 5.1. Test case 1 In the first test case we perform a fully explorative analysis of objective function (25)’s solution space, and how this varies for different sets of equipped controllers, on a very simple network, shown in Fig. 5. We consider OD demand flowing vertically along couples A-D, B-E and C-F, and we assume that link cost functions cl (fl ) are constant for all links apart from links 10–13, who instead are equipped with the BPR cost function cl (fl ) = 1 + α l · (fl /400)β , with parameter values as presented in Table 1. Specifically, we focus on how the global minimum’s location and objective function value directly depends on the chosen amount and location of controllers, how an insufficient amount of controllers implies a suboptimality in terms of global minimum function value with respect to system optimum, and how this behavior is correctly captured by computing the degree of controllability (21). This simple example and small network are chosen specifically to showcase how relevant the choice of controllers can be, even in extremely oversimplified scenarios. We begin by determining the Total Cost objective function values in both User Equilibrium (no control situation) and System Optimum conditions:

T CUE = 1.3322 · 105 T CSO = 1.3287 · 105 For such a small network and example, the difference between the two is rather trivial, it is however interesting to see how far from the System Optimal performance different controller configurations can get. We begin this exploration by examining the Total Cost objective function’s shape when a single pricing controller is equipped on either link 10, 11, 12 or 13, as shown in Fig. 6. Interestingly, equipping controllers g10 or g13 alone brings no gain to the system as a whole: the global minimum for both resulting objective functions is at g∗ = 0, thus yielding Total Cost equal, by definition, to the User Equilibrium value (marked in Fig. 6 by the black dashed line). Controllers g11 and g12 instead both exhibit individual global minima in the vicinity of g∗ = 2 (marked in Fig. 6 by the light green dashed line), bearing a visible reduction of Total Cost. For these four individual controllers, the percental distance from System Optimum and the corresponding degree of controllability are as

394

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 5. Three OD network.

follows:

T Cg10 = 0.2606%

T Cg11 = 0.1543%

T Cg12 = 0.1829%

T Cg13 = 0.2607%

θ (Bg10 ) = 1/3 θ (Bg11 ) = 1/3 θ (Bg12 ) = 1/3 θ (Bg13 ) = 1/3

We then consider all six possible couples of controllers, [g10 ,g11 ],[g10 ,g12 ],[g10 ,g13 ],[g11 ,g12 ],[g11 ,g12 ],[g12 ,g13 ], and show the respective Total Cost solution spaces in terms of contours in Fig. 7(a-f). The globally minimal points are easily identified by dark colouring, and are consistently found in the range g < 2.5. While not very clear to the naked eye, within these six combinations a clear “best case” arises, as can be seen when considering the percental distance from System Optimal total cost:

T Cg10 ,g11 = 0.1548%

T Cg10 ,g13 = 0.2616%

T Cg11 ,g13 = 0.1548%

T Cg10 ,g12 = 0.1835%

T Cg11 ,g12 = 0.006%

T Cg12 ,g13 = 0.1835%

The degree of controllability is again constant for all combinations, valued at θ (B) = 2/3. This is unsurprising: even though combination [g11 ,g12 ] is closest to the System Optimal value, for this network no less than three controllers are necessary to achieve full controllability. It’s important to notice that the degree of controllability θ , while perfectly suitable to describe whether or not a given set of controllers is capable of fully steering a given network towards optimality, fails to differentiate between partial controllability solutions, and thus gives no indication in terms of which of the six combinations above best approximates System Optimum. Future research will focus on developing metrics and heuristics that are also suitable for partial controllability applications, investigating whether it’s indeed possible to characterize not only the size of solution space which is rendered controllable by a chosen configuration, but its qualities in terms of objective function performance. This will be necessary especially when dealing with large-scale networks, where full controllability solutions might be impractical.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

395

Fig. 6. Total Cost objective function shapes for separate controllers.

5.2. Test case 2 To generalize the results shown on the simpler instance of Test Case 1, in this example we consider a larger network, exhibiting increased complexity, shown in Fig. 8. Origin-Destination demand flows again vertically, from origin nodes 1 through 5 towards destination nodes 92 through 96, respectively. We assume all links of this network to be equipped with the standard BPR cost function cl ( fl ) = ll /v fl + αl ·

( fl /ca pl )βl .

Values for the free flow speed v fl , the link capacity capl and the two BPR parameters α l ,β l are algorithmically defined for the topological shortest paths on an OD-by-OD basis, such that the OD pairs closer to the centre exhibit higher sensitivity to congested situations with respect to the OD pairs on the exterior of the network (e.g. the shortest path of OD pair 3–94 is more congestion sensitive than the shortest path of OD pair 1–92 or 5–96). Likewise, the free flow travel time conditions of said shortest routes are also devised such that the central ODs have shorter free flow travel times compared to the border ones. Links not pertaining to any OD’s free flow shortest paths are instead equipped with standard values of vl f = 50, ca pl =

10 0 0, αl = 0.25, βl = 2. This cost configuration induces conditions upon which the users’ route choice will naturally prefer routes crossing the central OD pair, even though this becomes very costly as flows increase. To achieve system optimal conditions, controllers on the network will therefore have to prevent flows coming from the external ODs from overflooding the centre, in order not to excessively penalise the weaker, congestion sensitive network portions. We further consider link lengths as random variables bearing the following distribution: ll ∼ N(5, 0.25). This source of stochasticity allows us to directly influence the topological shortest paths set, and, as a consequence, most of the link parameters on the network. By performing random draws from the link length distribution, we can therefore test our approach on considerably diverse network instances, and obtain a statistic of how different controller placement approaches perform in terms of Total Cost minimization, compared to the respective System Optimal solution. For each network instance we solve problem (26) considering four different controller sets: gf ,gv ,gc ,gr , obtained by applying, respectively, the first best pricing location approach (all links are equipped with a pricing controller), an implementation

396

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 7. Total Cost objective function solution space shape for different couples of controllers.

of the second best pricing location approach introduced in (Verhoef, 2002b), our own approach based on Algorithms 3.1 and 3.2, and finally a random configuration of controllers, extracted from a uniform distribution over all links in the network. For the sake of comparability, with the exception of first best pricing, the size of controller sets gv and gr is set equal to that of our own approach’s gc . For each network instance and controller set, we collect two indicators: the percental distance between the Total Cost value at the optimal point of optimization (26) and System Optimum, computed as TC = 100 · (TC(g∗ ) − TCSO )/TCSO , and the

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

397

Fig. 8. Multiple OD network.

total computational time for each instance, expressed in seconds. The results of these tests are summarized in Fig. 9(a,b), in terms of box plots for all four controller location approaches and random draws. Analysing the two figures, it becomes immediately clear that locating pricing controllers according to the first best tolling policy is, trivially, statistically similar to computing system optimum. This however comes at a considerable computational cost, in average more than twice that of all other approaches: first best tolling optimization always comprises of one variable per each link of the network (118), about 7-fold the average amount of controllers utilised by the other approaches. Moreover, first-best tolling is impractical to implement in real-life scenarios, due to prohibitive equipment costs. Comparing the remaining three approaches, while randomly placing controllers clearly yields little benefit, as only a few lucky combinations manage to get close to system optimal performances, locating controllers following our approach strongly outperforms all others, notably including second best pricing a-la Verhoef. This large gap in performances is notable, especially given the following consideration: the approach of Verhoef employs dc ( f ) exact knowledge both in terms of link flows fl and first order derivatives of link costs dl f l at equilibrium in order to l

determine the controller locations directly yielding the highest marginal gain in terms of Total Cost objective function. Our approach, instead, is based on topological information alone, both in terms of node to node adjacency and users’ considered route set, and infers optimal controller locations based solely on these inputs. Given the considerably less intensive information requirements, these test results are indeed extremely promising, especially given the fact that our approach is aimed at general networks and time dynamic scenarios, where information on either link flows or, especially, derivatives of link costs are simply not available. 5.3. Sensitivity of controller locations to available route information In general applications, employing a-priori generated route information might cause lossy performance, as indeed in operational conditions there is no guarantee that the utilised route set coincides with the enumerated one. In this Section we investigate whether our proposed methodology exhibits a satisfactory level of robustness to this condition, by introducing a mismatch between the amount of routes enumerated while placing controllers and the amount actually operationally considered by users. Specifically, while the route set for all User Equilibrium assignment procedures is generated through Yen’s K-Shortest Path algorithm considering a parameter k = 10, we compare the controller location solutions of both our proposed approach and the second-best pricing approach of Verhoef considering an increasing level of “route information error” .

398

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 9. Test Case 2 result statistics.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

399

Fig. 10. Sensitivity analysis of Controller Location solutions to route information mismatch.

We solve the controller location problem through the two algorithms considering an error level of 50% (k = 5), 70% (k = 3) and 90% (k = 1). The corresponding controller locations are then evaluated in terms of Total Cost optimization performance following exactly the same structure as the previous Section, throughout the same 100 replications of the network of Fig. 8. The corresponding results are shown in Fig. 10. Trivially, as the amount of route information mismatch increases (evolving towards the right hand side of Fig. 10), the statistics of both controller location algorithms worsen. However, our proposed approach clearly exhibits increased robustness compared to second-best pricing a-la Verhoef, maintaining a very solid level of performances even when the provided information is a fraction of the original, correct route choice. In future research we will investigate whether this property holds for general networks, and whether different route choice models (rather than parameters) have a harsher impact on the overall optimisation performance. 6. Conclusions In this work we investigated the extent to which choosing locations, kinds and amounts of controllers to be installed in a transportation network affect the performances of network-wide control approaches. To achieve this goal, we formulated the theory of (structural) controllability of complex network systems for the specific instance of transportation networks, detailing the underlying modelling assumptions necessary to correctly capture the emergent dynamics proper of traffic’s behavior, both in terms of vehicle propagation and route choice. Thanks to the newly introduced methodology, we then introduced a controller location problem aiming to i) achieving full controllability of the underlying network and ii) minimising the amount of controllers necessary. An exact approach to the pricing controller location problem has been implemented, based on current state-of-the-art, in order to perform validation through two test cases.

400

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Through the first, explorative test case we showcased how the solution space for a network-wide minimisation of Total Cost changes with respect to the quantity and location of pricing controllers installed thereon, especially in terms of accessible minima, and how the developed full controllability approach correctly identifies the amount of controllers necessary to reach System Optimal performances. A second test case instead focused on comparing our developed methodology with three other controller location approaches, by equipping a controller set according to each, and successively minimising the Total Cost objective function. The test was repeated on 100 replications of a given network, in which base link parameters were randomized. Test results showed how locating controllers following our proposed approach yields network-wide performances far outperforming those of a well-known second-best pricing algorithm, while being considerably more parsimonious in terms of information requirements. Furthermore, robustness characteristics to the supplied route information have been empirically investigated. Future research directions include, firstly, developing solution algorithms for the generic controller location problem, in order to correctly identify the minimum amount of controllers necessary to achieve full controllability of transportation networks irrespective of kind (thus including also traffic lights). Performance validation will further be performed in time-dynamic scenarios, comparing again different controller location approaches, including discrepancies in route choice behavior and demand fluctuations. As pointed out in the results section, the simple controllability metric we proposed here (21) is incapable of distinguishing different partial controllability solutions, a characteristic which becomes very unattractive when considering real life scenarios, where budget constraints might prevent full controllability installations. Correctly capturing the effects of partial controllability is therefore another topic of future interest. Finally, algorithms to detect and respond to loss of controllability in case of incidental conditions also represent a fundamental challenge to be tackled. Appendix A. implementation details of Verhoef’s second-best pricing We implement Verhoef’s second-best optimisation problem adapting the Lagrangian formulation of (Verhoef, 2002a) to our notation and instance. Verhoef’s original notation for the “shadow price of non-optimal pricing” lagrangian multipliers λp is as follows (Eq. (10) of the original work): J 

λp =

j=1



δ jp ·

P 

q=1

     J J P I P      δ jq · Nq · c j  − λq δ jp · δ jq · c j  + δip · δiq · λq · Di  − δ jp · δ jq · f j q=1,q= p

q=1,q= p

i=1

j=1

J  j=1

δ jp · c j  −

I  i=1

δip · Di

j=1

(27)



where i, k j p, q Nq cj

indices for OD pairs, denoted 1, ..., I index for links, denoted 1, ..., J index for paths, denoted 1, ..., P, Pi being a path for OD pair i the path flow on path q The average cost function for the use of link j A dummy that takes on the value of 1 if link j belongs to path p, and a value of 0 otherwise The level of the toll on link j, if a toll is levied The inverse demand function for trips for OD pair i J  δ jp · (c j (N j ) + f j ) − Di Ni )) ≤ 0 , and 0 otherwise A dummy equal to 1 if p ∈ Pi and

δ jp fj Di

δ ip

j=1

and where primes denote derivatives. For our tests we consider inelastic demand, therefore yielding Di  = 0 ∀i ∈ I. Applying our own notation, we can rewrite Eq. (27) as follows:



λp =

l∈ p

fl · dcl /d fl −

P  q=1,q= p



λq ·



dcl /d fl

l∈P∩Q

dcl /d fl

(28)

l∈ p

and solve it for λ (as a correctly determined system of linear equations) in order to determine the shadow cost for all links and their ranking, from which a quantity of controllers (chosen equal to that deemed necessary following Algorithm 3.2) is selected for the tests of Sections 5.2 and 5.3, following decreasing link ranks. Appendix B. LTV framework, congestion modelling and limitations In this Appendix we highlight through two simple theoretical examples the conditions under which the LTV model proposed in Eq. (11) can be considered exact, and those in which, conversely, the model fails to capture the network’s behavior. As discussed in Section 3, we assume traffic as obeying Newell’s Kinematic Wave Theory, and we furthermore consider the instance of triangular shaped fundamental diagram. To illustrate the LTV model’s capabilities in terms of congestion

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

401

modelling, we refer first to the simple example highlighted in Figs 11-12. The network in question is a straight stretch of road spanning a total distance of 6 km, the free flow speed vf is 60 km/h throughout the whole network, whereas capacity is 10 0 0 veh/h for the first four km but, due to some disruption, only 750 veh/h from that point on. Demand is 10 0 0 veh/h from time 0 [min] until time 10 [min], it then reduces to 500 veh/h from that point forward. The maximum backwards wave propagation speed w is 15 km/h. The two triangular diagrams of Fig. 11 express, from left to right respectively, the behavioral characteristics of the first 4 km of road and those of the remaining 2 km. In both Figures, capital letters highlight the specific flow/density statuses in which the different portions of road find themselves throughout time, the red dotted lines in Fig. 11 specifically remarking the transitional pathways between said statuses. Following KWT, as demand entering the network at km 0 reaches km 4, the capacity limitation triggers an increase in density and shifts the status of the road between km 3 and km 4 from free flow conditions (A) towards congestion (B), generating a backwards propagating wave. As time progresses, this wave spills back beyond km 3 and towards km 2. As demand decreases below the bottleneck’s capacity, a recovery wave is triggered at minute 12, and propagates forward until all congestion is cleared at minute 24. We model this network through Eq. (13) by considering four state variables, x1 ,x2 ,x3 ,x4 , placed at the corresponding kilometric markers along the road. As detailed earlier in Section 3, these variables represent the Cumulative Vehicle Number at each point in the road; since no diverging or merging intersection is present, route fractions for this example are all identical to 1. The full LTV model for this network can be written as follows:

x2 (k ) = a¯ 2,1 (k − 1 ) · x1 (k − 1 ) x3 (k ) = a¯ 3,2 (k − 1 ) · x2 (k − 1 ) x4 (k ) = a¯ 4,3 (k − 1 ) · x3 (k − 1 ) x1 ( 0 ) =  0 1 · 10 0 0 · x1 (k − 1 ), k ∈ [1, 10/Ts ] x1 (k ) = T1s · 500 · x1 (k − 1 ), k ∈ [11/Ts , +∞] Ts

(29)

State variable x1 is exogenously controlled, representing the demand pattern, whereas the others evolve according to the variable nature of the elements of matrix A¯ (k ). The chosen discretisation time period is Ts = 1 [min ].

Fig. 11. Fundamental Diagrams for the road stretch taken into consideration.

Fig. 12. Space-Time diagram of the example in question, showcasing congestion formation, propagation and dissipation.

402

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 13. Time-varying values for elements a¯ i, j .

Fig. 14. CVN dynamics.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

403

Fig. 15. Fundamental Diagram for the road stretch taken into consideration.

Fig. 16. Space-Time diagram of the example in question, showcasing congestion formation and propagation.

Given the simplistic nature of the example at hand, exact values for the travel times τ i.j (k) along the represented stretch of road can easily be computed, meaning that punctual values for elements a¯ i, j can be obtained exactly, for each time step k, as shown in Fig. 13. As time progresses, matrix A¯ (k ) correctly captures the fact that congestion arises firstly in the 3km-4 km stretch of the given road stretch (black line), and then propagates to the 2km-3 km stretch (red line). Once demand decreases, both values gradually return to 1. The example’s evolution in terms of Cumulative Vehicle Numbers (solid lines, representing physical evolution) and discrete status evolution (circled dots, representing LTV model’s evolution) are shown in Fig. 14. It’s clear that the model captures the physical evolution correctly throughout all stages of congestion formation, propagation and dissipation. For the sake of establishing controllability, as discussed earlier, values of the elements a¯ i, j should never become exactly zero. Congestion dynamics alone, while non-linear, will not always cause this effect (as clearly visible in the evolution of elements a¯ i, j shown in Fig. 13). This implies that controllability for this simple stretch of road can be computed considering a steady-state non-zero value for all elements a¯ i, j , with no loss of generality. Rather trivially, for a simple straight corridor, placing a single controller at any of the successive points will be sufficient to achieve full controllability. The following example showcases instead the specific situation and conditions in which model (11) is inapplicable, and thus the simplified LTI representation of Eq. (13) cannot be used to establish controllability. Given the same 6 km stretch of road, we now consider an instance where demand is constantly valued at 500 veh/h, but capacity at km 2 collapses to 0 veh/h after minute 5 due, for example, to a complete roadblock incident. Once again, the triangular fundamental diagram of Fig. 15 and the space-time diagram of Fig. 16 show the status in which different portions of the road stretch operate throughout time. In this instance, congestion begins forming from time 5 [min] and will spill back towards km 1 and km 0 gradually, without ever recovering until the roadblock is successfully removed. In these extreme conditions, the travel times experienced by vehicles in the congested status (B) are infinity, since vehicles in the queue are perfectly stationary due to lack of outflow. Elements a¯ 3,2 (k ), a¯ 4,3 (k ) become thus undefined after time k = 5 min, as shown in Fig. 17. Trivially, the LTV model of eq. (11) can no longer represent the system beyond this time, and thus the predicted evolution of state variables x3 ,x4 is simply missing, as shown in Fig. 18. In this instance, due to the extreme severity of congestion (complete roadblock), controllability cannot be established through our proposed approach. (Table 1 and Algorithm 3.1).

404

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Fig. 17. Time-varying values for elements a¯ i, j .

Fig. 18. CVN dynamics.

References Aboudolas, K., Papageorgiou, M., Kosmatopoulos, E., 2009. Store-and-forward based methods for the signal control problem in large-scale congested urban road networks. Transp. Res. Part C Emerg. Technol., Selected papers from the Sixth Triennial Symposium on Transportation Analysis (TRISTAN VI) 17, 163–174. doi:10.1016/j.trc.2008.10.002. Aboudolas, K., Papageorgiou, M., Kouvelas, A., Kosmatopoulos, E., 2010. A rolling-horizon quadratic-programming approach to the signal control problem in large-scale congested urban road networks. Transp. Res. Part C Emerg. Technol. 18, 680–694. doi:10.1016/j.trc.20 09.06.0 03.

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

405

Allsop, R.E., 1974. Some possibilities for using traffic control to influence trip distribution and route choice. In: Proceedings of the 6th International Symposium on Transportation and Traffic Theory held at the University of New South Wales, Sydney, Australia from 26-28 August 1974, 6, pp. 345–373. Bashirov, A.E., Mahmudov, N., S¸ emı, N., Etıkan, H., 2007. Partial controllability concepts. Int. J. Control 80, 1–7. doi:10.1080/0 020717060 0885489. Bekiaris-Liberis, N., Roncoli, C., Papageorgiou, M., 2017. Highway traffic state estimation per lane in the presence of connected vehicles. Transp. Res. Part B Methodol. 106, 1–28. doi:10.1016/j.trb.2017.11.001. Cantarella, G.E., Sforza, A., 1995. Network Design Models and Methods for Urban Traffic Management, in: Urban Traffic Networks, Transportation Analysis. Springer, Berlin, Heidelberg, pp. 123–153. doi:10.1007/978- 3- 642- 79641- 8_5. Castillo, E., Calviño, A., Lo, H.K., Menéndez, J.M., Grande, Z., 2014. Non-planar hole-generated networks and link flow observability based on link counters. Transp. Res. Part B Methodol. 68, 239–261. doi:10.1016/j.trb.2014.06.015. Castillo, E., Jimenez, P., Menendez, J.M., Conejo, A.J., 2008. The observability problem in traffic models: algebraic and topological methods. IEEE Trans. Intell. Transp. Syst. 9, 275–287. doi:10.1109/TITS.2008.922929. Castillo, E., Nogal, M., Rivas, A., Sánchez-Cambronero, S., 2013. Observability of traffic networks. Optimal location of counting and scanning devices. Transp. B Transp. Dyn. 1, 68–102. Contreras, S., Kachroo, P., Agarwal, S., 2016. Observability and sensor placement problem on highway segments: a traffic dynamics-based approach. IEEE Trans. Intell. Transp. Syst. 17, 848–858. doi:10.1109/TITS.2015.2491282. de Palma, A., Kilani, M., Lindsey, R., 2005. Comparison of second-best and third-best tolling schemes on a road network. Transp. Res. Rec. J. Transp. Res. Board 1932, 89–96. doi:10.3141/1932-11. Dotoli, M., Fanti, M.P., Meloni, C., 2006. A signal timing plan formulation for urban traffic control. Control Eng. Pract. 14, 1297–1311. doi:10.1016/j.conengprac. 2005.06.013. Evers, R., Proost, S., 2015. Optimizing intersections. Transp. Res. Part B Methodol. 71, 100–119. doi:10.1016/j.trb.2014.10.006. FGSV-Verlag, 2015. Handbuch für die Bemessung von Straßenverkehrsanlagen. Forschungsgesellschaft für Straßen- und Verkehrswesen e. V., Köln). Gartner, N.H., 1983. OPAC: a demand-responsive strategy for traffic signal control. Presented at the 62nd Annual Meeting of the Transportation Research Board. Transportation Research Board, pp. 75–81 ISSN: 0361-1981. Geroliminis, N., Daganzo, C.F., 2008. Existence of urban-scale macroscopic fundamental diagrams: Some experimental findings. Transp. Res. Part B Methodol. 42, 759–770. doi:10.1016/j.trb.20 08.02.0 02. Ghali, M.O., Smith, M.J., 1995. A model for the dynamic system optimum traffic assignment problem. Transp. Res. Part B Methodol. 29, 155–170. doi:10. 1016/0191-2615(94)0 0 024-T. Hahn, J., Edgar, T.F., Marquardt, W., 2003. Controllability and observability covariance matrices for the analysis and order reduction of stable nonlinear systems. J. Process Control 13, 115–127. doi:10.1016/S0959-1524(02)0 0 024-0. Hautus, M.L.J., Centre for Analysis, Scientific Computing, and Applications W&I, Research 1957/58 - 1978/79, 1969. Controllability and observability conditions of linear autonomous systems. Indag. Math. Proc. 72, 443–448. He, S., 2013. A graphical approach to identify sensor locations for link flow inference. Transp. Res. Part B Methodol. 51, 65–76. doi:10.1016/j.trb.2013.02.006. Hegyi, A., De Schutter, B., Hellendoorn, H., 2005. Model predictive control for optimal coordination of ramp metering and variable speed limits. Transp. Res. Part C Emerg. Technol. 13, 185–209. doi:10.1016/j.trc.20 04.08.0 01. Henry, J.J., Farges, J.L., Tuffal, J., 1984. The prodyn real time traffic algorithm. In: Proceedings of the 4th IFAC/IFIP/IFORS Conference, Baden-Baden, Federal Republic of Germany, 20–22 April 1983, pp. 305–310. doi:10.1016/B978- 0- 08- 029365- 3.50048- 1. Hoogendoorn, S., Landman, R., van Kooten, J., Taale, H., Schreuder, M., 2014. Integrated Network Management Amsterdam: Towards a Field Operational Test. TRB 93rd Annual Meeting Compendium of Papers Report/Paper Numbers: 14-2755. Hoogendoorn, S.P., Van Kooten, J., Adams, R., 2016. Lessons Learned from Field Operational Test of Integrated Network Management in Amsterdam. Transp. Res. Rec. J. Transp. Res. Board 2554, 111–119. doi:10.3141/2554-12. Hunt, P.B., Robertson, D.I., Bretherton, R.D., Winton, R.I., 1981. SCOOT - a traffic responsive method of coordinating signals. Transport and Road Research Laboratory (TRRL) ISSN: 0266-7045. Kalman, R.,E., Ho, Y.C., Narendra, S.K., 1963. Controllability of linear dynamical systems. Contrib. Differ. Equ. 1, 189 –23. Lin, C.-T., 1974. Structural controllability. IEEE Trans. Autom. Control 19, 201–208. doi:10.1109/TAC.1974.1100557. Little, J.D.C., Kelson, M.D., Gartner, N.H., 1981. MAXBAND: a versatile program for setting signals on arteries and triangular networks. Working papers 1185-81., Massachusetts Institute of Technology (MIT), Sloan School of Management. Liu, Y.-Y., Slotine, J.-J., Barabási, A.-L., 2011. Controllability of complex networks. Nature 473, 167–173. doi:10.1038/nature10011. Lowrie, P.R., 1990. Scats, sydney co-ordinated adaptive traffic system : a traffic responsive method of controlling urban traffic. Roads and Traffic Authority NSW, Darlinghurst, NSW Australia. Mauro, V., Taranto, C. di, 1990. UTOPIA. Selected Papers from the IFAC/IFIP/IFORS Symposium, Paris, France, 19-21 September 1989 IFAC Symposia Series 1990, pp. 245–252. doi:10.1016/B978- 0- 08- 037025- 5.50042- 6. Mora˘ rescu, I., Canudas-de-Wit, C., 2011. Highway traffic model-based density estimation, in: Proceedings of the 2011 American Control Conference. In: Presented at the Proceedings of the 2011 American Control Conference. IEEE, pp. 2012–2017. doi:10.1109/ACC.2011.5991298. Newell, G.F., 1993. A simplified theory of kinematic waves in highway traffic, part I: General theory. Transp. Res. Part B Methodol. 27, 281–287. doi:10.1016/ 0191-2615(93)90038-C. Ng, M., 2012. Synergistic sensor location for link flow inference without path enumeration: a node-based approach. Transp. Res. Part B Methodol. 46, 781–788. doi:10.1016/j.trb.2012.02.001. Olshevsky, A., 2014. Minimal Controllability Problems. IEEE Trans. Control Netw. Syst. 1, 249–258. doi:10.1109/TCNS.2014.2337974. Papageorgiou, M., Hadj-Salem, H., Blosseville, J.-M., 1991. Alinea: a local feedback control law for on-ramp metering. Transportation Research Board ISSN: 0361-1981. Rinaldi, M., Himpe, W., Tampère, C., 2016. A sensitivity based approach for adaptive decomposition of anticipatory network traffic control. Transportation Research Part C: Emerging Technologies 66, 150–175. doi:10.1016/j.trc.2016.01.005. Rinaldi, M., Tampère, C.M.J., 2015. An extended coordinate descent method for distributed anticipatory network traffic control. Transp. Res. Part B Methodol. 80, 107–131. doi:10.1016/j.trb.2015.06.017. Rinaldi, M., Tampere, C.M.J., Himpe, W., Holvoet, T., 2013. Centralized and decomposed anticipatory Model Predictive Control for network-wide Ramp Metering. In: 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013). IEEE, pp. 461–467. doi:10.1109/ITSC.2013.6728274. Rinaldi, M., Tampère, C.M.J., Viti, F., 2017. On characterizing the relationship between route choice behaviour and optimal traffic control solution space. Transp. Res. Part B Methodol. doi:10.1016/j.trb.2017.08.014. Rinaldi, M., Viti, F., 2017. Exact and approximate route set generation for resilient partial observability in sensor location problems. Transp. Res. Part B Methodol. 105, 86–119. doi:10.1016/j.trb.2017.08.007. Sen, S., Head, K.L., 1997. Controlled optimization of phases at an intersection. Transp. Sci. 31, 5–17. doi:10.1287/trsc.31.1.5. Sheffi, Y., 1985. Urban transportation networks: equilibrium analysis with mathematical programming methods. Prentice-Hall, Incorporated, Englewood Cliffs, NJ United States. Shepherd, S., Sumalee, A., 2004. A genetic algorithm based approach to optimal toll level and location problems. Netw. Spat. Econ. 4, 161–179. doi:10.1023/B: NETS.0 0 0 0 027771.13826.3a. Summers, T.H., Cortesi, F.L., Lygeros, J., 2016. On submodularity and controllability in complex dynamical networks. IEEE Trans. Control Netw. Syst. 3 (1). doi:10.1109/TCNS.2015.2453711. Summers, T.H., Cortesi, F.L., Lygeros, J., 2014. On Submodularity and Controllability in Complex Dynamical Networks. ArXiv14047665 Cs Math. Taale, H., 2008. Integrated Anticipatory Control of Road Networks. A Game-Theoretical Approach. Netherlands TRAIL Research School, Netherlands.

406

M. Rinaldi / Transportation Research Part B 118 (2018) 381–406

Taale, H., Hoogendoorn, S., 2012. Network-wide Traffic Management with Integrated Anticipatory Control. TRB 91st Annual Meeting Compendium of Papers. Transportation Research Board, NW Washington, DC United States Report/Paper Numbers: 12-3654. Transportation Research Board, 2010. Highway Capacity Manual 2010. National Research Council, Washington, DC. Ukkusuri, S., Doan, K., Aziz, H.M.A., 2013. A Bi-level Formulation for the Combined Dynamic Equilibrium based Traffic Signal Control. Procedia - Soc. Behav. Sci. 80, 729–752. doi:10.1016/j.sbspro.2013.05.039. Van De Weg, S.G., Keyvan-Ekbatani, M., Hegyi, A., Hoogendoorn, S.P., 2016. Urban Network Throughput Optimization via Model Predictive Control Using the Link Transmission Model. TRB 95th Annual Meeting Compendium of Papers. Transportation Research Board, Washington, DC United States Report/Paper Numbers: 16-1512. Verhoef, E.T., 2002a. Second-best congestion pricing in general static transportation networks with elastic demands. Reg. Sci. Urban Econ. 32, 281–310. Verhoef, E.T., 2002b. Second-best congestion pricing in general networks. Heuristic algorithms for finding second-best optimal toll levels and toll points. Transp. Res. Part B Methodol. 36, 707–729. Verhoef, E.T., Emmerink, R.H.M., Nijkamp, P., Rietveld, P., 1996. Information provision, flat and fine congestion tolling and the efficiency of road usage. Reg. Sci. Urban Econ. 26, 505–529. doi:10.1016/0166- 0462(96)02130- 8. Viti, F., Rinaldi, M., Corman, F., Tampère, C.M.J., 2014. Assessing partial observability in network sensor location problems. Transp. Res. Part B Methodol. 70, 65–89. doi:10.1016/j.trb.2014.08.002. Wallace, C.E., Courage, K.G., Reaves, D.P., Schoene, G.W., Euler, G.W., 1984. TRANSYT-7F USER’S MANUAL. University of Florida, Gainesville Transportation Research Center, Gainesville, FL United States/Federal Highway Administration Office of Safety and Traffic Operations, Washington, DC United States. Report/Paper Numbers: UF-TRC-U32 FP-06/07. Wang, W.-X., Ni, X., Lai, Y.-C., Grebogi, C., 2012. Optimizing controllability of complex networks by minimum structural perturbations. Phys. Rev. E 85 (026115). doi:10.1103/PhysRevE.85.026115. Webster, F.V., 1958. TRAFFIC SIGNAL SETTINGS. Road Research Lab Tech Papers, UK Report/Paper Numbers: No 39. Yan, H., Lam, W.H., 1996. Optimal road tolls under conditions of queueing and congestion. Transp. Res. Part Policy Pract. 30, 319–332. Yang, H., Bell, M.G., 1997. Traffic restraint, road pricing and network equilibrium. Transp. Res. Part B Methodol. 31, 303–314. Yang, H., Yagar, S., 1995. Traffic assignment and signal control in saturated road networks. Transp. Res. Part Policy Pract. 29, 125–139. Yen, J.Y., 1971. Finding the K Shortest Loopless Paths in a Network. Manag. Sci. 17, 712–716. doi:10.1287/mnsc.17.11.712. Yuan, Z., Zhao, C., Di, Z., Wang, W.-X., Lai, Y.-C., 2013. Exact controllability of complex networks. Nat. Commun. 4, 2447. doi:10.1038/ncomms3447. Zhao, C., Zeng, A., Jiang, R., Yuan, Z., Wang, W.-X., 2017. Controllability of flow-conservation networks. Phys. Rev. E 96 (012314). doi:10.1103/PhysRevE.96. 012314.