Data driven model free adaptive iterative learning perimeter control for large-scale urban road networks
Introduction
Urban traffic signal control is widely used to alleviate congestion all over the world. In the past decades, a number of methods for urban traffic control systems have been developed, such as SCOOT (Robertson and Bretherton, 1991), SCATS (Lowrie, 1982), TUC (Diakaki et al., 2002), and max pressure (Varaiya, 2013). However, accuracy and computational burden challenge the possible successful application of these optimization and control methods due to the requirement of the detailed models of the controlled urban networks. Moreover, these controllers make decisions at a local level without considering network effects, which may lead to suboptimal solutions due to congestion propagation. Therefore, developing an aggregated model for large-scale road networks is essential for both efficiency and reliability reasons. The macroscopic fundamental diagram (MFD) of urban traffic networks attracts increasing interest in literature owing to its potential in enabling the development of low complexity network-level traffic models, and ultimately the design of real-time management systems for large-scale urban networks.
The idea of MFD was first proposed in Godfrey (1969), but its existence with dynamic features has not been verified until 2008 by the empirical data of Yokohama (Geroliminis and Daganzo, 2008). Thereafter there has been an increased interest in MFD as numerous contributions have been made including theoretical analyses of the MFD, region partitioning via clustering, network-wide traffic modeling and control system design via actuation over perimeter control. Nevertheless, the heterogeneous distribution of congestion and hysteresis phenomena (Geroliminis and Sun, 2011, Gayah and Daganzo, 2011, Mahmassani et al., 2013) may affect the presence of a well-defined MFD. To overcome these drawbacks, one alternative way is to partition the large heterogeneous network into several homogeneous sub-regions. Various clustering algorithms for static and dynamic configuration have been developed to network partitioning, such as (Ji and Geroliminis, 2012, Saeedmanesh and Geroliminis, 2016, Saeedmanesh and Geroliminis, 2017, Lopez et al., 2017, Casadei et al., 2018). Clustering provides the possibility to build multi-region MFD models that can be integrated in perimeter control.
Based on MFD-based dynamical models, traffic control schemes for the perimeter control (i.e., regulating a set of traffic lights on the boundary between two neighbouring regions) are proposed to improve urban traffic conditions. The application of perimeter control to single region networks can be found in Daganzo, 2007, Haddad, 2017a, Haddad and Shraiber, 2014, Keyvan-Ekbatani et al., 2012, Keyvan-Ekbatani et al., 2015a. In Keyvan-Ekbatani et al. (2012), a proportional-integral (PI) feedback controller is involved. These results are extended to the case with time-delays in Keyvan-Ekbatani et al. (2015a). Considering traffic system uncertainty, a robust PI perimeter controller is designed by transforming the plant into a linear parameter-varying model in Haddad and Shraiber (2014). In Haddad, 2017a, the optimal control laws are proposed for both coupled and decoupled perimeter control inputs. Perimeter control is also applied to two-region urban networks with MFD-based dynamics. The two-region perimeter control problem is first formulated in Haddad and Geroliminis (2012) and stability conditions for stable equilibrium are derived. Boundary conditions and controllability issues are investigated in Zhong et al. (2018). In Geroliminis et al., 2013, Haddad, 2017b, model predictive control (MPC) is used to design the perimeter controller for two-region systems.
There are also recent works focusing on multi-region MFD networks. In Aboudolas and Geroliminis (2013), a multivariable linear quadratic regulator (LQR) is designed by linearizing the model at an equilibrium point. A practical multiple concentric PI controller is proposed in Keyvan-Ekbatani et al. (2015b). In order to adapt to different traffic conditions, the controller in Aboudolas and Geroliminis (2013) is enhanced by online adaptive optimization in Kouvelas et al. (2017), which is the first study in the literature where gains are updated based on historical data and performance of the controller from day to day. MPC is also applied to multi-region networks in Ramezani et al., 2015, Fu et al., 2017, Zhou et al., 2017, Sirmatel and Geroliminis, 2018. In Haddad and Mirkin, 2016, Haddad and Mirkin, 2017, the centralized and distributed adaptive perimeter control approaches are proposed, respectively. In the recent work Lei et al. (2019), a decentralized estimation and decentralized MFAC method is applied to perimeter control to solve the strong-coupled interconnections between sub-regions. The other kinds of MFD-based modeling and perimeter control methods could also be found in: integration of agent-based modeling with MFD (Kim et al., 2018a), modeling of macroscopic flows considering link level capacity (Kim et al., 2018b), hybrid PI control (Ding et al., 2018), control via vehicle routing (Menelaou et al., 2017), optimal control (Aalipour et al., 2019), congestion pricing in a connected vehicle environment (Yang et al., 2019), uncertainty modeling (Gao and Gayah, 2018) and robust control (Zhong et al., 2017, Ampountolas et al., 2017, Mohajerpoor et al., 2019).
However, almost all of the aforementioned methods are model-based. Due to the inevitable unmodelled dynamics and unknown demand uncertainty, the mismatch between the model predictions and the plant is a potential source of problems for model-based perimeter control. For this reason, it is worth investigating perimeter control methods independent of the dynamical model. On the other hand, traffic flow patterns usually show similarity from a macroscopic point of view (Hou and Xu, 2003, Hou et al., 2008, Hou and Li, 2016), even if individual roads might have strong daily variations. In such circumstances, the performance of the controller would improve greatly by making full use of the repeatability and similarity of a traffic system.
Iterative learning control is an effective method to deal with repeated control processes. It was originally proposed in Arimoto et al. (1984), and has been extensively developed in the past 30 years with various learning controllers including proportional derivative (PD) type ILC (Saab, 2003), norm optimal ILC (Amann et al., 1996) and Lyapunov-based adaptive ILC (Tayebi, 2004). In Hou and Xu (2003), ILC is first applied to the traffic system to solve the density control problem of freeways. For more detailed reviews of ILC, readers are pleased to read the survey papers (Bristow et al., 2006, Xu, 2011). Although ILC is widely studied, some fundamental limitations still exist. Controller design of the norm optimal or Lyapunov-based adaptive ILC requires model dynamics, and the selection of the learning gain of PD-type ILC is nontrivial. Recently, data-driven model-free adaptive ILC (MFAILC) (Chi and Hou, 2007, Chi et al., 2015, Bu et al., 2019) has been proposed on a new type of dynamic linearization (DL) data modeling technique (Hou and Jin, 2013). Distinguished from the existing other linearization methods, such as Taylor’s, input-output linearization method, etc. (Hou et al., 2017), the DL data modeling technique does not linearize the system at a fixed point but builds an equivalent linearization data model at each operation point. Using only input and output (I/O) data, all the parameters of this dynamic linearization data model can be estimated and then the controller is designed. Considering different memory lengths of the considered system, the DL data model can be classified into three forms: the compact-form dynamic linearization (CFDL) data model, the partial-form dynamic linearization (PFDL) data model, and the full-form dynamic linearization (FFDL) data model (Hou and Jin, 2013). By virtue of the DL data modeling technique, two critical assumptions, identical initial condition and identical trajectory, to traditional ILC, can be relaxed in the MFAILC method. It is also worthwhile pointing out that MFAILC is a branch of model free adaptive control (MFAC), which has been developed into a systematic framework in the past 20 years (see Hou et al., 2017, Hou and Jin, 2013, Hou and Xiong, 2019, Liu and Yang, 2019, Wu et al., 2019 for some recent progress of MFAC). Until now, MFAC has been applied successfully to over 150 different fields (see Hou and Wang, 2013, Hou et al., 2017 for the survey of MFAC approaches). Since the urban traffic network has following outstanding features, that is, repetitive operation pattern is obvious, the traffic system modeling is difficult, and the traffic system data are easy to be obtained, these features motivate us using the MFAILC method in this paper to deal with the perimeter control problem.
The main contributions are as follows.
(1) The time domain DL description used in Lei et al. (2019) is extended to the multi-input and multi-output (MIMO) iterative form, which presents the relationship of multi-region urban traffic network dynamics between successive iterations.
(2) Based on the novel iterative DL description, a data-driven model free adaptive iterative learning perimeter control (MFAILPC) strategy is employed in the multi-region urban traffic network in the first time. In particular, the proposed method is compared with other several typical perimeter algorithms in a comprehensive way.
Three attractive properties of this method are as follows.
(1) The data-driven feature: the controller design and implementation are independent from the traffic plant model, i.e., the system modeling and regional demand profiles are not needed in the perimeter control. Instead, it only requires I/O data of the traffic network, which means complication due to plant-model mismatch is avoided. Note that the critical accumulation used in the algorithm is an input to the method, which is separated from the controller design and not related to the data-driven feature.
(2) The iterative learning feature: the controller obtains ”experience” from the historical data stored in the data base. In other words, the perimeter control performance will improve over iterations.
(3) The adaptive feature: by estimating the pseudo Jacobian matrix, the learning gain is updated iteratively, which implies the learning gain selection problem in the tradition ILC method disappears.
The rest of the paper is organized as follows. In Section 2, the urban traffic system dynamics is described using MFD. In Section 3, the MFAILC based perimeter control strategy is introduced in detail. In Section 4, simulation results compared with other typical perimeter control strategies are given. The main conclusions and future works are summarized in Section 5.
Section snippets
Dynamics for multi-region urban traffic network system
The aggregated models (Ramezani et al., 2015, Sirmatel and Geroliminis, 2018) for the urban network considering inflow demands and transfer flows are presented in the following part. Note that the traffic model here is only used to simulate traffic reality and generate traffic data, and not for the perimeter controller design.
Consider a network with heterogeneous distribution of congestion, consisting of R homogeneous regions, i.e., , each with a well-defined MFD (see Fig. 1 for
Iterative dynamic linearization for traffic dynamics
Define following collective vectors:where , and stand for the vectors contain all the accumulation, perimeter control input, and demand data at sampling instant k, respectively.
For controller design purposes, only the information at the current control time instant is required. To distinguish data from
Case studies
To test the performance of the proposed method, six cases with different types of uncertainties are simulated on a three-region network (as the one in Fig. 1(a)). The cases are built from the simple to the complex: Case 1 is a basic one without errors; In Case 2–4, each case considers one typical uncertain factor with different levels; Case 5–6 give results under more realistic situations. More details are summarized in Table 1. It is worth noting that, for the proposed MFAILPC method, the
Conclusions
In this paper, the perimeter control problem for a multi-region network is addressed by a novel data-driven method named MFAILPC. By dynamically linearizing the controlled plant at each operation point, the system has been transformed into an iterative CFDL data model between successive iterations. With the help of the I/O data, the PJM is estimated and used in the controller design. There are three important features of MFAILPC: First, only the I/O data of the system is required in the
Acknowledgment
The first and second authors would like to acknowledge the support by the National Natural Science Foundation of China (Grant # 61433002, 61833001) and the Beijing Natural Science Foundation (Grant # L161007). The third and fourth authors acknowledge the support by the ERC (European Research Council) Starting Grant METAFERW: Modelling and controlling traffic congestion and propagation in large-scale urban multimodal networks (Grant # 338205).
References (66)
- et al.
Perimeter and boundary flow control in multi-reservoir heterogeneous networks
Transp. Res. Part B
(2013) - et al.
Macroscopic modelling and robust control of bi-modal multi-region urban road networks
Transp. Res. Part B
(2017) - et al.
Aggregation and travel time calculation over large scale traffic networks: an empiric study on the grenoble city
Transp. Res. Part C
(2018) - et al.
Dual-stage optimal iterative learning control for nonlinear non-affine discrete-time systems
Acta Automatica Sinica
(2007) - et al.
A unified data-driven design framework of optimality-based generalized iterative learning control
Comput. Chem. Eng.
(2015) Urban gridlock: macroscopic modeling and mitigation approaches
Transp. Res. Part B
(2007)- et al.
A multivariable regulator approach to traffic-responsive network-wide signal control
Control Eng. Pract.
(2002) - et al.
Hierarchical perimeter control with guaranteed stability for dynamically coupled heterogeneous urban traffic
Transp. Res. Part C
(2017) - et al.
An analytical framework to model uncertainty in urban network dynamics using macroscopic fundamental diagrams
Transp. Res. Part B
(2018) - et al.
Clockwise hysteresis loops in the macroscopic fundamental diagram: An effect of network instability
Transp. Res. Part B
(2011)
Existence of urban-scale macroscopic fundamental diagrams: some experimental findings
Transp. Res. Part B
Properties of a well-defined macroscopic fundamental diagram for urban traffic
Transp. Res. Part B
Optimal coupled and decoupled perimeter control in one-region cities
Control Eng. Pract.
Optimal perimeter control synthesis for two urban regions with aggregate boundary queue dynamics
Transp. Res. Part B
On the stability of traffic perimeter control in two-region urban cities
Transp. Res. Part B
Coordinated distributed adaptive perimeter control for large-scale urban road networks
Transp. Res. Part C
Robust perimeter control design for an urban region
Transp. Res. Part B
From model-based control to data-driven control: survey, classification and perspective
Inf. Sci.
An iterative learning approach for density control of freeway traffic flow via ramp metering
Transp. Res. Part C
On the spatial partitioning of urban transportation networks
Transp. Res. Part B
Exploiting the fundamental diagram of urban networks for feedback-based gating
Transp. Res. Part B
Agent-based network transmission model using the properties of macroscopic fundamental diagram
Transp. Res. Part C
Investigating transfer flow between urban networks based on a macroscopic fundamental diagram
Transp. Res. Rec.
Enhancing model-based feedback perimeter control with data-driven online adaptive optimization
Transp. Res. Part B
Urban network gridlock: theory, characteristics, and dynamics
Transp. Res. Part C
Constrained model predictive control: stability and optimality
Automatica
Controlling road congestion via a low-complexity route reservation approach
Transp. Res. Part C
Dynamics of heterogeneity in urban networks: aggregated traffic modeling and hierarchical control
Transp. Res. Part B
Clustering of heterogeneous networks with directional flows based on “snake” similarities
Transp. Res. Part B
Dynamic clustering and propagation of congestion in heterogeneously congested urban traffic networks
Transp. Res. Part B
Adaptive iterative learning control for robot manipulators
Automatica
Max pressure control of a network of signalized intersections
Transp. Res. Part C
Nonlinear multi-model predictive control
IFAC Proc. Vol.
Cited by (76)
Cooperative traffic signal control through a counterfactual multi-agent deep actor critic approach
2024, Transportation Research Part C: Emerging TechnologiesDemonstration-guided deep reinforcement learning for coordinated ramp metering and perimeter control in large scale networks
2024, Transportation Research Part C: Emerging TechnologiesRegional route guidance with realistic compliance patterns: Application of deep reinforcement learning and MPC
2024, Transportation Research Part C: Emerging TechnologiesNonlinear model predictive control of large-scale urban road networks via average speed control
2023, Transportation Research Part C: Emerging TechnologiesLane-changing trajectory control strategy on fuel consumption in an iterative learning framework
2023, Expert Systems with Applications