Optimal trac light control for a single intersection Bart De Schutter and Bart De Moor y ESAT-SISTA, KU Leuven, Kardinaal Mercierlaan 94, B-3 Leuven (

Size: px

Start display at page:

Download "Optimal trac light control for a single intersection Bart De Schutter and Bart De Moor y ESAT-SISTA, KU Leuven, Kardinaal Mercierlaan 94, B-3 Leuven ("

George Harrison
5 years ago
Views:

1 Katholieke Universiteit Leuven Departement Elektrotechniek ESAT-SISTA/TR 97- Optimal trac light control for a single intersection Bart De Schutter and Bart De Moor Proceedings of the 997 International Symposium on Nonlinear Theory and its Applications (NOLTA'97), Honolulu, Hawaii, pp 85{88, Nov 9{Dec 3, 997 Revised: December 997 This paper is also available by anonymous ftp from ftpesatkuleuvenacbe as le 97-psgz in the directory pub/sista/deschutter/reports ESAT/SISTA, KULeuven, Kardinaal Mercierlaan 94, B-3 Heverlee (Leuven), Belgium, tel , fax , URL: bartdeschutter@esatkuleuvenacbe, bartdemoor@esatkuleuvenacbe Bart De Schutter is a senior research assistant with the FWO (Fund for Scientic Research-Flanders) and Bart De Moor is a senior research associate with the FWO This research was sponsored by the Concerted Action Project of the Flemish Community, entitled \Model-based Information Processing Systems" and by the Belgian program on interuniversity attraction poles (IUAP P4- and IUAP P4-4), by the ALAPEDES project of the European Community Training and Mobility of Researchers Program, and by the European Commission Human Capital and Mobility Network SIMONET (\System Identication and Modelling Network")

2 Optimal trac light control for a single intersection Bart De Schutter and Bart De Moor y ESAT-SISTA, KU Leuven, Kardinaal Mercierlaan 94, B-3 Leuven (Heverlee), Belgium bartdeschutter@esatkuleuvenacbe, bartdemoor@esatkuleuvenacbe Abstract We consider a single intersection of two two-way streets with controllable trac lights on each corner We construct a model that describes the evolution of the queue lengths in each lane as a function of time, and we discuss how (sub)optimal PSfrag tracreplacements light switching schemes for this system can be determined I Introduction In this paper we study the optimal trac light control problem for a single intersection Given the arrival rates and the maximal departure rates of vehicles at the intersection we compute trac light switching schemes that minimize criteria such as average queue length, worst case queue length, average waiting time, : : :, thereby augmenting the ow of trac and diminishing the eects of trac congestion We show that for a special class of objective functions (ie, objective functions that depend strictly monotonously on the queue lengths at the trac light switching time instants), the optimal trac light switching scheme can be computed very eciently Moreover, if the objective function is linear, the problem reduces to a linear programming problem II The set-up and the model of the system We consider an intersection of two two-way streets (see Figure ) There are four lanes L, L, L 3 and L 4, and on each corner of the intersection there is a trac light (T, T, T 3 and T 4 ) For sake of simplicity we assume that the trac lights can be either green or red The average arrival rate of cars in lane L i is i When the trac light is green, the average departure rate in lane L i is i Let t, t, t, : : : be the time instants at which the trac lights switch from green to red or vice versa The trac light switching scheme is shown in Table Let l i (t) be the queue length (ie, the number of cars waiting) in lane L i at time instant t Normally, l i (t) can only take on integer values, but we use continuous approximations for the queue lengths Note that although in general this will not lead to an optimal trac light switching scheme, the trac light Senior Research Assistant with the FWO (Fund for Scientic Research { Flanders) y Senior Research Associate with the FWO T T L T 3 L Figure : A trac light controlled intersection of two two-way streets L 4 L 3 T 4 Period T T T 3 T 4 t { t red green red green t { t green red green red t { t 3 red green red green Table : The trac light switching scheme switching scheme that we shall obtain will be a good approximation to the optimal scheme Furthermore, in practice there will also be uncertainty due to inaccurate measurements and the time-varying nature of the arrival and departure rates Dene k = t k+? t k for k N Let us now write down the equations that describe the relation between the switching time instants and the queue lengths When the trac light T is red, there are arrivals at lane L and no departures Hence, d l (t) = dt for t (t k ; t k+ ) with k N, and l (t k+ ) = l (t k ) + k 85

3 for k = ; ; ; : : : When the trac light T is green, there are arrivals and departures at lane L Since the net arrival rate is? and since the queue length l (t) cannot be negative, we have: ( d l (t)? if l (t) > = dt if l (t) = for t (t k+ ; t k+ ) with k N So l (t k+ ) = max? l (t k+ ) + (? ) k+ ; for k = ; ; ; : : : Note that we also have l (t k+ ) = max? l (t k ) + k ; for k = ; ; ; : : :, since l (t) > for all t We can write down similar equations for l (t k ), l 3 (t k ) and l 4 (t k ) So if we dene x k = l (t k ) l (t k ) l 3 (t k ) l 4 (t k ) T then we have b =? 3 4? 4 T b =? 3? 3 4 T ; x k+ = max (x k + b k ; ) () x k+ = max (x k+ + b k+ ; ) () for k = ; ; ; : : : The model we have derived above is dierent from the models used by most other researchers due to the fact that we consider red-green cycle lengths that may vary from cycle to cycle Furthermore, we consider non-saturated intersections, ie, we allow queue lengths to become equal to during the green cycle For more information on other models that describe the evolution of the queue lengths at a trac-lightcontrolled intersection and on optimal trac light control the interested reader is referred to [4, 5, 6, 7, 9, ] and the references given therein III Optimal trac light control From now on we assume that the arrival and departure rates are known We want to compute an optimal sequence t, t, : : : t N of switching time instants that minimizes a given criterion J Possible objective functions are: (weighted) average queue length over all queues: J = i= t N l i (t) dt ; (3) (weighted) average queue length over the worst queue: Z tn J = max l i (t) dt ; (4) (weighted) worst case queue length: J 3 = max i; t? wi l i (t) ; (5) (weighted) average waiting time over all queues: J 4 = i= l i (t) dt ; (6) (weighted) average waiting time over the worst queue: J 5 = max i l i (t) dt ; (7) where > for all i Note that J and J 4 are in fact equivalent in the sense that for any weight vector w for J there exists a weight vector ~w for J 4 such that J and J 4 are equal This also holds for J and J 5 Dene the following sets: N? (N) = (N) = ; ; : : : ; ; ; : : : ; N? We can impose some extra conditions such as minimum and maximum durations for the red and green time, maximum queue lengths, and so on This leads to the following problem: minimize J (8) min;r 6 k 6 max;r for k (N) (9) min;g 6 k+ 6 max;g for k (N) () x k 6 x max for k = ; : : : ; N () x k+ = max(x k + b k ; ) for k (N) () x k+ = max(x k+ + b k+ ; ) for k (N) : (3) Now we discuss some methods to solve problem (8) { (3) Consider () for an arbitrary index k It is easy to verify that this equation is equivalent to: i= x k+? x k? b k > x k+ > (x k+? x k? b k ) i (x k+ ) i = (See also [3]) We can repeat this reasoning for (3) and for each k : 86

4 If we dene x = 6 4 x x N nally get a problem of the form and = 4 N? 3 7 5, we minimize J (4) Ax + B + c > (5) x > (6) Ex + D + f > (7) (Ax + B + c) T x = : (8) The system (5) { (8) is a special case of an Extended Linear Complementarity Problem (ELCP) [, ] In [, ] we have developed an algorithm to compute a parametric description of the complete solution set of an ELCP Once this description has been obtained, we can determine the values of the parameters for which J reaches a global minimum over the solution set of the ELCP It can be shown that J 3 is a convex function of the parameters [3] Furthermore, our computational experiments have shown that the determination of the minimum value of J, J, J 4 or J 5 over the solution set of the ELCP is a well-behaved problem in the sense that using a local minimization routine (that uses, eg, sequential quadratic programming) starting from different initial points always yields the same numerical result (within a certain tolerance) However, the general ELCP is an NP-hard problem [, ] Furthermore, the size of the parametric description of the solution set of an ELCP with p (nonredundant) inequality constraints in an n-dimensional space is O p b n c if p n This implies that the approach sketched above is not feasible if the number of switching cycles N is large In [3] we have developed a method to compute suboptimal solutions of problem (8) { (3) by considering several ELCPs of a smaller size Note that if there is no upper bound on x k, then problem (8) { (3) reduces to a constrained optimization problem in, which could be solved using a constrained minimization procedure The major disadvantage of this approach is that in general the resulting objective function is neither convex nor concave so that the minimization routine will only return a local minimum However, in [3] we have shown that J 3 is convex as a function of This implies that problem (8) { (3) with J = J 3 can be solved eciently (if there is no upper bound on the queue lengths or if we introduce a convex penalty term if one or more components of x max are nite) Now we dene some objective functions that can be considered as approximations to the objective functions J i (i = ; : : : ; 5) Although these approximations will lead to suboptimal solutions with respect to J i, the corresponding optimization problems can be solved much more eciently y Note that the objective functions J i do not depend explicitly on x since for given x, i 's and i 's, the components of x (ie, the x k 's) are uniquely determined by The approximate objective functions that we introduce now will depend explicitly on x and For a given x, x and we dene the function ~l i as the piecewise-linear function that interpolates in the points (t k,(x k ) i ) for k = ; : : : ; N The approximate objective functions Ji ~ (i = ; : : : ; 5) are dened as in (3) { (7) but with l i replaced by ~ l i We say that the vectors x and are compatible for a given x if the corresponding sequences fx k g N k= and f k g N? k= satisfy the recurrence equations () and () for all k Proposition If x and are compatible for a given x then we have J 3 () = ~ J3 (x; ) and J l () 6 ~ Jl (x; ) for l = ; ; 4; 5 Proof : See [3] Let the queue length vector x and the switching interval vector be compatible for a given x Note that if the queue lengths never stay in a time interval then J l ( ) and ~ Jl (x ; ) are equal In practice, an optimal trac light switching scheme implies the absence of long periods in which no cars wait in one lane while in the other lanes the queue lengths increase, ie, the periods during which the queue length in some lane is equal to are in general short for optimal trac light switching schemes This implies that for trac light switching schemes in the neighborhood of the optimal scheme ~ Jl will be a good approximation of J l Now we show that the use of the approximate objective function ~ J or ~ J4 leads to an optimization problem that can be solved more eciently than the original problem in which J or J 4 is used Let P be the problem (8) { (3) We dene the \relaxed" problem ~ P corresponding to the original problem P as follows: minimize ~ J (9) min;r 6 k 6 max;r for k (N) () y Recall that we have already introduced an approximation by considering continuous queue lengths Furthermore, there is also some uncertainty and variation in time of the arrival and departure rates, which makes that in general computing the exact optimal trac light switching scheme is not feasible Moreover, in practice we are more interested in quickly obtaining a good approximation of the optimal trac light switching scheme than in spending a large amount of time to obtain the exact optimal switching scheme 87

5 min;g 6 k+ 6 max;g for k (N) () 6 x k 6 x max for k = ; : : : ; N; () x k+ > x k + b k for k (N) (3) x k+ > x k+ + b k+ for k (N) : (4) Note that compared to the original problem we have replaced () { (3) by the relaxed equations 6 x k, (3) and (4), ie, we do not take the maximum condition into account Proposition If J is a strictly monotonous function of x ie, if for any with positive components and for all ~x ; ^x with 6 ~x 6 ^x and with ~x j < ^x j for at least one index j, we have J(~x ; ) < J(^x ; ) then any solution of the relaxed problem ~ P is also a solution of the original problem P Proof : See [3] In [3] we have shown that ~ J and ~ J4 are strictly monotonous functions of x, ie, they satisfy the conditions of Proposition Note that in general it is easier to solve the relaxed problem ~ P since the set of feasible solutions of ~ P is a convex set, whereas the set of feasible solutions of P is in general not convex since it consists of the union of a subset of the set of faces of the polyhedron dened by the system of inequalities () { (4) (See [, ]) In [3] we have also shown that if all the k 's are equal then ~ J and ~ J4 are linear, strictly monotonous functions of x, which implies that problem (8) { (3) then reduces to a linear programming problem, which can be solved eciently using (variants of the) simplex method or using an interior point method (see, eg, [8]) For a worked-out example of the computation of (sub)optimal trac light schemes for the set-up that has been considered in this paper the interested reader is referred to [3] IV Conclusions and further research We have derived a model that describes the evolution of the queue lengths at an intersection of two two-way streets with controllable trac lights on each corner We have indicated how an optimal trac light switching scheme for the given system can be determined We have shown that for objective functions that depend strictly monotonously on the queue lengths at the trac light switching time instants, the optimal trac light switching scheme can be computed very eciently We have discussed some approximate objective functions for which this property holds Moreover, if the objective function is linear, the problem reduces to a linear programming problem Topics for further research include: inclusion of an all-red phase and/or an amber phase, extension of the results to trac networks or clusters of intersections, extension to models with integer queue lengths, development of ecient algorithms to compute optimal trac light switching schemes for the objective functions discussed in this paper and for other more general objective functions Acknowledgment This research was sponsored by the Concerted Action Project of the Flemish Community, entitled \Model-based Information Processing Systems" (GOA-MIPS), by the Belgian program on interuniversity attraction poles (IUAP P4- and IUAP P4-4), by the ALAPEDES project of the European Community Training and Mobility of Researchers Program, and by the European Commission Human Capital and Mobility Network SIMONET (\System Identication and Modelling Network") References [] B De Schutter, Max-Algebraic System Theory for Discrete Event Systems PhD thesis, Faculty of Applied Sciences, KULeuven, Leuven, Belgium, 996 [] B De Schutter and B De Moor, \The extended linear complementarity problem," Mathematical Programming, vol 7, no 3, pp 89{35, Dec 995 [3] B De Schutter and B De Moor, \Optimal trac signal control for a single intersection," Tech rep 96-9, ESAT-SISTA, KULeuven, Leuven, Belgium, Dec 996 Submitted for publication [4] NH Gartner, JDC Little, and H Gabbay, \Simultaneous optimization of osets, splits, and cycle time," Transportation Research Record, vol 596, pp 6{5, 976 [5] HR Kashani and GN Saridis, \Intelligent control for urban trac systems," Automatica, vol 9, no, pp 9{97, Mar 983 [6] JH Lim, SH Hwang, IH Suh, and Z Bien, \Hierarchical optimal control of oversaturated urban trac networks," International Journal of Control, vol 33, no 4, pp 77{737, Apr 98 [7] D Lin, F Ulrich, LL Kinney, and KSP Kumar, \Hierarchical techniques in trac control," in Proceedings of the IFAC/IFIP/IFORS 3rd International Symposium, pp 63{7, 976 [8] Y Nesterov and A Nemirovskii, Interior-Point Polynomial Algorithms in Convex Programming Philadelphia: SIAM, 994 [9] ES Park, JH Lim, IH Suh, and Z Bien, \Hierarchical optimal control of urban trac networks," International Journal of Control, vol 4, no 4, pp 83{ 89, Oct 984 [] MG Singh and H Tamura, \Modelling and hierarchical optimization for oversaturated urban road trac networks," International Journal of Control, vol, no 6, pp 93{934,

Optimal traffic light control for a single intersection

KULeuven Department of Electrical Engineering (ESAT) SISTA Technical report 97- Optimal traffic light control for a single intersection B De Schutter and B De Moor If you want to cite this report, please