Traffic Effective Route Guidance in Traffic Networks Forecast: Number of cars will increase further Fact: Infrastructure will not be enhanced to the same extent Remedy: Improve the efficiency of traffic by other means Lectures developed by Andreas S. Schulz and Nicolás Stier May 12, 2003 c 2003 Massachusetts Institute of Technology � c 2003 Massachusetts Institute of Technology 3 2002 Urban Mobility Study Outline (http://mobility.tamu.edu/ums) • Lecture 1 “The bad news is that even if transportation officials do all the right things, the likely effect is that congestion will continue to grow. . . ” Route Guidance; User Equilibrium; System Optimum; User Equilibria in Networks with Capacities. • Total congestion “bill” in 2000 was $67.5 billion • Lecture 2 (= 3.6 billion hours delay + 5.7 billion gallons gas) time penalty for peak period travelers 1982 2000 16 hours 62 hours c 2003 Massachusetts Institute of Technology � Constrained System Optimum; Dantzig-Wolfe Decomposition; Constrained Shortest Paths; Computational Results. 4 c 2003 Massachusetts Institute of Technology � Problem 1 The Context • Olaf Jahn (Research Assistant). • Rolf H. Möhring (Principal Investigator). People travel (between 6% and 19%) too much because Collaboration with and support by DaimlerChrysler, Berlin. of an unfavorable selection of their route. (Beccaria & Bolelli 1992, Lösch 1995) • Nicolas Stier (Research Assistant). • Andreas S. Schulz (Principal Investigator). Supported by General Motors Innovation Grant and SMA. c 2003 Massachusetts Institute of Technology � 5 c 2003 Massachusetts Institute of Technology � 2 Shortest Path Routing Potential Remedies • Toll systems • Dynamic traffic signal control • Park and Ride • Traveller information systems Improved network performance, but . . . (Kaufman et al. 1991, Lee 1994) c 2003 Massachusetts Institute of Technology � Route Guidance 9 c 2003 Massachusetts Institute of Technology � 6 Shortest Path Routing II Route Guidance . . . the same simulations show the performance decreases as soon as many cars use the system. c 2003 Massachusetts Institute of Technology � Route Guidance 10 Proposed Solutions c 2003 Massachusetts Institute of Technology � Route Guidance 7 In-Car Navigation Systems • Multiple path routing: – k shortest paths – random perturbation • Feedback control: – iterative computation of shortest paths • Traffic assignment: – user equilibrium – system optimum – a new approach c 2003 Massachusetts Institute of Technology � Route Guidance 11 c 2003 Massachusetts Institute of Technology � Route Guidance 8 Modeling Assumptions Reality Our Model • microscopic → individual vehicles → exact position, speed • macroscopic → one abstract measure → traffic flow • dynamic → consider time → on a single point at any time • static → time independent → simultaneously at any point of the path • on-line → additional input over time • off-line → all data known in advance c 2003 Massachusetts Institute of Technology � The Traffic Model selfish users central planner the goal optimize own travel time optimize system welfare fair, not efficient efficient, not fair fair, efficient 15 Representation of the Road Network c 2003 Massachusetts Institute of Technology � 12 Route Guidance How much can one gain ? • Study worst-case ratios between guided / unguided traffic • Without guidance: use game theory to predict traffic (Wardrop 1952) • Users’ behavior modeled as user equilibrium (Nash eq.) • Price of anarchy is a measure of user equilibrium performance (Papadimitriou 2001) c 2003 Massachusetts Institute of Technology � The Traffic Model 16 c 2003 Massachusetts Institute of Technology � Route Guidance 13 Flows • Different drivers have different origins and destinations The Traffic Model • Flows on paths: fP is the traffic along path P • Flow on arcs: fa = � fP P :a∈P c 2003 Massachusetts Institute of Technology � The Traffic Model 17 c 2003 Massachusetts Institute of Technology � The Traffic Model 14 The Traffic Model Traversal Time Functions ta(fa) • Directed graph G = (V, A), k demands (oi, di) with rate ri 2 2 0 • Flows on paths fP . Can be non-integral. • Traversal times: latency functions ta(·) → continuous and nondecreasing → belong to a given set L (e.g. linear) 0 • The total travel time of a flow is: � ta(fa)fa C(f ) := x+1 1 2x 2 0 0 2 • Traversal time of an arc a depends on the flow fa on a • Dependence captured by the function ta(fa) 2 2 · 4 + 2 · 1 = 10 • Travel time along path P is denoted by tP (f ) = a∈A c 2003 Massachusetts Institute of Technology � fa 0 2 19 The Traffic Model c 2003 Massachusetts Institute of Technology � System Optimum � a∈P ta(fa) 18 The Traffic Model The Traffic Model Convex Multicommodity Min-Cost Flow Problem � min • Directed graph G = (V, A), k demands (oi, di) with rate ri ta(fa)fa 2 2 a∈A � s.t. for all a ∈ A fP = ri for all i = 1, . . . , k fP ≥ 0 for all P ∈ P P a � • Traversal times: latency functions ta(·) → continuous and nondecreasing → belong to a given set L (e.g. linear) P ∈Pi c 2003 Massachusetts Institute of Technology � 20 The Traffic Model 0 2 c 2003 Massachusetts Institute of Technology � 2 19 The Traffic Model (Pigou 1920) • Directed graph G = (V, A), k demands (oi, di) with rate ri 1 1 =min fb2 + 1 − fb = 3/4 s.t. 0 ≤ fb ≤ 1 2 2 0 • Flows on paths fP . Can be non-integral. • Traversal times: latency functions ta(·) → continuous and nondecreasing → belong to a given set L (e.g. linear) x and fa = fb = 1 2 1 2 0 x+1 1 2x • The total travel time of a flow is: � ta(fa)fa C(f ) := fa, fb ≥ 0 c 2003 Massachusetts Institute of Technology � 0 The Traffic Model 1 s.t. fa + fb = 1 x+1 a∈A Example of SO SO =min fa + fb2 0 1 2x • The total travel time of a flow is: � ta(fa)fa C(f ) := Pi : set of paths from oi to di P = ∪Pi where 0 • Flows on paths fP . Can be non-integral. fP = fa 2 2 0 0 2 2 2·1+2·3=8 a∈A The Traffic Model 21 c 2003 Massachusetts Institute of Technology � The Traffic Model 19 Braess Paradox • UE non-monotone with network improvements Example of SO 1 (Braess 1968) (Pigou 1920) 1/2 1 1 x 1 1 1 1 t 1/2 c x x SO =min fa + fb2 s.t. fa + fb = 1 =min fb2 + 1 − fb = 3/4 s.t. 0 ≤ fb ≤ 1 and fa = fb = 1 2 1 2 fa, fb ≥ 0 c 2003 Massachusetts Institute of Technology � The Traffic Model 23 c 2003 Massachusetts Institute of Technology � Braess Paradox • UE non-monotone with network improvements 1 1/2 1 1 User Equilibrium (Braess 1968) Definition : A flow is a UE iff nobody can switch to a path with smaller travel time. • Travel times of users between the same OD-pair are equal 1 x 21 The Traffic Model 1 1/2 x 3 1 C(f ) = 2 • UE always exists and is unique (Beckmann et al. 1956) 1 1 1 x c 2003 Massachusetts Institute of Technology � The Traffic Model 23 c 2003 Massachusetts Institute of Technology � Braess Paradox • UE non-monotone with network improvements 1 1/2 1 1 (Braess 1968) 3 1 C(f ) = 2 Definition : A flow is a UE iff nobody can switch to a path with smaller travel time. • UE always exists and is unique 1 1 1 1 x c 2003 Massachusetts Institute of Technology � (Beckmann et al. 1956) 1 1 0 1 1 1/2 x x 1 User Equilibrium • Travel times of users between the same OD-pair are equal 1 x 22 The Traffic Model x The Traffic Model 23 c 2003 Massachusetts Institute of Technology � The Traffic Model 22 Networks with Capacities Braess Paradox • UE non-monotone with network improvements • Latencies model capacity only implicitly 2 What is the impact of having explicit capacities on arcs? 0 1 x 0 1 0 cap=2 x Mx • Introduce capacities as hard constraints 0 • Straightforward to define SO 1 c 2003 Massachusetts Institute of Technology � 26 Networks with Capacities Equilibria in Networks with Capacities 1/2 1 0 1 1 1 1 c 2003 Massachusetts Institute of Technology � Mx • How good is the best / worst eq. ? (Papadimitriou 2001) Price of Anarchy γ := max 2 inst. • In general, γ unbounded 0 0 0 cap =2 0 0 2 c 2003 Massachusetts Institute of Technology � 23 The Traffic Model Price of Anarchy measures impact of lack of Central Coordination 2 • There may be multiple equilibria (UE w/o cap. was unique) 1 C(f ) = 2 x Definition : A flow is a capacitated UE iff nobody can switch to a path with smaller travel time and residual capacity • Travel times for users of same demand may differ (were constant w/o cap.) 3 1 C(f ) = 2 1 x x 2 1/2 1 1 0 2 • What is now a UE? (Braess 1968) 2 C(UE) C(SO) (Roughgarden & Tardos 2000) • If latencies are in L, γ ≤ α(L), where α(L) depends only on L x In particular, α({linear latencies}) = 4/3 (Roughgarden & Tardos 2000) (Roughgarden 2002) • Pigou’s and Braess’ examples are worst possible 2 27 Networks with Capacities c 2003 Massachusetts Institute of Technology � The Traffic Model 24 Multiple Equilibria 2 2 2 0 0 2 2 2 2 0 0 Mx 2 2 x cap=2 2 2 2 1+M 2 2M 1+M 2 1+M 2M 1+M Networks with Capacities 0 2 2 2 2 4 with costs of: 2 2 4M 2 2 4 M 1+M Worst UE can be unbounded! c 2003 Massachusetts Institute of Technology � Networks with Capacities 28 c 2003 Massachusetts Institute of Technology � Networks with Capacities 25 Network with Capacities: Guarantees Convex Optimization Review Theorem. For any instance of a network with capacities with latencies in L, we have • Let z be a continuously differentiable and convex function on a convex set. C(BUE) ≤ α(L) C(SO) • Then x∗ is a global minimum of z iff In particular, if latencies are linear, C(BUE) ≤ 43 C(SO) the gradient along all feasible directions is non-negative Guarantee does not change with introduction of capacities c 2003 Massachusetts Institute of Technology � Networks with Capacities 32 c 2003 Massachusetts Institute of Technology � Beckmann UE Proof (for the Linear Case) • Assume ta(fa) = qafa + ra for all a and let f = BUE � • Let C f (x) = a xa ta(fa) • Space of UE non-convex: Difficult to get Best UE • Instead, Beckmann UE (BUE) is • For all flows x: C(f ) ≤ C f (x) � (same as a(xa − fa)ta(fa) ≥ 0, the condition for ) � � � • C f (x) = a xa(qafa + ra) ≤ a xa(qaxa + ra) + 14 a fa2 qa min c 2003 Massachusetts Institute of Technology � � 0 fa ta(x)dx subject to f feasible flow capacity constraints because (x − f /2) ≥ 0 � � • Last, a xa(qaxa + ra) + 14 a fa2 qa ≤ C(x) + 14 C(f ) ⇒ � a∈A 2 3 4 • Opt. Cond. BUE: g feasible direction ⇒ � a gata(fa) 33 c 2003 Massachusetts Institute of Technology � Networks with Capacities • Assume ta(fa) = qafa + ra for all a and let f = BUE � • Let C f (x) = a xa ta(fa) Lemma. f is a BUE ⇒ f is a UE • For all flows x: C(f ) ≤ C f (x) � (same as a(xa − fa)ta(fa) ≥ 0, the condition for ) � � � • C f (x) = a xa(qafa + ra) ≤ a xa(qaxa + ra) + 14 a fa2 qa • Suppose f is not a UE ⇒ ∃ two paths P , Q s.t. flow can be re-routed from P to Q and tP (f ) > tQ(f ) Proof : • P and Q define a circulation g which is a feasible direction because (x − f /2)2 ≥ 0 � � • Last, a xa(qaxa + ra) + 14 a fa2 qa ≤ C(x) + 14 C(f ) c 2003 Massachusetts Institute of Technology � • � a gata(fa) <0 ⇒ f is not a BUE But BUE is not necessarily the best equilibrium C(f ) ≤ C(x) Networks with Capacities 30 Beckmann UE is an Equilibrium Proof (for the Linear Case) 3 4 ≥0 C(f ) ≤ C(x) Networks with Capacities ⇒ 29 Networks with Capacities 33 c 2003 Massachusetts Institute of Technology � Networks with Capacities 31 Non-convexity of UE 2 Proof (for the Linear Case) 2 • Assume ta(fa) = qafa + ra for all a and let f = BUE � • Let C f (x) = a xa ta(fa) w 0 1 1 2 0 SO cap =2 x 1 0 because (x − f /2)2 ≥ 0 � � • Last, a xa(qaxa + ra) + 14 a fa2 qa ≤ C(x) + 14 C(f ) feasible flows 0 0 1 0 2 UE BUE w z • For all flows x: C(f ) ≤ C f (x) � (same as a(xa − fa)ta(fa) ≥ 0, the condition for BUE) � � � • C f (x) = a xa(qafa + ra) ≤ a xa(qaxa + ra) + 14 a fa2 qa 2 z ⇒ 2 c 2003 Massachusetts Institute of Technology � Networks with Capacities 34 Best vs. Beckmann 0 C(BUE) = 7 0 1 x+1 2 and c 2003 Massachusetts Institute of Technology � cap =1 Networks with Capacities • For all flows x: C(f ) ≤ C f (x) � (same as a(xa − fa)ta(fa) ≥ 0, the condition for BUE) � � � • C f (x) = a xa(qafa + ra) ≤ a xa(qaxa + ra) + 14 a fa2 qa x because (x − f /2)2 ≥ 0 � � • Last, a xa(qaxa + ra) + 14 a fa2 qa ≤ C(x) + 14 C(f ) 0 2 C(best UE) = C(SO) = 6.875 Networks with Capacities ⇒ 35 Review No capacities With capacities UE unique Set of UE may be non-convex UE/SO ≥ α(L) UE/SO unbounded UE/SO ≤ α(L) BUE/SO ≤ α(L) Networks with Capacities 3 4 c 2003 Massachusetts Institute of Technology � C(f ) ≤ C(x) Networks with Capacities 33 Proof (for the Linear Case) • Assume ta(fa) = qafa + ra for all a and let f = BUE � • Let C f (x) = a xa ta(fa) • For all flows x: C(f ) ≤ C f (x) � (same as a(xa − fa)ta(fa) ≥ 0, the condition for BUE) � � � • C f (x) = a xa(qafa + ra) ≤ a xa(qaxa + ra) + 14 a fa2 qa because (x − f /2)2 ≥ 0 � � • Last, a xa(qaxa + ra) + 14 a fa2 qa ≤ C(x) + 14 C(f ) ⇒ c 2003 Massachusetts Institute of Technology � 33 • Assume ta(fa) = qafa + ra for all a and let f = BUE � • Let C f (x) = a xa ta(fa) 2 0 c 2003 Massachusetts Institute of Technology � C(f ) ≤ C(x) Proof (for the Linear Case) • The BUE does not need to be best UE: 2 3 4 36 c 2003 Massachusetts Institute of Technology � 3 4 C(f ) ≤ C(x) Networks with Capacities 33