Strategic Choice of Arrival Times in a Transient Queue Nahum Shimkin, Technion Joint work with Sandeep Juneja, TIFR Workshop on Optimization, Scheduling and Queues In Honor of Gideon Weiss The University of Haifa, June 2012 The Arrival-Time Problem We consider a (FCFS) queueing system that caters to a finite population of customers Service starts at a known time T0 Customers prefer to get served early (or before others), but dislike waiting in queue The customer's dilemma: When to arrive at the queue? • arrive early and (possibly) wait longer in queue, or • arrive later: wait less but finish late Nahum Shimkin Strategic Choice of Arrival Times 2 Some Examples Lining up … for tickets: e.g. for a football match or rock concert (early buyers get better seats) for a new product (new iPad, Harry Potter book) for lunch in a busy cafeteria Nahum Shimkin Strategic Choice of Arrival Times 3 The Queueing Model N customers, FCFS I.I.D. service demands (V j , j = 1,… , N ) , exponentially distributed with E (V j ) = µ −1 Service starts at T0 = 0 Each customer j can choose her arrival time t j ∈ (−∞, ∞) , possibly randomly, according to a CDF F j (t ) Nahum Shimkin Strategic Choice of Arrival Times 4 Cost Function We consider a linear cost function for each customer: C j = α j w j + β jτ j where • w j is j 's waiting time in the queue • τ j his service completion time Homogeneous population case: α j ≡ α , β j ≡ β Each customer wishes to minimize his own expected cost E (C j ) , by appropriately choosing his arrival time distribution F j (t ) Due to obvious dependence on arrival decision of others – we obtain a non-cooperative game problem, and consider the Nash Equilibrium Point (NEP) for this game. Nahum Shimkin Strategic Choice of Arrival Times 5 Related Literature Strategic arrivals in queues: Glazer & Hassin (1983): exponential system with opening time, no late-service cost. Similar models: Rapoport et. al. (2004), Hassin & Kleiner (2009), Haviv (2012) Lariviere & van Mieghem (2004): competition leads to Poisson arrivals. Transportation literature: Extensive work exists on the bottleneck model (Vickrey 1969), also known as the morning commute problem. This essentially comprises a fluid queueing model, with individual costs related to required arrival times. Nahum Shimkin Strategic Choice of Arrival Times 6 The Fluid Model Motivated by the large-population scenario ( N → ∞ ), we first consider a simplified fluid model of our stochastic queueing system. Consider a continuous population of (infinitesimal) customers, represented by the set [0, Λ ]. Deterministic service rate µ , starting at t = 0 . Thus, with no idleness, all customers may be served within T f = Λ µ Suppose there are I ≥ 1 customer classes, indexed by i , with respective masses {Λ i }, ∑ i Λ i = 1. The cost function for a class- i customer is Ci ( w,τ ) = α i w + βiτ ( w : waiting time in queue, τ : service completion time α i > 0, βi > 0 : class-dependent cost parameters) Nahum Shimkin Strategic Choice of Arrival Times 7 Fluid model: Equilibrium arrival profile Define mi = αi α i + βi , and suppose mi < mi +1 The equilibrium arrival density is then unique and has the following form (Jain-Juneja-S. 2011) Λ F '(t ) 3 µ m3 class 3 µ m3 class 1 Q (t ) µ m2 µ m1 T0 T1 T2 T3 = Λµ t Evidently: Those who do not mind waiting (smaller mi ) arrive earlier. Nahum Shimkin Strategic Choice of Arrival Times 8 Fluid Model: Single-Class Suppose I = 1 (homogeneous cost functions) Then the equilibrium arrival density is uniform: F '(t ) µ αα+ β T0 = − αβ T1 −α 0 T1 = cost C ( t ) β t Λ µ t The Price of Anarchy (PoA, ratio of social cost at equilibrium to optimal social cost) is exactly 2 Nahum Shimkin Strategic Choice of Arrival Times 9 Fluid Model: Single-Class Case (2) Note: While the aggregate arrival profile is unique, it may be composed in many different ways from individual arrival densities. Hence, there is no uniqueness on the individual level. aggregate t individual t Nahum Shimkin Strategic Choice of Arrival Times 10 Back to the Stochastic Queueing Model We return to the N -customer queueing model, with homogeneous costs (i.e., single-class). Main results: ♦ Existence and uniqueness of a (symmetric) NEP ♦ Characterization and computation of the NEP ♦ Convergence of to the fluid limit as N → ∞ Nahum Shimkin Strategic Choice of Arrival Times 11 Notation Strategy profile: F = {F j }Nj =1, F −i = {F j , j ≠ i} Q(t ) ≡ Q(t ; F ) , Q −i (t ) ≡ Q(t ; F −i ) Ci (t ) = E (α wi (t ) + βτ i (t )) : expected cost for i given that he arrives at t and the others follow F −i Queue size: As τ i (t ) = t + wi (t ) , we have Ci (t ) = (α + β ) E ( wi (t )) + β t Nahum Shimkin Strategic Choice of Arrival Times 12 Equilibrium Conditions Let T i denote the support of (the measure represented by) Fi (⋅) Then, at a Nash equilibrium, Ci (t ) = ci for t ∈ Ti , and Ci (t ) ≥ ci for t ∉ Ti Therefore, on the interior of T i , d C (t ) dt i Nahum Shimkin =0 Strategic Choice of Arrival Times 13 Equilibrium Conditions We next observe that Ci (t ) = (α + β ) E ( wi (t )) + β t = (α + β )( d dt E (Q −i (t )) = E (Q −i (t )) µ d dt + max{0, −t}) + β t E ( A−i (t ) − S −i (t )) = ∑ j ≠i F j' (t ) − µ1{t ≥0} P (Q −i (t ) > 0)) d C (t ) = 0, t ∈ T o is equivalent to So that dt i i 1 µ ' F ∑ j ≠i j (t ) = Nahum Shimkin α α +β − 1{t ≥0} P(Q −i (t ) = 0) , Strategic Choice of Arrival Times t ∈ Ti o 14 Main Results: Symmetry Proposition 1 Any NEP is symmetric, namely Fi = F ∀i [Established under mild technical conditions on { F }] i We next characterize the symmetric equilibrium distribution F . Nahum Shimkin Strategic Choice of Arrival Times 15 Existence, Uniqueness and Characterization Proposition 2 1. F (t ) admits a density F '(t ) , and is supported on a single interval F '( t ) [ta , tb ], with ta < 0 and tb > 0 µ 2. For ta ≤ t < 0 , F '(t ) = N −1 αα+ β , a constant ta tb 0 3. For 0 ≤ t < tb , F (t ) satisfies the differential relation N −1 F '(t ) µ = αα+ β − P{Q −i (t ; F ) = 0} (*) −i Here Q (t ; F ) is the queue size due to arrivals of the (other) N − 1 customers with arrival-time distribution F 4. F (tb ) = 1 (by definition), and F '(tb ) = 0 5. Properties 1-4 uniquely characterize the equilibrium dist. F Nahum Shimkin Strategic Choice of Arrival Times 16 t Note: Existence of solutions to the (functional) differential equation (*) follows by Lipschitz continuity of the RHS in F (tb ) = 1. Existence and uniqueness of the solution with the prescribed boundary conditions follows by monotonicity properties of the RHS. Nahum Shimkin Strategic Choice of Arrival Times 17 Random N Suppose that N , the number of arriving customers, is random 2 with known distribution p N = ( p N ( n), n ≥ 1) , with E ( N ) < ∞ . Then the results above are remain valid, with the deterministic N replaced by 2 E ( N ) N E( N ) Nahum Shimkin Strategic Choice of Arrival Times 18 Computation Explicit computation of the equilibrium distribution F is possible for N = 2 only. In that case, F '(t ) is linear for 0 ≤ t ≤ tb . For N > 2 , a numerical solution is required. ♦ ♦ P{Q −i (t ; F ) = 0} may be computed, for given F , from the evolution equations for the Markov chain ( k (t ), m(t )) , where k is the number in the queue and m the number of arrival by time t , 0 ≤ k ≤ m < N . These equations may be integrated together with (*) to obtain F from any initial value F (0) . A search over F (0) is required to find the solution that satisfies the boundary condition. Nahum Shimkin Strategic Choice of Arrival Times 19 Computation results The following equilibrium arrival densities F '(t ) are obtained for α = 2 , β = 1, µ = N , for several values of N . Nahum Shimkin Strategic Choice of Arrival Times 20 The Fluid Limit Consider the N -th system, with N customers, service rate µ N = µ N , and equilibrium distribution FN . Recall that for the fluid model with volume Λ = 1 and rate µ , we β had f (t ) ≡ F '(t ) = µ αα+ β for t ∈ [Ta , Tb ] = [− µα , µ1 ] . Proposition 3: F N (t ) converges to F (t ) at a rate o( log N ), N in the sense that: taN = Ta − o( N −1 N log N ) N tbN = Tb + o( log N ) N f N (t ) = µ αα+ β − µ P0 (t )1{t ≥0} , t ∈ [taN , tbN ] log N where 0 ≤ P0 (t ) ≤ N1 for t ≤ Tb − o( ) N Nahum Shimkin Strategic Choice of Arrival Times 21 Further Work Heterogeneous cost parameters Nonlinear costs, due-date arrival times. Nahum Shimkin Strategic Choice of Arrival Times 22 Nahum Shimkin Strategic Choice of Arrival Times 23