The 20th International Symposium on Transportation and Traffic Theory Noordwijk, the Netherlands, 17 – 19, July, 2013 A Bayesian approach to traffic estimation in stochastic user equilibrium networks Chong WEI Beijing Jiaotong University Yasuo ASAKURA Tokyo Institute of Technology 1 Purpose Estimating Traffic flows on congested networks Path Flows O-D Matrix Link Flows 2 Background • Likelihood-based methods - Frequentist: Watling (1994), Lo et al. (1996), Hazelton (2000), Parry & Hazelton (2012) - Bayesians: Maher (1983), Castillo et al. (2008), Hazelton (2008), Li (2009), Yamamoto et al. (2009), Perrakis et al. (2012) 3 Background • On congested networks: Bi-level model Bi - level Likelihood Link count constraint equilibrium constraint 4 Background • On congested networks: Single level model Bayesian Likelihood Link count constraint equilibrium constraint 5 Highlights • Use a likelihood to present the estimation problem along with equilibrium constraint • Exactly write down the posterior distribution of traffic flows conditional on both link count data and equilibrium constraint through a Bayesian framework • Develop a sampling-based algorithm to obtain the characteristics of traffic flows from the posterior distribution 6 Primary problem • On a congested network, estimating π² based on π± ∗ and πͺ. π²: vector of route flows; π± ∗ : vector of observed link counts; πͺ: pre-specified O-D matrix ; • equilibrium constraint: the network is in Stochastic User Equilibrium. 7 Representation • Bi-level approach: min π( π², π± ∗ ) s.t. πͺ and π π’π • Our approach: π(π²|π± ∗ , πͺ, π π’π) π . . denotes a conditional probability density; π± ∗ , πͺ, π π’π are the given conditions; π π² π± ∗ , πͺ, π π’π → E(πͺ), Var(πͺ), E(π±), Var π± . 8 Decomposition π π² π± ∗ , πͺ, π π’π πππ π‘πππππ π΅ππ¦ππ ′ π‘βπππππ π(π²|πͺ) πππππ π π π’π, π± ∗ π², πͺ) ππππππβπππ π(π± ∗ |π²) π π π’π π², πͺ) ππππ πππ’π‘π πππ’πππππππ’π 9 Equilibrium constraint • π π’π and π π’ππ (see Hazelton et al. 1998): π π’π βΊ π π’ππ ∀π ∈ πΌ π π’ππ : user π displays Stochastic User Behaviour i.e., user π selects the route that he or she perceives to have maximum utility; πΌ: set of users on the networks; • The equilibrium constraint can be obtained as: π π π’π | π², πͺ = ∀π∈πΌ π(π π’ππ |π², πͺ) 10 An illustrative example • Two-route network 110 Link A detector O 200 A Link B ? (90) D Proposed model Equilibrium model 91.81 105.15 True value = 90 True value = 90 11 Path flow estimation problem • The representation of the problem: π π² |π π’π, π± ∗ here, πͺ is no longer a given condition. • Using Bayes’ theorem ∗ π(π π’π, π± | π²)π(π²) ∗ π π²|π π’π, π± = π(π π’π, π± ∗ ) • The constant term π π π’π, π± ∗ = ππππ π‘πππ‘ 12 Path flow estimation problem • The posterior distribution π π²|π π’π, π± ∗ ∝ π(π π’π, π± ∗ | π²)π(π²)/π(π π’π, π± ∗ ) ππ ! ∗ π π π’π, π± π², πͺ) Likelihood ∀π∈π ∀π∈π π π¦π ! βπ Prior probability: the principle of indifference 13 Prior knowledge of O-D matrix • Dirichlet distribution π ππ² = Γ ∀π∈π ππ ∀π∈π Γ ππ ∀π∈π ππ ππ −1 ππ : the relative magnitude of the demand of the O–D pair in the total demand across the network • Do estimation with prior knowledge π π²|π π’π, π± ∗ , π ∝ π π π² π π²|π π’π, π± ∗ ∝ π π π² π π π’π, π± ∗ π²)π(π²) 14 Estimation • Sampling-based algorithm π π²|π π’π, π± ∗ E(πͺ) E(π±) Var(πͺ) Var(π±) 15 Blocked sampler 0 (1) Specify initial samples π²10 , … , π²|π| for π²1 , … , π²|π| , set π‘ ← 1 and π ← 1. (2) For the O–D pair π: draw π²ππ‘ using the Metropolis–Hastings (M–H) algorithm. (3) If π < |π| then π ← π + 1, and go to step (1); otherwise, go to step (3). (4) If π‘ < π then π‘ ← π‘ + 1, π ← 1, and go to step (1); otherwise, stop the iteration. 16 Test network 60 O-D pairs 53 unobserved links 23 observed links (about 30% of the links) 17 Test network “observed” flow on link π, π₯π∗ may be different from the “true” flow, π₯π# due to observational errors, so that inconsistencies can arise in the “observed” link flows. For illustrative purposes, we created the “observed” flow, π₯π∗ by drawing a sample from the Poisson distribution as π₯π∗ ~Poisson(π₯π# ). we created π by introducing Poisson-perturbed errors to the true O–D matrix 18 Link estimates without prior knowledge 19 O-D estimates without prior knowledge 20 Link estimates with prior knowledge 21 95% Bayesian confidence interval 22 O-D estimates with prior knowledge 23 Conclusions • A likelihood-based statistical model that can take into account data constraint and equilibrium constraint through a single level structure. • Therefore, the proposed method does not find an equilibrium solution in each iteration. • The proposed model uses observed link counts as input but does not require consistency among the observations. 24 Conclusions • The probability distribution of traffic flows can be obtained by the proposed model. • No special requirements for route choice models. The National Basic Research Program of China (No. 2012CB725403) 25 QuestionsοΌ Chong WEI chwei@bjtu.edu.cn Yasuo ASAKURA asakura@plan.cv.titech.ac.jp 26