Microscopic traffic simulators are useful tools for designing, evaluating and optimizing transportation systems. In order for a simulator to accurately describe reality, the corresponding traffic model must be properly calibrated. However, the calibration can be rather difficult when the model is computationally expensive and has many parameters. To overcome these difficulties, Sensitivity Analysis (SA) can be applied as an essential instrument for supporting model calibration. Through SA the

more »

... actitioners can obtain valuable information about the relationship between model inputs and outputs, and hence focus on the proper set of most influential parameters for the calibration. Notice, however, that many quantitative SA techniques may also fail with computationally expensive models. To address the above issues, in this paper we developed the quasi-OTEE method, an efficient and qualitative SA approach based on the Elementary Effects (EE) method. The quasi-OTEE approach is able to screen the most influential parameters of a complex model through computing and comparing their Sensitivity Indexes. With the improved sampling strategy, this approach is much more efficient than the original EE method. The approach is validated through a numerical analysis, and a case study of a small synthetic network in Aimsun. A more detailed case study of the Zurich network in VISSIM is then provided to illustrate its application. The results demonstrate that the proposed approach is an effective SA tool for the computationally expensive microscopic traffic models, as well as other complex models in the general scientific community. 50 Copyright ⓒ 2014 SERSC expected to have significant impacts on the model output). Through fine tuning these parameters, it is expected that the model outputs can be efficiently adjusted towards the target values 0. To select the correct set of influential parameters from a complex model, one widely adopted instrument in the general scientific community is the Sensitivity Analysis (SA) [1] . SA can be defined as the study of how the variations in the model outputs can be attributed to the variations in the model inputs [2] . Through providing the modeler qualitative and/or quantitative information regarding the effects of parameters and their variations on the simulation results, SA can help to efficiently identify the most critical parameters for the subsequent calibration. Unfortunately, to the authors' knowledge, there are few examples in which the SA of microscopic traffic models is performed: the authors of [3] and [4] used the One-At-a-Time (OAT) method to evaluate the variance of the VISSIM output. The variance-based SA approach was employed in [5] for calibrating VISSIM; in [6], [7], and [8] for Aimsun; and in [9] and [10] for Paramics. In [1], Kriging metamodelling was employed to approximate and simplify the original complex model in order to investigate the effects of different optimization algorithms. Despite these examples, there appears to be no previous research suggesting a general method for the sensitivity analysis of traffic simulation models. Furthermore, when the traffic model is complex and has many parameters, many traditional SA techniques (e.g., Monte Carlo approaches, for details see [2] ) are unfeasible because of the high computational demand of the traffic model itself. This paper aims at providing a practical and efficient SA approach, which is especially useful for computationally expensive microscopic traffic models. The proposed approach is developed based on the Elementary Effects (EE) method [11], but represents a significant improvement in terms of efficiency. The paper is organized as follows. Section 2 presents a brief review of the EE Method and other modifications made until now. Section 3 explains the methodology for the proposed SA approach, and demonstrates two validation experiments. Section 4 illustrates a case study based on a real calibration project with the City of Zurich. Section 5 presents the conclusions and some recommendations for future research. 51 input space based on random samples. By randomly generating a certain number (e.g., m) of X in the input space, and each time only changing the i-th parameter by Δ, then m EEs of the i-th parameter can be obtained according to the definition above. The mean μ, the standard deviation σ, and the absolute mean μ* of these m EEs, which are called the Sensitivity Indexes (SI), can be derived as: