“No Free Lunch” Theorems Applied to the Calibration of Traffic Simulation Models

摘要: In 1997, Wolpert and Macready derived “No free lunch theorems for optimization.” They basically state that “the expected performance of any pair optimization algorithms across all possible problems is identical.” This to say there no algorithm outperforms the others over entire domain problems. other words, choice most appropriate depends upon specific problem under investigation, a certain algorithm, while providing good (both in terms solution quality convergence speed) on problems, may reveal weak others. apparently straightforward concept not always acknowledged by practitioners. A typical example, field traffic simulation, concerns calibration models. this paper, general method verifying robustness procedure (suitable, general, simulation optimization) proposed based test with synthetic data. The main obstacle methodology significant computation time required necessary simulations. For reason, Kriging approximation model instead. tested case study, where effect different combinations parameters, algorithms, measures goodness fit, levels noise data also investigated. Results show clear dependence between study analysis ascertain need global solutions