On September 15th, 2021, Francisco Antunes successfully defended his Ph.D. thesis entitled “Active Learning Metamodels for Transport Simulation Problems.”

Congratulations, Francisco!

He was supervised by Professors Bernardete Ribeiro (University of Coimbra – UC, Portugal) and Francisco C. Pereira.

Examiners were Prof. Gonçalo Correia (TUDelft), Prof. Luís Paquete (UC, Portugal), Prof. Margarida Coelho (University of Aveiro, Portugal), Prof. Teresa Galvão (Faculty of Engineering of the University of Porto, Portugal), and Prof. Ana Bastos (UC, Portugal).

The defense session was chaired by Prof. António Pais Antunes (UC, Portugal).

Find below the abstract of the thesis! For more information, feel free to contact Francisco at franant@dtu.dk, fnibau@uc.pt, or directly via his Linkedin page.

Transport systems constitute a fundamental structure deeply hardwired into today’s increasingly denser cities. The proper, voluntary, and efficient movement of people, goods, and services is quintessential to healthy and sustainable social and economic advancements.

Urban environments, and their inseparable transport infrastructures, are inherently intricate and highly dynamic. These deeply integrated systems frequently prove to be challenging to model and study due to the multitude of involved variables, external factors, unknown stochastic phenomena, and inevitably human behavior. This overwhelming complexity is, in most cases, not easily encapsulated into closed-form and tractable mathematical formulae, if not indeed infeasible in the first place. As a result, simulation approaches are traditionally employed as modeling tools to explore virtual representations of actual or planned systems to assess their performances ultimately.

Nevertheless, despite encompassing simplified representations of real-world systems, simulation models can too become rather complex and therefore computationally expensive to implement and run, especially if designed with sufficient detail. Even in situations where the simulation runtimes do not represent a significant hindrance to the computer experimentation, input variable spaces with reasonable dimensions can render the exploration of the simulator’s output behavior tiresome to attain systematically. In this sense, simulation metamodels often emerge as easy-to-implement solutions to address the just mentioned shortcomings of simulation modeling. Simulation metamodels are essentially functions that aim at approximating the input-output mappings inherently defined by simulation models. These models, which are generally recognized by their computing speed and simple functional structure, are fitted to previously simulated data and then used for output prediction purposes.

In an effort to further minimize the computational burden of exhausting computer experimentation, active learning can be employed in strategic conjunction with simulation metamodels, as it proposes a more efficient modeling approach by aiming at high predictive performance with as few data points as possible. This is achieved by providing the metamodel, or associated algorithm, with the ability to choose the most informative data points in an iterative manner to be run by the simulation model, thereby reducing data redundancy and runtimes while increasing learning efficiency at the same time.

In this thesis, we develop and explore an integrated active learning metamodeling methodology in the context of transport systems simulation. This methodology seeks to improve the exploration process of the simulation input space in order to allow a more efficient description and understanding of the corresponding output behavior. To this end, the Gaussian Process modeling framework is employed as a simulation metamodel due to its nonlinear and Bayesian properties, which in turn provide a natural platform for developing native active learning strategies. Several transport-related simulation models and active learning settings are analyzed and discussed.

Despite the clear advantages that both simulation metamodeling and active learning, in the context of expensive or systematic computer experiments, can provide to the transport simulation field, to the best of our knowledge, such techniques remain seldom applied in an integrated manner or even known to transport simulation modelers, practitioners, and related professionals. Many simulation-based studies still rely on somewhat manual processes, primarily focusing on the design of static what-if approaches. We are confident that the adoption of active learning metamodeling schemes can serve as a reliable and enhancing complement to the traditional simulation analysis both in research and industry. Furthermore, we believe that this work will stimulate further discussion, developments, and more applications within the field of transportation.