OR Seminar

Date: 24/11/2017

Hour: 11h

Venue: Sala de Seminários, ed. VII

Dealing with messy data: optimization in an uncertain world

Cláudia Soares (ISR and IST)

In this talk I'll focus on a few large-scale learning problems, namely big data processing and distributed indoor localization. We live in an increasingly data-driven world. We base our driving on Google maps or waze, we buy online using recommendations from Amazon, we share our lives in social media, like Facebook or Instagram. Optimization is a key tool to solve important problems arising in today's digital world. From this premise I will provide an overview of several ongoing and future research projects that explore this perspective on the future of data science, with an emphasis on distributed data processing.

Calibration of parameters in Dynamic Energy Budget models using Direct Search methods

Jéssica Morais (MARETEC and IST)

Dynamic Energy Budget (DEB) theory aims to capture the quantitative aspects of metabolism at the individual level for organisms of all species.

The choice of a DEB model is based on the phylogenetic information of the species. The parameterization of that model is then based on information obtained through the observation of natural populations and experimental research. The resulting set of parameters defines the model and reflects the energetic performance of the organism.

Currently the Add-my-pet toolbox estimates these parameters with the help of the Nelder-Mead simplex algorithm (available in DEBtool), a popular direct search method which at- tempts to compute the minimum of a nonlinear function, without any derivative information.

However, it has some limitations regarding convergence and how to address constraints.

In this work we will present the results of a numerical comparison between the Nelder-

Mead simplex method and the SID-PSM algorithm, a directional direct-search method for which convergence can be established both for unconstrained and constrained problems. We will propose different approaches to obtain a robust and efficient algorithm, able to solve the constrained optimization problems resulting from the estimation of the DEB model parameters for different species.