IPS 1031 - Advances in Optimal Design Techniques
Category: IPS
Thursday 9 October 10:50 a.m. - 12:30 p.m. (Europe/Amsterdam)
Room - Kilimanjaro 2
Participants
Organizer: Jesus Lopez-Fidalgo (Instituto de Ciencia de los Datos e Inteligencia Artificial, Universidad de Navarra, Spain)
Chair: Jialiang Li (Department of Statistics and Data Science, National University of Singapore)
1. Presentation Title: An innovative proposal for a joint multi-response Kriging modelling and optimization with an application to freight trains
Authors: Nedka Dechkova Nikiforova1, Rossella Berni1, Luciano Cantone2
1Department of Statistics Computer Science Applications “G. Parenti”, University of Florence, Italy
2Department of Engineering for Enterprise “Mario Lucertini”, University of Rome “Tor Vergata”, Rome, Italy
Abstract: In this talk, we deal with a proposal for a joint multi-response Kriging modelling and optimization, by specifically considering Universal Kriging models involving a non-constant trend and anisotropic covariance functions. More specifically, we propose a joint Kriging modelling for a multiple response situation, aiming also to study the association among the responses.
2. Presentation Title: A design optimality criterion based on the AUC for classification
Authors: Carlos de la Calle, Jesus Lopez-Fidalgo, Pablo Urruchi
Speaker: Jesus Lopez-Fidalgo (Instituto de Ciencia de los Datos e Inteligencia Artificial, Universidad de Navarra, Spain)
Abstract: An appropriate estimation method for classification and then an optimality criterion for searching for an optimal subsampling procedure are proposed. An example for classifying tweets coming either from humans or from bots illustrates the results
3. Presentation Title: Optimal designs for models under the Flexible Closed Skew Normal Distribution
Authors: G. González Farías, J. López Fidalgo, J. U. Márquez Urbina
Speaker: Graciela Gonzalez-Farias
Abstract: We introduce a parsimonious and flexible subclass of the Closed Skew Normal (CSN) distribution. We have derived and proved several important properties for this subclass, showing that it is identifiable and closed under marginalization and conditioning, and that zero correlation implies independence. We discuss why these random fields serve as valid models, and we investigate least squares estimators within this framework. In addition, using the Fisher information matrix of a particular parametrization of the model, we show the optimal locations to improve model estimators under the normal distribution. Finally, to assess the potential impact of this optimization process in linear models, we present a comparative simulation study incorporating different levels of bias.
4. Presentation Title: Optimal Exact Designs for Small Studies in Toxicology with Applications to Hormesis via a Metaheuristic Algorithm
Speaker: Ray-Bing Chen, PhD (Department of Statistics & Institute of Data Science, National Cheng Kung University, Tainan, Taiwan)
Abstract: There are theory-based methods for constructing model-based optimal designs when the sample size is large. The problem becomes challenging when the sample size is small. The theory may no longer apply and even if it did, the optimal design may not be implementable. We provide examples and also show that a simple rounding procedure of the weights from an optimal approximate design to an optimal exact design can produce the wrong optimal exact design. To solve this longstanding, serious and practical problem, we propose a state-of-the-art nature-inspired metaheuristic algorithm to find efficient designs for an experiment with a small sample size. As an application, we use the algorithm to find an optimal design for a toxicology experiment to detect existence of hormesis in a dose response study, and an optimal design to estimate the hormesis threshold. Being a metaheuristic algorithm, it can be used to find different types of optimal designs for various statistical models. We demonstrate its flexibility by finding locally D-optimal designs for estimating model parameters in logistic models for small experiments, along with user-friendly codes to produce all designs in the paper.
Abstracts and papers
For more details on registrations and submissions for the 65th ISI World Statistics Congress, please first login to your account. If you do not have an account then you can create one below: