Transcript – Sequential Minimum Energy Designs for Model Selection

Our poster is entitled Sequential Minimum Energy Designs for Model Selection and our searches on a statistical method to design experiments for the purpose of model selection or hypothesis testing, which is obviously a very common goal in science, engineering and other fields. For example, in 3D printing, we may want to design an experiment to determine if one printing model is more reliable than another or one printer printing material is better. And the general public we’re interested in is how should we decide what data should be collected to figure out the answer to such questions? More specifically, our work addresses some common challenges which arise when designing experiments for model selection. For example, designs that are highly optimized for distinguishing two or more models may only focus on settings where the models differ most and so are often poor for checking whether any of the models are more generally adequate for capturing the phenomena of interest. Our approach tackles such issues by ensuring a tradeoff between selecting to collect data that will best distinguish key models and choosing to collect data under varied settings that will allow a more general assessment of model adequacy. This wider exploration of different settings is called a space drilling design, and the reference to minimum energy in our title is because, to motivate our design approach, we make an analogy with charged particles, which also tend to spread out this goal of separating key models while also checking them more. General adequacy has wide ranging applicability, and going forward, we’re very excited to have the chance to try out our new design approach in a number of scientific and engineering settings.