Anna Schartman

I am a paleoclimatologist investigating the history and drivers of the savanna biome development through Earth’s past, in order to enhance our understanding of the response of these critical ecosystems to climate and environmental change. Here I integrate records of plant-wax biomarker distributions, carbon, and hydrogen isotopic compositions from Miocene Northwest Africa to understand how changes to aridity affected the development and expansion of savanna ecosystems in the region.

Most real-world data-driven decision-making problems involve elements of uncertainty. Take, for example, the problem of portfolio optimization, where future returns on investments are unknown. However, we often have additional context that may help resolve some uncertainty. In portfolio optimization, examples of context include recent investment returns, economic indicators and current events. To solve problems with both uncertainty and context, the traditional approach is “predict-then-optimize:” predict unknown parameters based on context with a machine learning model, then use those predictions to find an optimal decision with mathematical programming. Recently, researchers have developed machine learning algorithms for integrating the prediction and optimization stages, with strong evidence of integration boosting decision quality. However, these algorithms are often unusable with large-scale problems due to computational complexity. In our work, we leverage the shared structure of many practical decision-making problems in order to develop scalable algorithms that integrate optimization for decision-making directly into machine learning pipelines.