Name of the participant: Ogün Yurdakul
Description of the IT research project: The day-to-day operation of companies is fraught with numerous instances where a decision needs to be taken in the presence of uncertain parameters. Most studies in the literature employ stochastic optimization (SO) techniques to take decisions under uncertainty. While SO techniques require that the probability distribution of the uncertain parameters be specified a priori, companies cannot know the true distribution of uncertain parameters. On the flip side, companies have at their disposal copious amounts of data on the past realizations of uncertain parameters and on the covariates that influence the realization of uncertain parameters. However, most SO approaches do not effectively leverage contextual information and represent uncertainty with one-size-fits-all probability distributions. The studies in the machine learning (ML) literature brought forth several supervised learning models that can accurately forecast the realization of uncertain parameters. Nevertheless, the generated forecast is typically a single value in the form of a point forecast that cannot capture the stochastic nature of an uncertain parameter.
To address these shortcomings, this project aims to develop a framework that utilizes ML techniques in conjunction with SO models so as to aid small-scale enterprises in taking decisions under uncertainty. Following in the footsteps of the contextual SO paradigm, also referred to as predictive prescription, a key pillar of the project is to work out an overarching methodology that jointly leverages supervised ML techniques and SO models. The proposed methodology will be designed in a generic way that effectively exploits contextual information and efficiently solves a conditional SO problem given an observation on covariates. Further, specific use cases will be identified in consultation with the industry partner and the proposed framework will be tailored to the features, uncertain parameters, and the objective function of the identified use cases.
Software Campus partners: TU Berlin, DATEV
Implementation period: 01.04.2022-31.03.2024
