Optimization Problems Under Endogenous Uncertainty Observation: Application to Pharmaceutical Research and Development Planning

This thesis deals with the modeling and solution of stochastic optimization under exogenous uncertainty and endogenous observation (STOXUNO) problems with specific emphasis on the planning of the pharmaceutical research and development (R&D) pipeline. STOXUNO refers to problems in which distribution of the underlying uncertainty is known prior to the model formulation, but the time at which the uncertainty is observed depends upon the decision maker. Traditionally, STOXUNO problems were intractable due to excessive formulation size, particularly with regards to non-anticipativity constraints (NACs), and solved either with heuristics or fixing the observation of uncertainty. In this thesis, we present a multi-stage stochastic programming (MSSP) framework for solving this class of problems.;The thesis begins with an overview of the pharmaceutical R&D planning problem and review of solution techniques for both this and general STOXUNO problems.;Next, a basic model for the pharmaceutical R&D problem is presented in the form of a stochastic resource constrained project scheduling problem. Using this basic model, general theoretical properties for greatly reducing the number of non-anticipativity constraints (NACs) are presented.;The fourth chapter of the thesis focuses on solution of the resulting reduced formulation. Recognizing formulation size is the largest hurdle to solution of STOXUNO problems, we explore ways to solve the problem over a reduced formulation while maintaining feasibility and global optimality. Two different solution approaches are explored; a rolling horizon approach and a novel branch and cut algorithm. Both approaches are designed to optimize only over a fraction of the NACs, but represent a majority of all constraints.;In the fifth chapter of the thesis, the basic model is revisited and extended. We show how a number of practical considerations including generalized precedence structure, resource variability, uncertainty interdependence and risk management can be incorporated into the model while retaining the existing properties. Illustrative examples of handling these aspects are presented.