A computational architecture to address combinatorial and stochastic aspects of process management problems

This thesis considers the problem of portfolio selection and task scheduling arising in research and development (R&D) pipeline management, where several projects compete for a limited pool of various resource types. Each project (product) usually involves a precedence-constrained network of testing tasks prior to product commercialization. If the project fails any of these tasks, then all the remaining work on that product is halted and the investment in the previous testing tasks is wasted. Further, there is significant uncertainty in the task duration, task resource requirement, task costs/rewards and task success probabilities. A two-loop computational architecture, Sim-Opt, which combines discrete event simulation and mathematical programming, has been developed by viewing the underlying stochastic optimization problem as the control problem of a performance-oriented, resource-constrained, stochastic discrete event dynamic system. Sim-Opt introduces the concept of a time line, which is a controlled, simulated trajectory that represents a specific combination of the realization of the various sources of uncertainty in the system. Multiple time lines are explored in the inner loop of Sim-Opt to accumulate information, which is subsequently used in the outer loop to obtain improving solutions to the system. Methods have been developed to integrate information from the inner loop with respect to portfolio selection and resource management. Industrially motivated case studies have been investigated using Sim-Opt to evaluate the effectiveness of different policies of operation, to evaluate the value of outsourcing of resources, and to obtain improving solutions in the outer loop. Basic algorithm and software engineering methods to achieve significant improvements in the performance of formulation generation and the generation of a heuristic lower bound along with identification of cut families for effective application of branch-and-cut methods for solution have been described. Lastly, the data complexity of the pipeline problem has been addressed by defining an XML-based structured input language for modeling the data needs in a formatted and extensible manner. This thesis demonstrates the benefit of explicitly viewing the R&D pipeline as the control problem of a discrete-event dynamic system and the effectiveness of Sim-Opt as a practical approach for addressing stochastic optimization