Research and development planning: Selecting and scheduling projects with approximate solutions to a Markov decision model
This research investigated approximate solution methods for solving the research and development (R&D) project selection and scheduling problem. Globally, industry and government invested a projected one trillion dollars in R&D in 2006, yet detailed decision models have not been widely used in practice. The key research contributions were: a comprehensive survey of the R&D planning research literature, a Markov Decision Process (MDP) model based on the R&D planning tool of technology roadmaps, and an extensive evaluation of approximate solution methods for solving different types of project networks.
Many R&D planning models have been created, including scorecards, linear programs, dynamic programs and various heuristics. Technology roadmaps or networks of projects that lead to technology goals have increased in use in recent years. Unlike models of serial or parallel sets of projects often seen in the literature's models, these technology roadmaps may have complex forms with complicated project precedence relationships. To model this increased complexity, the problem was formulated as a discrete, dynamic sequential stochastic program using a Markov decision model.
Due to the "curse of dimensionality," models of large and complex project networks cannot be solved with dynamic programming in a reasonable time. Recently, researchers have proposed genetic algorithms (GAs) for near-optimal solutions to project selection and scheduling problems. Using experiments of simulated project networks, this research tested three approximate solution methods and compared them to the GA approach.
Three approaches were examined for solving the MDP formulation: state aggregation, problem decomposition, and heuristic methods. For state aggregation, the system states were aggregated by grouping together similar project states. In problem decomposition, a project network was split into sub-networks. The sub-network solutions were combined into an overall funding solution. The most effective heuristic was the goal contribution heuristic, where projects were selected based on how much each project contributed toward estimated expected utility.
An integrated heuristic combined state aggregation and the goal contribution heuristic for an overall recommendation. The goal contribution and integrated heuristic were usually the best of the approximate solution methods. In empirical results, both methods had a statistically significant higher utility than the GA benchmark.
Research & development--R&D;
Decision making models;
0790: Systems design
0796: Operations research