A genetic programming approach for heuristic selection in constrained project scheduling

The resource-constrained project scheduling problem (RCPSP) with cash hows investigates the scheduling of activities that are linked by precedence constraints and multiple resource restrictions. Given the presence of cash hows which represent expenses for initiating activities and payments for completed work, maximizing the Net Present Value of the project is a practical problem. This is a complex combinatorial optimization problem which precludes the development of optimal schedules for large projects. Many heuristics exist for the RCPSP, but it has proven difficult to decide in advance which heuristic will provide the best result, given a problem characterization in terms of parameters such as size and complexity. In this paper we discuss the use of genetic programming (GP) for heuristic selection, and compare it directly to alternative methods such as OLS regression and neural networks. The study indicates that the GP approach yields results which are an improvement on earlier methods. The GP solution also gives valuable information about project environments where a given heuristic is inappropriate. In addition, this approach has no problem evolving complex nonlinear functions to capture the relationship between problem parameters and heuristic performance. Thus the results given in this paper shed light on the logical domains of applicability of the various heuristics, while at the same time provide an improved heuristic selection process.