Connectionist approaches for solver selection in constrained project scheduling

Solver selection for many complex practical applications is a difficult problem due to the availability of a large number of heuristic procedures and the resulting heavy requirements on the collection, storage and retrieval of information needed by the solvers. This paper presents the application of connectionist methods to aid the process of heuristic selection, control and management for the resource-constrained project scheduling problem with cash flows (RCPSPCF) which is a difficult combinatorial optimization problem. Many heuristic procedures have been developed for the RCPSPCF, with differing performance characteristics in different problem environments. This makes the task of choosing the most appropriate heuristic or heuristic category for a given instance of the problem a complex task. We apply neural networks to induce the relationship between project parameters and heuristic performance to guide the selection under different project environments. We also compare the results of the neural network approach with those from traditional statistical procedures. An innovative feature of our approach is the integration of statistical and optimization methods with neural networks to address data preprocessing, thereby improving the performance of the neural network. We demonstrate that neural network methodology can be employed both to extract information about project conditions as well as to provide predictions for novel cases. Extensive experimentation with network topologies and learning parameters indicates that this approach has significant promise in identifying categories of heuristics that are appropriate for any instance of the problem, rather than selecting a single best heuristic.