Maximizing the Net Present Value of Resource-Constrained Project Scheduling Problems using Recurrent Neural Network with Genetic Algorithm

Abstract

Scheduling long-term and financially dependent projects constrained by resources are of the utmost significance to project and finance managers. A new technique based on a modified Recurrent Neural Network (RNN) employing Parallel Schedule Generation Scheme (PSGS) is proposed as heuristics method to solve this discounted cash flows for resource-constrained project scheduling (RCPSPDC). To resolve the gradient exploding/vanishing problem of RNN, a Genetic Algorithm (GA) is employed to optimize its weight matrices. Our GA takes advantage of p-point crossover and m-point mutation operators besides utilizing elitism and tournament strategies to diversify and evolve the population. The proposed RNN architecture implemented in Julia language is evaluated on sampled projects from well-known 17,280 project instances dataset. This article, establishes the superior performance of our proposed architecture when compared to existing state-of-the-art standalone meta-heuristic techniques, besides having transfer learning capabilities. This technique can easily be hybridized with existing architectures to achieve remarkable performance.