Real-Time Project Scheduling Under Uncertainty

Abstract

Project scheduling is a crucial management task across many fields, such as manufacturing, software development, and logistics. In a dynamic business environment, a real-time project scheduling approach can significantly reduce the time and effort project managers spend on planning and monitoring projects. This thesis studies an integrated proactive-reactive scheduling approach using deep reinforcement learning to deal with stochastic activity duration. A genetic algorithm, combined with the critical chain method, is employed to generate a robust baseline schedule proactively. A Markov decision process is developed as a sequential decision making problem for reactive scheduling. To enhance model training, moment matching, and fast-forward selection methods are used to identify the most representative subset of uncertain activity durations. Double deep q-network with a convolution neural network is then used to train the model. The goal is to generate a reactive schedule that minimizes the total tardiness from the baseline schedule. Computational results on benchmark instances were performed to demonstrate the effectiveness of the proposed model and sensitivity analysis was conducted to evaluate the performance.