Project Risk Management using Deep Reinforcement Learning

Project scheduling under uncertainty involves balancing acceleration costs against the risk of delay. Due to the uncertainty in activity durations, the project manager must decide before each task whether to allocate extra resources to reduce expected lateness. This classical time-cost trade-off problem (TCTP), has been widely studied both in deterministic and stochastic settings. Recently, the problem has been formulated as a stochastic shortest-path decision process, providing a natural basis for learning-based control. Advances in reinforcement learning (RL) now enable adaptive scheduling without explicit stochastic models. Previous studies applied deep RL to resourceconstrained scheduling and activity acceleration. Building on these foundations, the present paper formulates project scheduling as a decision process in which each activity has an uncertain duration described by a probability distribution, and an optional acceleration action with associated cost. A deep reinforcement learning approach is developed to determine, at each milestone of the project, whether acceleration is economically justified given the expected risk of overall project delay. The proposed framework thus combines classical time-cost optimisation with modern risk-aware learning to provide a flexible tool for managing project risk under uncertainty. In the paper both simple sequential projects as well as more complex projects with parallel activities are considered. The proposed methodology is illustrated by a few numerical examples.

Keywords: Project risk management, scheduling under uncertainty, deep reinforcement learning.