Offshore Pipeline Routing Optimization via Probabilistic Reinforcement Learning for Varying Landslides' Stability

Optimizing offshore pipeline routes is a challenging problem, especially in environments with significant uncertainty, where site characterization data is limited. Current approaches in offshore pipeline design and optimization typically involves three stages: identifying the components of relevant geohazards, creating composite maps by pixel-per-pixel aggregation techniques of geohazards, and computing likely least-cost paths between any two points of reference (i.e., origin and destination). Dijkstra's algorithm remains the algorithm of choice in pipeline optimization as it is the default least-cost path determination implemented in most GIS software. However, such implementation can be computationally intensive for a large grid problem. Recently, Q-learning, a Reinforcement Learning (RL) method, was proposed to optimize subsea pipeline route design. Nonetheless, the current approaches assume that the `cost' map that is used as the foundation for the optimization algorithm is fixed. This study proposes the modelling of pipeline routing based on a physics-based model (i.e., infinite slope model), the Bayesian paradigm, and the use of Policy Iteration. This results in a dynamic programming algorithm which uses a probability of failure field as a prior, and then optimizes any pipeline routing given a designated origin and destination positions. Preliminary results indicate that the optimized model consistently identifies the most optimal path between the given origin and destination positions, and that the training procedure is more computationally efficient than other competing algorithms.

Keywords: Offshore, Pipeline routing, Reinforcement learning, Stochastic, Probability of failure.