Enterprise AI Analysis

PD-RSAC formulates the problem as a hex-grid semi-Markov Decision Process (semi-MDP) with mixed actions (discrete service/reposition/charge and continuous charging power) and variable action durations. A key innovation is the feasible-action projection layer, which uses a time-limited rolling Mixed-Integer Linear Program (MILP) to guarantee physical feasibility at every decision step. To handle distributional shifts, PD-RSAC optimizes a Soft Actor-Critic (SAC) agent against a Wasserstein-1 ambiguity set, employing a novel graph-aligned Mahalanobis ground metric to capture spatial correlations in uncertainty. This robust backup utilizes a Kantorovich-Rubinstein dual with a projected subgradient inner loop for primal-dual risk budget updates. The architecture integrates a two-layer Graph Convolutional Network (GCN) encoder for state representation, twin critics, and a value network driving the adversary.

Experiments on a large-scale EV fleet simulator built from NYC taxi data demonstrate that PD-RSAC significantly outperforms strong heuristic, single-agent RL, and multi-agent RL baselines. It achieved the highest net profit of $1.22M, compared to $0.58M-$0.70M for baselines, while critically maintaining zero feeder-limit violations. The ablation study confirmed that all components—MILP projection, Wasserstein DRO, and graph-aligned metric—contribute positively to the robust performance. The MILP layer had the largest impact, underscoring the importance of feasibility guarantees.

This research offers a powerful framework for managing complex cyber-physical systems where hard constraints and real-world uncertainties are prevalent. For enterprise EV fleet operators, PD-RSAC promises not only enhanced profitability through optimized dispatch and charging but also guaranteed operational safety, preventing costly infrastructure overloads. The method's ability to adapt to unseen demand patterns makes it highly robust for real-world deployment, reducing risks associated with distributional shifts and providing a strong competitive advantage in urban mobility services. Future work will focus on optimizing the computational overhead of the MILP layer for even greater scalability.