Nour-Eddine Moumni

Efficient allocation of electric vehicles (EVs) to charging stations in dense urban environments requires decision-making under partial observability and timevarying traffic demand. We present a controlled empirical comparison of two multiagent reinforcement learning paradigms for cooperative EV charging assignment in a SUMO-based urban network: Independent Q-Learning (IQL; DTDE) and QMIX (CTDE). Across 7 random seeds and 200 training episodes per seed (2,800 episodes in total), QMIX attains a mean training acceptance rate of 81.5±17.1%, versus 66.0 ± 16.8% for IQL. Under pure-greedy exploitation (ε=0, frozen weights, 32 episodes per seed), QMIX reaches 83.8±20.5% acceptance with 95.4% successper- accept, versus 64.4 ± 15.5% and 93.6% for IQL (Wilcoxon: T = 0, p ≤ 0.016; paired Cohen’s dz = 8.6). Non-RL baselines (NEAREST, WEIGHTED) accept all requests (100%) but incur higher timeout rates (9.3% and 4.1%), highlighting the value of learned selective acceptance. An ablation analysis suggests that, within this tabular-IQL vs. deep-QMIX comparison, CTDE/deep mixing— rather than reward shaping alone—better explains the gap.

Reproducible offline benchmarks for electric vehicle charging require explicit transition definitions and auditable data processing. We introduce IoEV-MARL, comprising 157,071 decision-level transitions from SUMO simulations of the road network of Tangier, Morocco, organized in a 62-column table. The corpus covers 700 collection episodes, seven seeds, four traffic regimes, and five behavioral arms: NEAREST, WEIGHTED, tabular Independent Q-Learning, and two QMIX checkpoint/exploration configurations. Learned collection policies originate from 120-episode training campaigns. A three-layer contract distinguishes decision-time observations, inter-decision elapsed time, and outcome/evaluation fields. Elapsed time is missing when no subsequent decision is recorded. The ETL protocol documents horizon censoring, missing successors, dimensional grain, and file-level provenance. Seven-fold leave-one-seed-out evaluations benchmark Behavior Cloning and offline Implicit Q-Learning on accept/reject decisions conditional on the selected station. Mean Behavior Cloning balanced accuracy ranges from 0.758 to 0.928 across tiers and feature sets, measuring imitation rather than operational performance. In matched greedy evaluation, QMIX has lower station-load dispersion in 7/7 seeds under Normal traffic, versus 2/7, 4/7, and 4/7 under Light, Peak, and Surge. Discrete-state audits characterize coverage gaps and structural reachability constraints. IoEV-MARL supports offline decision-learning research without establishing closed-loop performance gains.

Deep reinforcement learning (DRL) is widely assumed to outperform simpler rule-based and tabular baselines for sequential decision problems. We test this assumption for electric vehicle (EV) charging assignment using Simulation of Urban MObility (SUMO) simulations of real Rabat and Tangier road networks, with a protocol-amended, trained-vs-untrained diagnostic of Deep Q-Network (DQN) and Double DQN (DDQN) agents against untrained controls and an adaptive heuristic under a corrected pipeline and a not-yet-field-validated station-power regime. The primary endpoint, mean queue waiting time, is an assignment-to-arrival access delay. Trained policies are classified against a pre-specified 60 s rule using multiplicity-adjusted bootstrap intervals: of 12 pre-defined contrasts, none reaches the rule, two show a small detectable advantage, one is detrimental (softening to inconclusive with ten seeds), and nine are inconclusive. A legacy 210-run campaign predating the pipeline corrections is exploratory and shows no Holm-corrected difference. A reward/Markov decision process (MDP) audit identifies four structural defects and one replay sampling limitation; of the five audit-derived interventions, three were already active, one was structurally inert although enabled, and one showed no improvement. Exact per-contrast tests of the isolated ablations cannot reject at seven seeds; restoring R2 yields a small DDQN effect in one scenario that does not reproduce in a second. Two post hoc exploratory comparators, a deterministic minimum expected completion time dispatcher (MECT) and a discrete action proximal policy optimization (PPO) agent, are reported descriptively. The diagnostic does not separate under-training from a non-discriminative environment. The broader contribution is a reusable audit workflow for applied reinforcement learning benchmarks: endpoint semantic validation, trained-vs-untrained controls, multiplicity-adjusted inference and reward specification inspection.