IoEV-MARL: A Reproducible Offline Dataset and Benchmark for Electric Vehicle Charging Coordination Under Centralized and Decentralized Training

1. Introduction

The mass adoption of Electric Vehicles (EVs) introduces variable loads on urban power grids and traffic networks [1]. To mitigate congestion and balance grid demand, Multi-Agent Reinforcement Learning (MARL) has become a prominent approach for decentralized EV charging management [2]. While algorithmic architectures for MARL have matured, deploying such systems in real-world infrastructure via online interaction carries substantial cost and risk [3]. Consequently, research is shifting toward Offline Reinforcement Learning, which aims to optimize policies using strictly pre-recorded data [4]. However, this shift reveals a major gap in the resources examined here: the pipeline transforming raw IoT telemetry into learning-ready transition matrices is systematically under-specified. This observation builds directly on a companion single-agent audit of RL benchmarking practice for EV charging assignment [5], which formalized a reusable checklist of reproducibility and evaluation-protocol defects (e.g., undiagnosed training instability, selection-bias-prone comparator baselines) discovered while benchmarking trained, untrained, and heuristic policies; the present work extends that single-agent audit approach to the decentralized, multi-station, offline setting. A separate related study [6] compares centralized (CTDE) and decentralized (DTDE) training paradigms for cooperative EV charging coordination under a controlled multi-seed online protocol. That work uses a distinct simulation environment and reports entirely independent online-training results; it shares no data, experiments, or offline transition records with the dataset and benchmark introduced here. A further, independent conference contribution [7] similarly contrasts QMIX and Independent Q-Learning for cooperative EV charging assignment under urban traffic uncertainty, but in an online-only setting with its own simulation environment, seeds, and metrics. As with [6], it shares no data, experiments, or offline transition records with the present work.

A structured analysis of the literature reveals five research gaps limiting progress in offline MARL in urban mobility. There is a lack of standardized, RL-ready EV coordination datasets (G1): open data exists for load modeling and day-ahead forecasting [8,9], and existing offline MARL resources [4,10,11] cover diverse simulated tasks, but none of the resources examined here, nor the cited EV simulation resources [12,13,14], document a fixed transition corpus for the coupled road-network and EV-station admission setting studied in this work. True decentralized learning from logged EV data also remains under-explored (G2), and real-world RL requirements – such as partial observability and non-stationarity – are rarely systematically mapped to the underlying data model (G5) [15].

Beyond the scarcity of data, the structural integrity of existing datasets is at risk. Recent offline MARL benchmarking work documents substantial baseline and evaluation-protocol inconsistencies across studies [16], but the specific link between upstream data-quality treatment and its downstream impact on the sequential decision-making process remains, to our knowledge, unaddressed in this examined literature (G3). Leakage audits tailored specifically to transition-based datasets similarly remain a targeted gap in the resources examined here, and are addressed by this work (G4). Data leakage—where future information inadvertently pollutes the current state observation ($s_t$)—has been documented as a widespread problem across machine learning disciplines [17].

To address these gaps, this paper introduces IoEV-MARL, a dataset and benchmark designed to support reproducible offline decision-learning evaluation using data collected from heuristics and policies trained under centralized and decentralized paradigms. This work makes four contributions:

1.

A causal, non-duplicative Dec-POSMDP Transition Contract ($x_t=(s_t,o_t^i,a_t,$$r_t,s_{t+1},\Delta t,\mathrm{done})$) integrating decision-time UTM coordinates and the empirical per-vehicle inter-decision elapsed time ($\Delta t$, when a subsequent decision exists), which supports future duration-aware SMDP analyses.

2.

An analytical Relational Star Schema Architecture: the currently delivered artifact is a single denormalized FACT_TRANSITIONS Parquet warehouse ($157,071$ rows, 62 columns, verified directly on the released file), five of which are physical-successor fields added to an earlier 57-column schema core, with specifications for future multi-agent tensor reconstruction ($\mathbf{S}\in {\mathbb{R}}^{B\times T\times 27}$, $\mathbf{O}\in {\mathbb{R}}^{B\times T\times N\times 15}$).

3.

A multi-tiered offline behavioral benchmark spanning $157071$ empirical SMDP transitions across 700 collection episodes (five arms × seven seeds × 20 episodes) with a 4-stage non-stationary traffic curriculum, contrasting decentralized tabular policies against monotonic deep value factorization (QMIX); the source IQL and QMIX collection policies are themselves learned over separate 120-episode training campaigns, distinct from the 700-episode collection corpus.

4.

A Reproducible Validation Protocol establishing strict temporal right-censoring handling, a seven-fold leave-one-seed-out evaluation across the same road-network topology, and detailed state-space audits identifying cyber-physical saturation dynamics.

2. Related Work

The development of the IoEV-MARL benchmark draws on three research areas: Electric Vehicle (EV) data modeling, Offline Multi-Agent Reinforcement Learning (MARL), and rigorous methods for validating machine learning models.

2.1. Open Datasets in EV Coordination

Comprehensive reviews by Amara-Ouali et al. [8,9] highlight a structural limitation: the vast majority of existing public datasets are specifically designed for supervised learning tasks (e.g., forecasting). Recent charging-load modeling studies confirm this pattern, applying machine learning and statistical forecasting to predict demand from order or session data without releasing a decision-level transition corpus [18,19]. While other recent studies have applied RL to EV charging management [2], they rely on closed online simulation loops. Among the resources examined here, no open corpus of decision-level transitions for decentralized, multi-station coordination under a coupled road-network simulation is available for this specific task; a notable exception synthesizes PHEV charging-session data via conditional tabular GANs specifically to train offline RL algorithms, though without a decentralized multi-agent transition contract [20]. The most directly related EV charging resources – ACN-Data/ACN-Sim [12,21], EV2Gym [13], SustainGym [14], the session-level transaction dataset of [22], and the synthetic offline-RL-ready corpus of [20] – are compared against IoEV-MARL on task, data origin, fixed-corpus availability, and documented protocol in Table 1 below. Among the resources examined here, IoEV-MARL combines SUMO-grounded road-network physics, a decentralized multi-station Dec-POSMDP transition contract, and an assembled fixed offline corpus for this problem class.

2.2. Offline MARL Benchmarks and Algorithms

Offline RL eliminates active environment interaction by training policies on static datasets; D4RL [23] established this paradigm’s standard single-agent benchmarking practice. Extending this to the multi-agent setting introduces a coordination-specific failure mode absent from single-agent offline RL: conservative, per-agent value regularization (as in CQL [24]) can independently push each agent toward a locally conservative action that is jointly uncoordinated, degrading rather than improving the team outcome. OMAR [10] addresses this directly by combining first-order policy gradients with zeroth-order optimization to better optimize conservative value functions over actor parameters; CFCQL [25] and OMIGA [11] also aim to preserve coordination through counterfactual conservative regularization and implicit global-to-local value regularization, respectively. IoEV-MARL addresses the coupled EV-charging/SUMO-road admission task with a SUMO-grounded urban mobility simulation, explicitly following the Dec-POSMDP, macro-action formalization [26,27] that underlies this offline-MARL coordination literature; OG-MARL [4] (evaluation-protocol standardization across MAMuJoCo, SMAC, PettingZoo, Flatland, among other environment families) and the SUMO-grounded traffic-signal-control benchmarks sumo-rl/RESCO [28] are positioned against IoEV-MARL in Table 1.

2.3. Comparative Positioning

Table 1 summarizes, for each resource discussed above, its task, data origin, whether it releases a fixed offline transition corpus (rather than an online environment that generates trajectories on demand), and whether a benchmarking protocol with baselines is documented. Entries marked "not identified in the cited resource(s)" reflect an absence we could not verify from the cited sources, not a claim that the capability is impossible.

Relative to the resources compared in Table 1, IoEV-MARL’s specific contribution is an audited decision-level transition corpus for this EV-charging admission setting, linking records to physical outcomes, censoring indicators, and provenance, with seed-held-out BC and offline Implicit Q-Learning evaluations. These evaluations address the logged admission task rather than complete station selection or joint offline multi-agent coordination. The assembled corpus is publicly deposited, as stated in Data Availability. This comparison establishes task-specific positioning, not exhaustive priority or operational superiority.

2.4. Real-World RL Challenges and Data Leakage

RL deployment in cyber-physical systems faces non-stationarity, partial observability, and execution delays [15]. Furthermore, dataset construction errors frequently lead to temporal data leakage, compromising reproducibility [17]. IoEV-MARL explicitly encodes the per-vehicle inter-decision elapsed time ($\Delta t$) and separates observations from post-hoc KPIs to ensure rigorous AI validation.

3. Task Formalization and Transition Contract

Fixed-step transition representations do not directly express the variable intervals between vehicle decisions that arise in cyber-physical urban mobility; we therefore log event-based transitions and their observed inter-decision elapsed time while the simulation continues to evolve, following the semi-Markov temporal-abstraction framework [29]. The environment is formalized as a Decentralized Partially Observable Semi-Markov Decision Process (Dec-POSMDP) [26,27].

3.1. Agent Definition, State Space, and Structural Dead Zones

In this framework, each EV operates as a decentralized routing requester $i\in \{1,\dots ,N_{veh}\}$, interacting with $N=6$ heterogeneous charging stations in Tangier (UTM Zone 30N). An agent is triggered to make a discrete allocation decision at timestamp t when its State of Charge (SoC) falls below a critical threshold ($\mathrm{SoC}\le 65\%$). The global cyber-physical state $s_t\in \mathcal{S}$ encompasses per-station occupancy/queue states, aggregate demand, fleet-level statistics, time-of-day, and spatial load-balance features – the encoded global vector does not include electrical-grid or road-traffic-density variables.

Row-level definition. One logged transition row corresponds to one EV’s decision instant, identified by the composite key defined in Section 5.1 and audited in Section 5.4. The action $a_t\in \{0,1\}$ is a binary accept/reject gate applied to a single already-selected candidate station $c_j$, chosen upstream by the deployed station-selection policy; it is not a joint action over the full candidate set. QMIX’s per-station value heads score the $N=6$ candidates before this gate decision is taken; they do not generate separately logged station-agent rows. Alternative candidates are not exported, so only the downstream gate, not upstream station selection, can be learned or evaluated offline from this release.

To enforce decentralized execution and prevent data leakage, each deciding agent i receives a local observation $o_t^i\in {\mathcal{O}}^i\subseteq s_t$ comprising:

1.

Internal Vehicle State: Current battery level $\mathrm{SoC}\in [0,100]\%$, and Cartesian decision-time coordinates $(x_t,y_t)$ in the local UTM zone 30 WGS84 road projection.

2.

Infrastructure V2I Signals: Network driving distance to station $d(v_i,c_j)\le 5000\mathrm{m}$, access queue length $q_j\in [0,6]$, and active charger occupancy $l_j\in [0,3]$.

3.1.1. Discrete State Formulation and Structural Reachability

In the tabular formulation and state-indexer module, the multi-agent state space is discretized across seven cyber-physical dimensions:

$$\begin{array}{cc}\hfill |{\mathcal{S}}_{\mathrm{expanded}}|& =\underset{\mathrm{SoC}}{\underbrace{4}}\times \underset{\mathrm{Distance}}{\underbrace{3}}\times \underset{\mathrm{Charger}\mathrm{Load}}{\underbrace{4}}\times \underset{\mathrm{Queue}}{\underbrace{3}}\hfill \\ \hfill & \times \underset{\mathrm{Time}-\mathrm{of}-\mathrm{Day}}{\underbrace{4}}\times \underset{\mathrm{Prior}\mathrm{Action}}{\underbrace{2}}\times \underset{\mathrm{Wait}-\mathrm{Zone}\mathrm{Buffer}}{\underbrace{4}}=4608\mathrm{nominal}\mathrm{states}\hfill \end{array}$$

(1)

yielding $4608\times 2=9216$ theoretical state-action pairs $(s,a)$. The seventh dimension encodes the upstream physical waiting buffer status: $\mathrm{wait}\left[0\right]$ (unlocked/unequipped station, $1152$ states) and $\mathrm{wait}[1..3]$ (active equipped buffer: empty, partially occupied, and saturated buffer, $3\times 1152=3456$ states). In fully equipped station topologies, $\mathrm{wait}\left[0\right]$ remains dormant.

However, empirical topological audits of the Tangier road network reveal that significant portions of this theoretical space constitute structural dead zones governed by physical infrastructure mechanics:

• High Battery Boundary ($\mathrm{SoC}>65\%$): Because routing decisions are initiated exclusively when $\mathrm{SoC}\le 65\%$ (so $65\%$ itself is eligible), the upper range of battery charge ($(65\%,100\%]$) is never invoked, restricting active states to lower energy tiers; this is not in tension with the $\mathrm{soc}\left[3\right]$ bin threshold ($\ge 60\%$) reporting non-zero support, since $[60\%,65\%]$ remains eligible.
• Queue and Buffer Dead Zones: Station admission controllers and reservation timeouts ($720\mathrm{s}$ soft / $1440\mathrm{s}$ hard limit) reduce demand before long physical queues can form. Consequently, severe queue states ($\mathrm{queue}\left[2\right]$) and full wait-zone buffers ($\mathrm{wait}\left[3\right]$) require extreme traffic injection (surge regimes) to be activated.
• Spatio-Temporal Boundary Constraints: Vehicles enter via peripheral highway gates, so by road-network design EVs are not expected to be located within $500\mathrm{m}$ of central downtown stations during early episode timestamps ($t<900\mathrm{s}$); this expectation follows from gate placement and minimum travel time, which make the corresponding spatial-temporal cells sparsely populated early in episodes. We have not exhaustively audited all episodes to establish zero occurrences.

Each vehicle identifier is tokenized per episode as $\tau \left(v_i\right)=\mathtt{anon}\_\Vert \mathrm{SHA}-256(v_i\Vert \mathrm{seed}\Vert \mathrm{episode})[:8]$, an episode-scoped salted digest: not a keyed HMAC; strong anonymization and non-re-identification have not been demonstrated. This supports local intra-episode trajectory linkage via the composite key (Section 5.1), without claiming that identifiers persist across runs, which has not been verified.

3.2. The SMDP Transition Contract

In physical road networks, an action $a_t\in \{0,1\}$ ($0=\mathrm{Reject}$, $1=\mathrm{Accept}$) precedes a variable duration before the subsequent decision epoch, during which concurrent traffic continues to evolve on the network (Figure 1). Physical service execution (en-route travel, wait-zone queueing, and plug charging) is tracked in separate physical telemetry fields (result_delay_sec for decision-to-result latency, and charge_duration_sec strictly for the active plug connection interval), whereas the SMDP transition duration $\Delta t$ records the inter-decision interval strictly when a subsequent decision epoch exists. We establish a causal, strictly non-duplicative SMDP transition contract:

$$x_t=\left(s_t,o_t^i,a_t,r_t,s_{t+1},\Delta t,\mathrm{done}\right)$$

(2)

where:

• $s_t,o_t^i$ denote the global state and local observation captured strictly at the decision timestamp $t_{dec}$ on the road network (separate from later result-writing steps).
• $a_t\in \{0,1\}$ is the discrete allocation action.
• $r_t\in \mathbb{R}$ represents the immediate reward signal. The implemented reward is outcome-conditional rather than a continuous grid/queue penalty, and is strictly piecewise in $a_t$: a rejected decision receives a flat baseline with no physical penalty terms, while an accepted decision accrues the outcome base plus four physical shaping terms:

  $$r_t=\left\{\begin{array}{cc}{R}_{\mathrm{reject}}\hfill & \mathrm{if}{a}_t=\mathrm{Reject}\mathrm{or}\mathrm{station}\mathrm{rejects}\hfill \\ \begin{array}{cc}\hfill & R_{\mathrm{outcome}}\left(a_t\right)-{\lambda }_q·\tilde{q}\left(o_t^i\right)-{\lambda }_{\mathrm{dist}}·\tilde{d}\left(o_t^i\right)\hfill \\ \hfill & +{\lambda }_{\mathrm{soc}}·\left(1-\frac{{\mathrm{SoC}}_t}{100}\right)-{\lambda }_s·\mathrm{sat}\left(o_t^i\right)\hfill \\ \hfill & +r_{\mathrm{coop}}\left({\mathbf{s}}^{\mathrm{final}}\right)\hfill \end{array}\hfill & \mathrm{if}{a}_t=\mathrm{Accept}\mathrm{and}\mathrm{station}\mathrm{accepts}\hfill \end{array}\right.$$

  (3)

  where $R_{\mathrm{outcome}}\left(a_t\right)\in \{R_{\mathrm{success}},R_{\mathrm{failure}},R_{\mathrm{timeout}}\}=\{+1.00,-0.80,-0.60\}$ is selected by the decision status determined when the outcome is finalized, and $R_{\mathrm{reject}}=-0.10$ is a fixed baseline applied identically whether the agent itself rejected or the station declined an attempted accept, so the agent is not double-penalized relative to a proactive reject. ${\lambda }_q=0.2$, ${\lambda }_{\mathrm{dist}}\in [0.05,0.10]$ (configuration-dependent), and ${\lambda }_{\mathrm{soc}}=0.10$ weight the normalized queue, distance, and SoC-urgency terms respectively. ${\lambda }_s·\mathrm{sat}\left(o_t^i\right)$ is the proportional saturation penalty (defined in Equation (8)), and $r_{\mathrm{coop}}\left({\mathbf{s}}^{\mathrm{final}}\right)={\lambda }_{\mathrm{coop}}·(\overline{L}-L_i)/C$ is the cooperative load-balancing bonus (Section 7.4), applied only on the accept branch. The raw sum is clipped to $[{\mathrm{reward}}_{\min },{\mathrm{reward}}_{\max }]$. On the released Parquet, 100% of logged reject decisions across Tiers 1–3 carry $r_t=-0.10$ with zero variance, consistent with this piecewise structure. No electrical-grid or substation-capacity term is modeled in the current SUMO scenario; an earlier tri-criterion delay/queue/grid formulation was superseded by this outcome-conditional scheme and is not what generated the released reward column.
  Reward timing differs by component: $R_{\mathrm{outcome}}$ is determined once the outcome is known, potentially hundreds of simulation steps after $t_{dec}$; queue, distance, SoC-urgency, and primary-branch saturation terms use the captured decision-time observation $o_t^i$; and $r_{\mathrm{coop}}$ uses a fresh snapshot of all stations’ loads at finalization (station_store.get_snapshot()). The resolved reward is attached to the originating decision row. Consequently, the released $r_t$ is not purely a function of $(s_t,a_t)$ evaluated instantaneously at $t_{dec}$.
• $s_{t+1}$ is the subsequent observed state in the vehicle’s trajectory, or an unassigned null absorption state ∅ upon trajectory completion.
• $\Delta t=t_{k+1}^{\left(i\right)}-t_k^{\left(i\right)}\ge 0$ is the continuous elapsed simulation time in seconds (delta_t, with discrete step delta delta_t_steps) until the same vehicle’s subsequent logged decision epoch, computed in the ETL as $\mathrm{bad}\mathrm{hbox}-\mathrm{bad}\mathrm{hbox}$. On the final decision of a vehicle trajectory—including episodes right-censored at the simulation horizon where truncated = True—no subsequent decision exists and $\Delta t$ is strictly exported as missing (NA).
• $\mathrm{done}\in \{\mathrm{False},\mathrm{True},\mathrm{NA}\}$ records trajectory boundary status, explicitly separating natural task completion (terminated) from simulation horizon cutoff (truncated / right-censored).

Physical charging outcomes (battery energy added $\Delta \mathrm{kWh}$, charge duration $T_{\mathrm{charge}}$, and completion status physical_charge_completed) are recorded in separate evaluation fields to distinguish initial reservation admission from physical charging completion.

3.3. Temporal Transition Logging and Discounting Considerations

The delivered collection policies and offline value-learning baseline use fixed discrete-epoch discounting without duration weighting. Logging empirical $\Delta t$ in seconds supports future duration-aware SMDP analyses using, for example, ${\gamma }^{\Delta t/c}$ with normalization scale $c=60\mathrm{s}$. Such discounting could account for unequal intervals between decisions; it is not evaluated in this work.

4. Collection Protocol and Dataset Construction

The end-to-end cyber-physical data generation, curation, and benchmark evaluation pipeline of IoEV-MARL is illustrated in Figure 3, presented in full below (Section 4.5). The pipeline bridges the physical urban mobility layer to standardized offline reinforcement learning through four modular stages: (1) realistic microscopic co-simulation with physical admission guards, (2) multi-tiered behavioral data generation with unified exploration decay, (3) causal ETL extraction and cryptographic verification, and (4) a production-grade relational star schema in Apache Parquet, intended to support future multi-agent tensor reconstruction.

4.1. Data Provenance, Simulation Environment, and Sanitization

The physical mobility layer is executed via Eclipse SUMO (v1.23.1) [30], replicating the urban road topology of Tangier, Morocco ($2.7\mathrm{km}\times 2.95\mathrm{km}$ bounding box, UTM Zone 30N). The road network is extracted from OpenStreetMap [31] (osm.net.xml) and populated with six heterogeneous charging stations partitioned into slow (11 kW AC, $\eta =0.90$), medium (22 kW AC, $\eta =0.93$), and fast (50 kW DC, $\eta =0.95$) chargers, two of each class (Table 2).

4.1.1. Seed Protocol: Disjoint Monte Carlo Replications

Motivated by the broader reproducibility concerns documented for RL experiments [32], the executed protocol (Campaign A) identifies each pseudo-random stream by replication slot, absolute episode, and stream identifier, using NumPy’s SeedSequence with $\mathrm{bad}\mathrm{hbox}=2026101000$ to separate physical SUMO mobility from algorithmic exploration. The resulting integer seeds are persisted in a manifest and collision-checked before execution: no collisions were found among the $7\times 120\times 2=1680$ values used. This check verifies only the seed table itself, not universal stream independence or bit-identical reproduction across platforms and NumPy versions. Supplement S2 documents the non-cryptographic, probabilistic guarantees, superseded seed designs, and secondary Campaign B.

4.1.2. Spatial Sanitization and TraCI Sentinel Elimination

During vehicle spawning, insertion, and transient boundary hops, TraCI returns a negative sentinel coordinate ($\mathrm{bad}\mathrm{hbox}=-1073741824.0$). Unvalidated pipelines square this negative sentinel on both coordinates in their Euclidean fallback calculation, producing an erroneous distance of $1.97\times {10}^9\mathrm{m}$ (about $1.97$ million kilometers). This silently pushes $8\%$ to $12\%$ of decisions into the distant bin ($\ge 1000\mathrm{m}$). The IoEV-MARL pipeline instead validates coordinate ranges strictly ($\left|x\right|,\left|y\right|\le 50000\mathrm{m}$, $\mathrm{dist}\le 20000\mathrm{m}$), rejecting out-of-network sentinels and preventing sentinel values from contaminating spatial features.

4.1.3. CPython Determinism, Hash Seed Randomization, and Process Isolation

CPython’s set iteration and dictionary construction from unordered containers are subject to per-process hash randomization, a source of silent non-determinism in decentralized MARL tie-breaking and route assignment. IoEV-MARL enforces reproducibility within the documented execution environment via a three-point contract: setting the PYTHONHASHSEED environment variable before spawning child workers, canonical (immutable-key) sorting of all dynamic entity sets prior to policy scoring, and a separate, deterministically seeded random-number generator for each agent. Full engineering detail is reported in Supplement S6.

4.1.4. Multi-Hour Extensibility (Protocol D3-b)

Beyond the standardized $7200\mathrm{s}$ episodic horizon used to generate the released dataset, a separate sequential chaining harness was engineered and tested to verify that the underlying simulation infrastructure can scale to extended multi-hour operations without altering the episodic training protocol; technical verification over consecutive hourly slices confirmed zero invariant violations. A 24-hour diurnal demand scenario (D2) was also specified for future macro-cycle evaluations. Neither the chaining tests nor the diurnal scenario are part of the released 700-episode dataset; both are reported in full, with their engineering detail and scope caveats, in Supplement S3.

4.2. The ETL Pipeline and Asynchronous Alignment

The extraction pipeline transforms raw simulation artifacts into AI-ready semi-Markov transitions through a strict, causal sequence. Unlike legacy multi-table joins that suffer from combinatorial explosion and temporal leakage, the TransitionBuilder operates directly at the decision level ($1:1$ trace-driven mapping).

The ETL pipeline proceeds through four stages (a detailed block diagram of this architecture is provided in Supplement S4):

1.

Cyber-Physical Ingestion:DataExtractor reads the XML trip/stop logs, physical lifecycle events (charge_sessions.jsonl), completion logs (charging_end.jsonl), and decision traces (trace.csv).

2.

Causal Sequential Alignment: Decisions are sorted by vehicle ID, decision step, and trace row index, then forward-shifted within each vehicle trajectory to get the successor state $s_{t+1}$ and the duration $\Delta t=t_{k+1}-t_k$. Trajectories with no subsequent decision have nothing to shift from. Their successor fields remain missing rather than receiving guessed values, as nullable integers for the discrete fields and as missing for $\Delta t$, while the terminal flags themselves are still resolved via the episode’s unfinished-vehicle roster (see the done/terminated/truncated convention below).

3.

Decoupled Physical Match: Accepted allocations ($a_t=\mathrm{Accept}$) link $1:1$ to a station lifecycle (ASSIGNED→ARRIVED→CHARGING→COMPLETED/TIMED_OUT). A completion counts only once vehicle, station, and terminal step ($v_i,c_j,t_{\mathrm{end}}$) all match. Battery energy comes from vehicle-side telemetry ($\Delta \mathrm{Wh}$), not station dispenser meters.

4.

Columnar Export and Provenance:PrivacyTokenizer hashes each identifier (salted SHA-256 of vehicle ID, episode, and seed, truncated to 8 hex characters – not a keyed HMAC), then the result is typed into nullable Parquet/CSV columns and sealed with cryptographic manifests.

The algorithm for this leakage-free semi-Markov extraction is presented in Algorithm 1.

Algorithm 1:Leakage-Safe SMDP Transition Extraction and Physical Verification

Table 3 traces, for each of the four functional stages of the ETL pipeline just described, which source files are read, which scripts and transformation/join rules are applied, and which columns are produced – verified against the extraction scripts.

Privacy tokenization (PrivacyTokenizer, Section 4) and the rejection of raw spatial/PII column names are applied as a final pass over the per-arm output before the merge above, producing tokenized_veh_id and purging any banned raw coordinate column.

On thedone/terminated/truncatedConvention. Each vehicle’s final logged row is resolved via the episode’s unfinished_vehicle_ids roster: $(\mathrm{terminated}=\mathrm{True},\mathrm{truncated}=\mathrm{False})$ if the vehicle’s trajectory ended before the episode did, or $(\mathrm{terminated}=\mathrm{False},\mathrm{truncated}=\mathrm{True})$ if the episode horizon cut it off first; either way $\mathrm{done}=\mathrm{True}$ for that row (counts in Table 6). This administrative classification describes how logging ended, using episode-level records – it is not, by itself, physical proof that every terminated row’s vehicle actually completed charging (Table 6’s physical-match funnel tracks that separately).

Adaptive Dwell Caps and Interpretation of Horizon Censoring (Protocol D4). Under realistic operations, vehicles cannot monopolize public dispensers indefinitely. To limit prolonged dispenser occupancy while allowing for slower chargers, IoEV-MARL replaces static timeouts with an adaptive dwell cap ($T_{\mathrm{dwell}}$) derived from physical battery energy requirements:

$$T_{\mathrm{dwell}}(v_i,c_j)=\min \left(H_{\mathrm{rem}},\max \left(60\mathrm{s},\min \left(1.25\times 3.6\times {\displaystyle \frac{\Delta E_{\mathrm{target}}\left(v_i\right)}{P_{\mathrm{eff}}\left(c_j\right)·\eta \left(c_j\right)}},3600\mathrm{s}\right)\right)\right)$$

(4)

where $\Delta E_{\mathrm{target}}$ (in Wh) is the energy deficit to reach $70\%$ SoC, first converted to kWh (dividing by $1,000$), then divided by power in kW to get a duration in hours. The $3.6$ factor converts that duration to seconds: $T_{\mathrm{expected},\mathrm{s}}=3.6·\Delta E_{\mathrm{Wh}}/(P_{\mathrm{kW}}·\eta )$, equivalently $3,600·\Delta E_{\mathrm{Wh}}/(P_{\mathrm{W}}·\eta )$ if power is in W. $P_{\mathrm{eff}}\left(c_j\right)\in \{11,22,50\}$ kW is the dispenser’s own rated power and $\eta \left(c_j\right)\in \{0.90,0.93,0.95\}$ its own conversion efficiency, both per-charger-class rather than a single global value. The engineering estimate is bounded to $[60\mathrm{s},3600\mathrm{s}]$ before the remaining simulation horizon $H_{\mathrm{rem}}$ is applied as a final hard cap. This horizon constraint can legitimately push $T_{\mathrm{dwell}}$ below the $60\mathrm{s}$ floor near the end of an episode, since the engineering bounds constrain the estimate, not the physical horizon. We do not report a specific comparison of completion rates under the legacy and adaptive caps here, as no such controlled comparison with a stated N and traffic condition is available for this released campaign.

When a session terminates, the transition contract distinguishes physical completion from timeout truncation:

• is_horizon_censored$\in \{\mathrm{True},\mathrm{False}\}$: an episode-level flag, attached to every transition of that episode rather than signaling a per-transition dwell cap. It indicates that the episode itself reached its maximum simulation horizon ($7200$ or $9000\mathrm{s}$, scenario-dependent) while vehicles remained unfinished, as opposed to clearing all demand beforehand. This differs from truncated (Algorithm 1), which flags an individual vehicle’s own last transition being cut off by that same horizon. A $T_{\mathrm{dwell}}$ timeout is different again: it is recorded in the source-level diagnostic request_outcome = timed_out_administrative (not a delivered Parquet column; see below), not through is_horizon_censored.
• physical_charge_completed$\in \{\mathrm{True},\mathrm{False},\mathrm{NA}\}$: nullable – True/False if and only if a physical charging-completion match was resolved for that decision; NA when no such match was attempted or possible (e.g., rejected or reassigned decisions); field-level NA counts are reported in Table 6. When resolved, True iff $\mathrm{SoC}\ge 70\%$ was achieved prior to $T_{\mathrm{dwell}}$.
• request_outcome: a multi-class diagnostic enum (completed_soc_target, timed_out_administrative, rejected_queue_full, interrupted_boundary) present in the per-source simulation/ETL diagnostics, not a column delivered in the released master Parquet itself; readers reconstructing this distinction from the release should use outcome/reason instead.

This metadata enables offline algorithms (such as conservative Bellman backups or survival-based credit assignment) to adjust discount horizons appropriately and avoid pessimistic value penalization on artificially truncated trajectories.

4.3. Relational Star Schema and Analytical Storage Architecture

Following the analytical Star Schema design pattern [33], this release favors a denormalized layout over a traditional 3NF relational model, optimized for Apache Arrow / Parquet columnar scanning. The physical release is delivered as a single consolidated, denormalized master table (FACT_TRANSITIONS, 62 typed columns on the currently released Parquet, five of which were added to an earlier 57-column schema core). Each charging hub has 3 charging plugs and a dedicated waiting buffer for 3 vehicles, establishing an effective combined physical ceiling of 6 vehicles per station site. Each record represents one audited decision epoch (1 row per decision, ratio $1:1$ with zero duplication). The dimensional grain and composite keys of the delivered artifact are made explicit in Figure 4 below; peripheral dimension attributes (e.g., station electrical ratings, wait-zone buffer capacities, simulation traffic profiles) represent contextual metadata rather than separately materialized physical tables.

FACT_TRANSITIONS’s fields are grouped by role: Context, Keys, Observations ($s_t,o_t$), Action ($a_t$), Reward ($r_t$), Evaluation Diagnostics (logged alongside $r_t$, not summed into it), Next Step ($s_{t+1}$), SMDP Delay ($\Delta t$), Termination, Physical Ground Truth, and Metadata/Tiers. The full field-by-field dictionary (stored type, unit, source file/field, transformation rule, null semantics) is given in FIELD_LINEAGE.csv (Supplement S4). One recurring pitfall: the state_id observation uses a per-tier codec (Tier 1/IQL’s 4,608-state encoder vs. Tiers 2–3/QMIX’s separate 1,152-state encoder) and must never be pooled or compared directly across tiers without converting through each tier’s own codec first. Station-level context (power_kw: 11/22/50 kW; charging_capacity: 3 plugs; wait-zone capacity: 3 vehicles) and vehicle/episode context (route_label, planned_vehicles, horizon_sec, seed_manifest_sha256) are logical dimension attributes, not materialized as separate tables in the delivered Parquet.

The dimensional grain, composite keys, and the end-user deliverable inventory are described in Section 5 below.

4.4. Multi-Tiered Behavioral Policy Taxonomy and Two-Stage Routing Architecture

To evaluate offline policy optimization across varying data qualities (inspired by OG-MARL [4]), the benchmark generates distinct behavioral data tiers:

1.

Tier 0 (Deterministic Baselines): Heuristic policies (Nearest Station greedy routing and Congestion-Weighted score routing) providing deterministic comparison baselines without learning exploration.

2.

Tier 1 (Suboptimal Exploratory Policy): Tabular Independent Q-Learning (IQL) under a frozen exploration rate ($\epsilon =0.30$), yielding noisy transitions characterized by high local queue saturation and rejection dominance.

3.

Tier 2 (Intermediate-Schedule Policy): QMIX checkpoint frozen at episode 60 of the 120-episode training schedule (fixed mid-training point, not empirically verified as behaviorally distinct from the converged policy) evaluated under controlled exploration noise ($\epsilon =0.15$).

4.

Tier 3 (Expert Factorized Policy): A QMIX checkpoint [34] featuring decentralized agent execution with monotonic mixing network factorization ($\partial Q_{tot}/\partial q_i\ge 0$) in pure exploitation mode ($\epsilon =0.0$). The "Expert" label denotes this construction (a late-training, deterministic-exploitation checkpoint), not a measured performance ranking: Tier 3 has the lowest mean per-decision reward of the three learned tiers (Section 8), so readers should not infer superior behavioral quality from the label alone.

The complete hyperparameter configuration across all four behavioral data generation tiers is summarized in Table 4.

4.4.1. Decoupled Admission Filtering and Policy Dispatch

IoEV-MARL separates capacity-based feasibility, policy-specific station selection and admission decisions, and physical execution (illustrated in Figure 2), verified directly against the active controller and station-state-store code (agent_controller_realtime_native_waitzone_campaign_fixaudit_20261015.py, station_state_store_waitzone_campaign.py, decision_module_nearest.py, decision_module_weighted.py).

1.

A – Shared Feasibility Condition (all four policies): All four control policies (NEAREST, WEIGHTED, IQL, QMIX) share the same feasibility test computed by the station store, combining a boolean charger-availability flag with wait-zone access:

$$\mathrm{Feasible}\left(c_j\right)\iff \mathrm{bad}\mathrm{hbox}\left(c_j\right)\vee \mathrm{bad}\mathrm{hbox}\left(c_j\right)$$

(5)

where, for the wait-zone-equipped stations used throughout this campaign, $\mathrm{bad}\mathrm{hbox}\left(c_j\right)\iff \left|\mathrm{bad}\mathrm{hbox}\left(c_j\right)\right|<C_j$, with $\mathrm{bad}\mathrm{hbox}$ the union of vehicles occupying or promised a charging place. Waiting-buffer availability accounts for parked occupants plus outstanding waiting-slot claims, without double-counting a vehicle already parked there. Each site has $C_j=3$ charging places and $C_{\mathrm{wait}}=3$ separate off-road waiting places. Six is a combined physical ceiling, not a guarantee of six immediately available admissions (Section 3). All four policies use the same availability condition, though implementation may filter candidates or mask acceptance differently. If no eligible feasible candidate remains, no allocation is made in that decision cycle, and the controller handles the subsequent retry or cooldown – a condition distinct from declining a particular candidate despite available capacity. Historical note: the released Tier 0 (NEAREST/WEIGHTED) campaign reflects a corrected, unified filter; the pre-fix filter definition is detailed in Supplement S7.

2.

B – Policy-Specific Selection and Admission Decision: Among feasible candidates:

• NEAREST Baseline: Greedily routes to the closest feasible infrastructure:

  $$c^{*}=\arg \underset{c_j\in {\mathcal{C}}_{\mathrm{feasible}}}{\min }d(v_i,c_j)$$

  (6)
• WEIGHTED Baseline: Scores feasible stations using a 3-term normalized objective plus an urgency discount, verified directly against decision_module_weighted.py:

  $$c^{*}=\arg \underset{c_j\in {\mathcal{C}}_{\mathrm{feasible}}}{\min }\left[0.40·\tilde{d}\left(c_j\right)+0.35·\tilde{L}\left(c_j\right)+0.20·\tilde{q}\left(c_j\right)-0.05·\mathbb{I}(\mathrm{SoC}<20\%)\right]$$

  (7)

  where $\tilde{d}\left(c_j\right)=\min (d(v_i,c_j)/1000,1)$, $\tilde{L}\left(c_j\right)=\min (({\mathrm{charging}}_j+{\mathrm{waiting}}_j+{\mathrm{reserved}}_j)/3,1)$, and $\tilde{q}\left(c_j\right)=\min ({\mathrm{waiting}}_j/4,1)$; the fixed coefficients ($0.40$/$0.35$/$0.20$/$-0.05$) are not learned or tuned per regime.
• IQL / QMIX (Learned Collection Policies): Online Independent Q-Learning uses tabular station-specific values, whereas QMIX uses neural value estimates. Both resolve binary accept/reject decisions, but their selection rules differ. IQL ranks candidates using the value of the resolved action, with execution-time reservation penalties. QMIX instead selects using the accept–reject value difference, with an accept-first shortlist, then applies its exploration rule to the selected station. Rejection may reflect value estimates, decision margins, tie-breaking, or exploration; it does not establish negative operational value. The controller may exclude the declined candidate and try another station, so declining one candidate does not establish final request failure. Heuristics instead propose allocation to their selected feasible station without a learned rejection gate.

An accept decision is an allocation proposal. The controller verifies route reachability and rechecks availability before reserving a charging place or atomically claiming a waiting slot. Failed verification, reservation, or stop placement can prevent physical admission after acceptance; charging completion is a later lifecycle outcome, established separately by the ETL physical matching. Figure 2 shows this execution layer separately from the two decision components.

Physical Admission Filter and the Role of the Waiting Buffer. IoEV-MARL’s deployed topology equips every station with an off-road waiting buffer, admitting vehicles to it precisely when $\mathrm{bad}\mathrm{hbox}=C_j$, rather than pruning saturated stations from ${\mathcal{C}}_{\mathrm{admissible}}$ as a buffer-less topology would (full counterfactual comparison in Supplement S7). Consistent with this, the released corpus shows $\mathrm{load}=3$ (Full) as the most, not least, frequently visited load bin (Table 9: $150,790$ decisions, $30.7\%$ of visited load-3 states). The legitimate reachability correction in this dataset is the $1152$-state exclusion of the dormant $\mathrm{wait}\left[0\right]$ (unequipped-station) slice, yielding the $3456$-state eligible space already used as the denominator in Table 9; no further correction (e.g., on load) is warranted or applied.

Physical Off-Road Wait Zones and State Space Expansion ($4608$ States). To resolve the admission deadlock where fully occupied stations ($\mathrm{bad}\mathrm{hbox}=3$) prematurely blocked incoming vehicles from reaching the charging premises, IoEV-MARL integrates dedicated upstream off-road buffer zones. Each charging hub is equipped with an adjacent off-road parkingArea accommodating up to $C_{\mathrm{wait}}=3$ waiting vehicles. When all active dispensers are occupied but the upstream buffer has available capacity, incoming EVs are physically admitted via TraCI’s setParkingAreaStop primitive rather than queuing on the active road network or being discarded. Once an active charging plug becomes vacant, the queuing vehicle is resumed via TraCI’s resume command and transitions to the physical charging bay.

The waiting buffer supplies the seventh observation dimension in the $4608$-state tabular formulation defined in Section 3. While this 7-dimensional discrete expansion governs tabular IQL policies and dataset state indexing, deep neural architectures (QMIX) process continuous local and global observation vectors ($d_{\mathrm{local}}=15$, comprising 9 continuous base features plus 6 one-hot station identities; $d_{\mathrm{global}}=27$). Crucially, the saturation penalty ${\lambda }_s·\mathrm{sat}\left(c_j\right)$ in Equation (3) is proportional, not binary, and begins as soon as active dispensers are full, continuing to grow as the waiting buffer fills:

$$\mathrm{sat}\left(c_j\right)=\min \left({\displaystyle \frac{\max (0,\mathrm{load}\left(c_j\right)-C_j+1)}{C_{\mathrm{wait}}+1}},1\right)$$

(8)

where $\mathrm{load}\left(c_j\right)=\mathrm{bad}\mathrm{hbox}\left(c_j\right)+\max (\mathrm{bad}\mathrm{hbox}\left(c_j\right),\mathrm{bad}\mathrm{hbox}\left(c_j\right))+\mathrm{bad}\mathrm{hbox}\left(c_j\right)$ is the combined active-plus-queued-plus-waiting-buffer occupancy. $\mathrm{sat}\left(c_j\right)=0$ while $\mathrm{load}\left(c_j\right)<C_j$ (dispensers available), rises linearly from $1/(C_{\mathrm{wait}}+1)$ (dispensers just full, wait zone empty) to 1 (dispensers and wait zone both full). Because the penalty rises gradually rather than triggering only at total site saturation, it penalizes early waiting-buffer use during transient traffic spikes less sharply than outright saturation.

4.5. End-to-End Operational Lifecycle: From Training to Offline Benchmark

To establish strict chain-of-custody and prevent silent data drift across preprocessing passes, IoEV-MARL implements a Double Cryptographic Manifest architecture based on SHA-256 digests: a Code and Pipeline Manifest (MANIFEST_CODE.sha256) freezing all source components governing the extraction contract (agent controller, XML/CSV extractor, transition builder, privacy tokenizer, campaign runner, schema specification – any modification to the shifting or matching logic invalidates the digest), and a Data and Artifact Manifest (MANIFEST_DATA.sha256) sealing the emitted dataset artifacts (Parquet and CSV mirrors, episode execution metadata, row counts, schema dimensions of 62 typed columns, byte-level checksums), supporting bit-exact file-integrity verification of the released artifacts.

To ensure end-to-end scientific transparency, the workflow for data generation, policy optimization, and artifact evaluation comprises six separate stages (illustrated in Figure 3):

The six stages reduce to two tracks sharing a common RNG and curriculum setup (1,680 collision-checked seed values, four traffic regimes). Track A trains IQL and QMIX online (Stages 1–3) under a shared power-law exploration schedule

$$\epsilon \left(ep\right)=\max \left({\epsilon }_{\min },{\epsilon }_{\mathrm{coeff}}·{(1+ep)}^{-\alpha }\right)$$

(9)

(${\epsilon }_{\mathrm{coeff}}=0.218247$, ${\epsilon }_{\min }=0.080$, $\alpha =0.2093$) over 120 training episodes, then evaluates converged checkpoints deterministically ($\epsilon =0$) across 280 Stage 3 episodes (Section 7.5). Track B collects the released offline corpus: Stage 4 generates five behavioral arms (Nearest, Weighted, Tier 1 IQL $\epsilon =0.30$, Tier 2 QMIX ep060 $\epsilon =0.15$, Tier 3 QMIX ep120 $\epsilon =0$) over 700 episodes, with Tier 3 reusing the 140 Stage 3 episodes under the identical converged checkpoint rather than re-simulating. Stage 5 then extracts and consolidates this corpus via the two-level ETL described in Section 4 – per-arm causal extraction, then multi-tier master consolidation – and delivers the single FACT_TRANSITIONS warehouse. This warehouse provides the data for the Behavior Cloning and offline Implicit Q-Learning benchmarks below, conditional on the logged accept/reject gate rather than a full station-selection or joint multi-agent evaluation (Supplement S5).

5. Data Exploration and Analytics

The following analysis describes the master dataset’s dimensional grain, behavioral-arm composition, and integrity checks. Table 5 provides the per-arm counts and outcome distributions; Table 6 verifies row conservation, uniqueness, and expected missingness.

5.1. Dimensional Grain and Composite Keys of the Delivered Artifact

Figure 4 makes explicit the grain of FACT_TRANSITIONS and the composite natural keys that identify each logical dimension – verified directly against the delivered $157,071$-row, 62-column master Parquet. The grain of FACT_TRANSITIONS is: one logged decision for one vehicle, within one episode, one replication seed, and one behavioral arm. The composite key $(\mathrm{bad}\mathrm{hbox},\mathrm{bad}\mathrm{hbox},\mathrm{bad}\mathrm{hbox},\mathrm{episode},\mathrm{bad}\mathrm{hbox},\mathrm{bad}\mathrm{hbox})$ is complete (no missing values) and unique across all $157,071$ rows; policy_tier alone does not distinguish NEAREST from WEIGHTED (both Tier 0), and omitting behavior_policy from the key produces $6,748$ spurious repetitions. For Tiers 1–3, policy_tier already separates the three learned arms, so behavior_policy adds nothing to their uniqueness key. Removing policy_tier from that reduced key instead groups $103,907$ of the $136,911$ learned-tier rows as duplicates (counting every member of each group, not just the repeats after the first) – but these are legitimate recurrences across different experimental arms, not transitions that should be deduplicated.

Two details matter here. tokenized_veh_id is not a global identifier – the same token can reappear across different seed/episode contexts, which is why the full composite scope above, not the token by itself, defines a trajectory. Also, station_id and physical_station_id answer different questions: the former is the candidate station selected for the decision, the latter is the station actually matched to a resolved physical charging segment (nullable, $100,943$ missing, with all populated values falling among the same six known stations). Reading physical_station_id as "the station where charging finished" is a mistake unless physical_charge_completed is checked alongside it. Episode-level attributes – route_label, episode_status, is_horizon_censored, demand_cleared – were confirmed constant within every (RUN, episode) key, with no exceptions found. Successor and physical-outcome fields stay inside FACT_TRANSITIONS rather than being split into separate dimensions, since they vary per transition, not per logical entity.

5.2. Deliverable Inventory

The prepared dataset package contains three release artifacts: a master Parquet table (IoEV_MARL_Master_transitions.parquet, $157,071$ rows, 62 columns), a row-oriented CSV mirror of the same table, and a SHA-256 manifest (MANIFEST_DATA.sha256) for file-integrity verification; these artifacts are publicly deposited, as stated in Data Availability. Upstream intermediate artifacts (per-arm Parquet files, per-source provenance manifests) support ETL traceability (Table 3) but are not themselves part of the release package; the full artifact inventory, including these intermediate artifacts, is given in Supplement S10. An example of the grain defined in Section 5.1 is the following released row (Tier 1/IQL, seed 1, episode 2): $\mathrm{bad}\mathrm{hbox}=\mathtt{anon}\_\mathtt{0}\mathtt{eb}\mathtt{83}\mathtt{c}\mathtt{2}\mathtt{b}$, $\mathrm{bad}\mathrm{hbox}=1499$, $\mathrm{soc}=33.7$, $\mathrm{bad}\mathrm{hbox}=1146.5$, $\mathrm{bad}\mathrm{hbox}=3$, $\mathrm{bad}\mathrm{hbox}=0$, $\mathrm{bad}\mathrm{hbox}=1917$, $\mathrm{decision}=\mathtt{accept}$, $\mathrm{outcome}=\mathtt{timeout}$, $\mathrm{reward}=-0.6316$, $\Delta t=1240.0\mathrm{s}$, $\mathrm{done}=\mathrm{terminated}=\mathrm{truncated}=\mathrm{False}$.

Supplement S5 specifies a future tensor-batch export for deep offline MARL frameworks (PyTorch/JAX); array files have not yet been serialized. The delivered artifact remains the single tabular Parquet table described above.

5.3. Stage 4 Behavioral Arm Composition

Table 5 reports the composition of all five behavioral arms, computed directly on the released master dataset (140 episodes per arm: 7 seeds × 4 traffic regimes × 5 episodes/regime). Note that this dataset, by construction, contains only vehicles that reached a logged decision; vehicles never requiring a charging decision (e.g., departing before any admission request) are not represented here and would require a separate SUMO-level count, not reconstructable from this Parquet alone.

Tier 0’s $100\%$ accept rate is structural, not a policy property: the heuristics never log a reject row (Section 6.2’s Tier 0 exclusion discussion), so every logged Tier 0 row is, by definition, an accepted decision whose outcome is subsequently resolved. Tier 3’s accept-conditional outcome distribution is markedly different from those of all other arms: $37.2\%$ of its accepted decisions end in failure, versus $\le 1.2\%$ for all other arms, and only $45.9\%$ end in success, the lowest of any arm. This pattern is consistent with Tier 3’s documented lowest mean reward among the three learned tiers (Section 8) and likely contributes directly to it; we have not traced the specific causal mechanism (e.g., a particular saturation regime or candidate-selection pattern under $\epsilon =0$) producing this failure rate, and leave this observed pattern for further investigation.

Vehicles With No Tripinfo-Matched Decision.Table 5’s "Vehicles" column counts only vehicles with at least one logged decision row. A separate, larger set of controlled-fleet vehicle IDs (veh_MARL_*) appears in each episode’s tripinfo-output.xml but never in that column at all. This is consistent with each arm’s own logging rules rather than a difference in driver demand: NEAREST/WEIGHTED log a row only for vehicles that complete a full accept/outcome cycle (Section 6.2’s Tier 0 exclusion discussion – no reject rows are ever logged for these two heuristics), so a vehicle whose request never reaches that point leaves no row at all, while the learned policies’ decision modules log every evaluated decision, accepted or not. This count is also distinct from the Tier 0 unserved rate documented in Supplement S7, which counts vehicles that received a decision but did not reach accept_success; here we isolate vehicles that never received any logged decision. A missing decision does not by itself mean "no request" or "rejected" – the cause could be that no charging need ever arose, an early departure or teleport before any admission attempt, or an un-logged administrative path, and this release does not distinguish between these at the per-vehicle level. Per-policy percentages (ranging from $34.2$–$34.3\%$ for both QMIX tiers to $70.6$–$70.8\%$ for NEAREST/WEIGHTED) and the loaded-vs.-tripinfo-matched denominator distinction are detailed in Supplement S7.

The traceability between ETL source files and output columns (Table 3) is presented alongside Algorithm 1 in Section 4.

5.4. Stage 5 Dataset Integrity Audit

Table 6 reports a direct, reproducible audit of the released master dataset, verified through direct Parquet queries rather than the consolidation script’s printed summary.

To maximize utility, the dataset is aligned with FAIR principles, exported in both .csv (for Pandas/R EDA) and .parquet formats (optimized for PyTorch/TensorFlow).

6. Benchmark Protocol

Randomly shuffling transition tuples can introduce future traffic states into the training set [17]. Evaluation splits should use the dataset’s embedded metadata to separate training and test observations:

• Episode-Level Temporal Splits: Data partitioning must occur strictly at the episode or campaign seed level, rather than by individual transition tuples: steps within the same trajectory are statistically dependent, so the trajectory, not the individual step, is the correct unit for a held-out split [35]. The same strict-separation principle applies to temporal splits in correlated, non-i.i.d. settings such as time series forecasting [36]. Training on early simulation episodes and evaluating strictly on subsequent episodes prevents the identical stochastic realization from contaminating both training and testing partitions.
• Prospective Spatial Generalization Tests: Station-level hold-outs could be investigated in future work; they were not performed in the present benchmarks, which use seed-level hold-outs on the same road-network topology.

Furthermore, the dataset delivers raw, unscaled physical values, leaving users responsible for normalization within their modeling pipeline to prevent joint-preprocessing leakage.

6.1. Offline RL Statistical Support Threshold ($N\ge 5$)

In offline reinforcement learning architectures (e.g., Conservative Q-Learning [24], Batch IQL), out-of-distribution action penalization relies on empirical coverage of visited state-action pairs. Sparse state-action support motivates coverage diagnostics; any minimum-count cutoff is an operational audit criterion, not a demonstrated threshold for value overestimation, policy collapse, or learnability. The delivered dataset spans seven replication seeds; state-action support remains heterogeneous, and any downstream analysis must explicitly state the operational minimum-count threshold it uses.

6.2. Behavior Cloning Protocol

To benchmark sequential decision learnability on the final $157071$-transition master dataset under the leakage-safe protocol above, we execute a multi-tier Behavior Cloning (BC) evaluation across behavioral Tiers 1–3 using 7-fold Leave-One-Seed-Out (LOSO) cross-validation. In each fold, a Random Forest classifier (200 trees, maximum depth 8, random seed fixed to 42) is trained on six simulation seeds and evaluated on the disjoint held-out seed. Models are trained under two distinct feature representations: (i) a 5-feature Physical set (SoC, routing distance $d_j$, active charging count $c_j$, reserved queue count $q_j$, and wait-buffer occupancy ratio), and (ii) a Context set appending vehicle coordinates $(x,y)$, decision epoch $t_{\mathrm{dec}}$, and the categorical route label – 8 numeric columns plus the route label prior to encoding, not a fixed 9-dimensional vector, since the route label is one-hot encoded (fit on the six training seeds, applied unchanged to the held-out seed) rather than entering as a single numeric column. Crucially, the target station identifier $\mathrm{bad}\mathrm{hbox}$ is excluded from the input features, as alternative candidate stations are not exported in the logged tuples and conditioning directly on the chosen hub would induce circular label leakage.

Exclusion of Tier 0 Baselines: Heuristic baselines (Greedy Nearest and Congestion-Weighted) log an assignment record per vehicle rather than per evaluated candidate station as in learned MARL policies. Consequently, unserved departures are not logged as explicit rejection transitions in this tier’s table, yielding zero rejection events across all $20160$ Tier 0 transitions in the master dataset. Because this results in a single-class task with zero target variance due to the logging design, Tier 0 is excluded from the binary accept/reject benchmark without synthetic proxy relabeling.

Comparators and Metric Aggregation: Two trivial comparators are reported alongside the Random Forest: a majority-class classifier and a classifier sampling at the training set’s observed class prevalence. Balanced accuracy is computed per fold on the held-out seed and aggregated as the mean ± standard deviation across the 7 folds, per tier and feature set, following established guidance on reporting result variability rather than single point estimates across runs [37].

6.3. Offline Implicit Q-Learning Protocol

To benchmark value-based reinforcement learning on the consolidated dataset without active simulation interaction, we implement an offline baseline using Implicit Q-Learning (IQL) [38]. To avoid confusing the two algorithms, we distinguish between the online data-collection policy—Independent Q-Learning (decentralized tabular IQL, Section 7.2)—and the offline benchmark—Implicit Q-Learning (in-sample neural IQL) evaluated here.

Architecture and Training Protocol: Following Kostrikov et al. [38], the offline IQL baseline operates on the 62-column nextstate_v2 transitions, framing station assignment as a per-candidate binary decision ($a\in \{0,1\}$). The architecture comprises a value network $V_{\psi }\left(s\right)$, twin action-value critics $Q_{{\theta }_1}(s,a),Q_{{\theta }_2}(s,a)$ with exponential moving average targets ($Q_{{\overline{\theta }}_1},Q_{{\overline{\theta }}_2}$, ${\tau }_{\mathrm{target}}=0.005$), and a stochastic categorical actor ${\pi }_{\varphi }\left(a\right|s)$. The value network is trained via asymmetric expectile regression ($L_2^{\tau }$) strictly under $\tau =0.7$:

$${\mathcal{L}}_V\left(\psi \right)={\mathbb{E}}_{(s,a)\sim \mathcal{D}}\left[|\tau -\mathbb{I}(Q_{\overline{\theta }}(s,a)-V_{\psi }\left(s\right)<0)|·{(Q_{\overline{\theta }}(s,a)-V_{\psi }\left(s\right))}^2\right]$$

(10)

where $Q_{\overline{\theta }}(s,a)=\min (Q_{{\overline{\theta }}_1}(s,a),Q_{{\overline{\theta }}_2}(s,a))$, evaluating value targets strictly within the distribution of logged transitions without querying out-of-distribution actions. The twin critics are optimized via mean squared temporal difference regression against the live value network:

$${\mathcal{L}}_Q\left({\theta }_i\right)={\mathbb{E}}_{(s,a,r,s^{\prime })\sim \mathcal{D}}\left[{(r+\gamma V_{\psi }\left(s^{\prime }\right)-Q_{{\theta }_i}(s,a))}^2\right]$$

(11)

with discount $\gamma =0.99$. Trajectory boundary conditions are strictly enforced: terminated transitions ($s^{\prime }$ absent) drop the bootstrap term and regress on immediate reward r, while right-censored/truncated transitions lacking valid subsequent epochs are excluded from the loss. Downstream policy extraction trains the actor via advantage-weighted maximum likelihood with inverse temperature $\beta =3.0$ and clipped exponential weights $\min (\exp \left(\beta (Q_{\overline{\theta }}(s,a)-V_{\psi }\left(s\right))\right),100.0)$.

Scope and Multi-Agent Limitations: In this benchmark, a single Implicit Q-Learning actor is fit per tier/fold (not six separately fitted per-station actors) and deployed as a decentralized binary decision model without centralized value factorization. Accordingly, this baseline does not perform multi-agent credit assignment across concurrent vehicles; factorized offline coordination (e.g., combining monotonic value factorization with in-sample or counterfactual regularization) remains an open challenge for future research.

Evaluation Budget, Feature Sets, and Preprocessing: Evaluation spans the 3 learned behavioral tiers and 2 feature sets (Physical and Context, identical to those defined for BC in Section 6.2), 7 held-out-seed folds each (42 evaluations total), fitted for 50 IQL epochs plus 50 FQE epochs per fold – a fixed protocol choice, not tuned post hoc, and distinct from the separate FQE numerical-stability audit (Section 7.8), which uses frozen actor checkpoints at epochs 100 and 150. Per fold, a scikit-learn StandardScaler is fit strictly on the six training seeds and applied unchanged to the held-out seed’s states and successor states; the categorical route label is one-hot encoded using the same procedure: fit on training data only, then transform. $\mathrm{bad}\mathrm{hbox}$ is the fraction of held-out test decisions where the extracted actor’s arg-max action matches the logged accept/reject decision; $\mathrm{bad}\mathrm{hbox}$ is FQE’s own estimate of the extracted actor’s expected discounted return.

7. Results and Discussion

This section distinguishes online collection-policy results (IQL and QMIX learnability and coordination), the Stage 3 greedy-evaluation cross-algorithm contrast, and offline learnability results (Behavior Cloning and Implicit Q-Learning) on the consolidated master dataset, following the protocol defined in Section 6. We report leave-one-seed-out Behavior Cloning and offline Implicit Q-Learning results below. An earlier pilot benchmark, evaluated on a legacy 4-feature, 27,353-transition dataset predating the wait-zone campaign, motivated the move to the master-dataset protocol below; it is reported in full in Supplement S1 and is not directly comparable to the results that follow (differing dataset scale, feature dimensionality, and validation split).

7.1. Multi-Tier Behavior Cloning Benchmark on the Consolidated Master Dataset

Table 7 reports the multi-tier Behavior Cloning results for behavioral Tiers 1–3 on the final $157071$-transition master dataset, following Section 6.

Imitation fidelity (Table 7) differs across behavioral arms and does not establish an ordering of operational quality – Section 8 notes Tier 3 has the lowest mean reward of the three despite the highest balanced accuracy. Trivial baselines (majority-class and train-prevalence random sampling) achieve $\approx 0.50$ by construction. Stratifying by operational mode reveals a performance gap not visible in the table: balanced accuracy drops during exploratory decision steps relative to exploitative steps in Tier 1 ($0.628$ vs $0.811$) and Tier 2 ($0.765$ vs $0.889$), indicating that matching the exploratory component is harder – a property of the logged decision task, not evidence of improved charging performance. Context provides a descriptive gain for Tiers 2 and 3 (positive across $7/7$ and $6/7$ folds respectively), while Tier 1 remains essentially unchanged; given the seven-seed sample size, this is a descriptive finding; neither statistical significance nor causality is established.

Nonlinear Context Predictability Probe: An out-of-sample nonlinear evaluation assessed whether spatiotemporal coordinates $(x,y,t_{\mathrm{dec}})$ improve prediction of the logged reward beyond local physical metrics, without inducing overfitting (full protocol and per-tier results in Supplement S11). Incorporating spatiotemporal context yielded a strictly positive out-of-sample $R^2$ improvement across all evaluated tier-by-held-out-seed folds. This finding provides observational evidence of local reward predictability only; consistent with the causal distinction between predictive relevance and causal feature attribution [39], it confirms that context features contain predictive information, but does not demonstrate improved global policy performance or higher off-policy returns.

Critical Interpretation Caveat: The decision-only BC learns accept/reject conditional on the retained station and its logged state. Without the other candidates’ states, it does not reproduce the policy’s full selection process. Scores describe fidelity to logged decisions, not operational policy performance. This benchmark confirms that logged sequential decisions across Tiers 1–3 contain learnable structure, motivating offline policy optimization via regularized architectures (such as Counterfactual Conservative Q-Learning, CFCQL-D [25]) as future work.

7.2. Multi-Seed Tabular IQL Learnability and Increasing Action-Value Separation

Beyond supervised imitation, we evaluate the learnability of the decentralized coordination problem using multi-agent Tabular Independent Q-Learning (IQL). We conduct an empirical evaluation across 7 independent seeds over 120 episodes, deploying a 4-regime non-stationary traffic curriculum (Light, Surge, Normal, Peak) rotated cyclically in 3-episode batches over 10 repetitions (30 cumulative episodes per regime). Both online policies use the unified power-law exploration schedule defined in Section 4.5, yielding an executed exploration rate of approximately $\epsilon \left(1\right)=0.1888$ at episode 1 and reaching the operational floor ${\epsilon }_{\min }=0.080$ at episode 120.

As summarized in Table 8, the policy confidence gap—defined as the absolute Q-value separation between admission actions $\left|Q\right(s,\mathrm{Accept})-Q(s,\mathrm{Reject}\left)\right|$—expands monotonically across all $7/7$ seeds, rising from an early baseline of $0.0838\pm 0.0028$ to a converged late average of $0.1442\pm 0.0094$ ($\Delta \in [+0.0498,+0.0717]$). This consistent widening indicates non-trivial temporal-difference learning: individual agents increasingly differentiate between acceptable and hazardous admission requests rather than retaining nearly equal action values ($\Delta \approx 0$).

However, examining the episodic reward trajectories stratified by traffic regime reveals a notable divergence. While rewards under Light traffic improve across seeds ($\Delta =+0.0133$, 4/7 seeds improving), performance consistently regresses under congested regimes: Normal ($\Delta =-0.0223$, 6/7 regressed), Peak ($\Delta =-0.0246$, 6/7 regressed), and Surge ($\Delta =-0.0236$, 7/7 regressed). A classical Dec-POSMDP coordination failure could explain this pattern, although the mechanism was not directly measured. Under severe congestion, charging stations may locally discover that rejecting incoming EVs is the least-penalized action ($Q(s,\mathrm{Reject})>Q(s,\mathrm{Accept})$), avoiding immediate saturation penalties ${\mathcal{P}}_{\mathrm{sat}}$, with uncoordinated stations collectively deflecting demand back onto the road network and suppressing macroscopic system rewards despite sharpened individual decision confidence. We did not measure per-station rejection coordination directly in this study; this remains a plausible mechanism, not an established one (Section 8).

7.3. State-Space Support and Cyber-Physical Boundary Dynamics

To evaluate dataset richness and structural coverage across the 7-dimensional cyber-physical observation space ($|{\mathcal{S}}_{\mathrm{expanded}}|=4608$ nominal combinations, $3456$ eligible states in equipped topologies), we perform a detailed state-support audit. The online IQL training logs contain $210306$ decision records across the seven seeds. The discrete state-support audit uses $210304$ records with valid encoded states; two runtime reservation-abort records have $\mathrm{bad}\mathrm{hbox}=-1$ and are excluded from this binning analysis. These online training records are distinct from the $157071$ curated offline transitions collected across the five behavioral arms. Table 9 reports the empirical stored support alongside the operational decision split ($a_t\in \{\mathrm{Accept},\mathrm{Reject}\}$) across all discretized observation bins.

The operational distributions in Table 9 characterize the recorded admission behavior:

• Monotonic Admission Tightening: Admission probability decays strictly as infrastructure becomes congested, falling from $91.6\%$ when chargers are idle ($\mathrm{load}\left[0\right]$) down to $26.3\%$ when all plugs are occupied ($\mathrm{load}\left[3\right]$). Similarly, when the upstream wait buffer is empty ($\mathrm{wait}\left[1\right]$), $79.5\%$ of vehicles are admitted, dropping sharply to $23.9\%$ once the buffer is partially occupied ($\mathrm{wait}\left[2\right]$).
• Activation of Critical Rare Bins: When pooled across all 7 seeds ($210304$ valid decisions, Table 9), total unique state coverage reaches $30.0\%$ ($1038/3456$ reachable combinations; per-dimension breakdown of this pooled coverage in Supplement S8). Crucially, the 4-stage curriculum with peak and surge regimes reaches previously unvisited cyber-physical boundary states: pooled support for severe queues ($\mathrm{queue}\left[2\right]$) reaches $121/1152$ states ($10.5\%$, comprising 382 decisions, representing a $\times 13$ to $\times 60$ expansion over legacy low-flux baselines), while buffer saturation ($\mathrm{wait}\left[3\right]$) attains $23/1152$ states ($2.0\%$, comprising 27 decisions, $\times 11.5$ gain). This indicates that the IoEV-MARL dataset activates severe stress conditions relevant for robust offline policy regularizers, within the online training history examined here (Section 8 notes this coverage figure is not the released corpus’s own).

7.4. Cooperative Value Factorization: QMIX Learning Dynamics and Spatial Load Balancing

To investigate whether centralized value factorization mitigates the reward regression observed under independent tabular learning, we evaluate monotonic deep multi-agent reinforcement learning via QMIX [34]. Each of the $N=6$ per-station value heads conditions on a continuous 15-dimensional local observation vector ${\mathbf{o}}_t^i\in {\mathbb{R}}^{15}$ (encompassing local SoC profiles, spatial arrival distances, charger occupancy, and station identity), while a centralized mixing network parameterized by hypernetworks aggregates individual utilities $Q_i({\tau }^i,a^i)$ into a joint action-value $Q_{\mathrm{tot}}(\mathbf{s},\mathbf{a})$ conditioned on the global road-network state $\mathbf{s}\in {\mathbb{R}}^{27}$; this factorization scores the candidate pool for the single EV-level accept/reject decision defined in Section 3; it does not introduce N separately logged agents.

Crucially, because QMIX optimizes over continuous observation spaces, discrete state-space coverage metrics (such as the tabular $3456$-state scope) are mathematically inapplicable. Instead, we evaluate QMIX using corresponding empirical criteria: increasing action-value separation across seeds, episodic reward progression across the 4-stage non-stationary traffic curriculum, behavioral admission projections on physical variables, and macro-level spatial coordination metrics.

Reward Progression Contrast Under Heavy Traffic Congestion.Table 10 shows positive QMIX reward deltas in all seven seeds under every traffic regime, contrasting with tabular IQL’s predominantly negative deltas under congestion (Section 7.2). This is a descriptive association between architecture and reward trajectory, not a controlled ablation isolating the mixing network: discount factor and network capacity/representation also differ (Section 8).

Spatial Load Balancing. A candidate contributing factor is illustrated by the spatial load variance metric ($\mathrm{bad}\mathrm{hbox}={\sigma }_{\mathrm{load}}/{\mu }_{\mathrm{load}}$ across the 6 charging stations, per-episode trajectory plotted in Supplement S9). Across all 7 seeds, the spatial coefficient of variation decreases substantially (${\Delta }_{\mathrm{cv}}=-0.0557$ overall, with $7/7$ seeds improving balance under Light, Normal, and Surge). One possible, unverified explanation is that the mixing network, conditioned on global state $\mathbf{s}$, penalizes localized bottlenecks and favors demand routing toward underutilized peripheral hubs; we did not measure per-station routing directly. Note that $\mathrm{bad}\mathrm{hbox}$ is not itself a term of $r_t$ (Equation (3)). Its cooperative bonus $r_{\mathrm{coop}}={\lambda }_{\mathrm{coop}}·({\mu }_{\mathrm{load}}-{\mathrm{load}}_i)/\mathrm{capacity}$ is computed from the same underlying per-station load distribution that $\mathrm{bad}\mathrm{hbox}={\sigma }_{\mathrm{load}}/{\mu }_{\mathrm{load}}$ summarizes, which justifies caution about reward/evaluation-metric independence. However, $r_{\mathrm{coop}}$ is not a monotonic function of CV: at capacity 3, load vectors $(1,1,1,1,1,1)$ and $(1,0,0,1,2,2)$ give identical $r_{\mathrm{coop}}=0$ but very different CV. Shared inputs therefore do not guarantee that optimizing one quantity improves the other.

Non-Stationarity as a Possible Explanation for IQL Reward Regression. Independently updating, decentralized tabular learners (Tabular IQL) are well known in the multi-agent RL literature to suffer from non-stationarity: because every other agent’s policy is also changing during training, the transition distribution perceived locally by any single learner is not stationary, which invalidates the standard convergence guarantees of single-agent Q-learning [40,41,42]. QMIX’s monotone centralized factorization is different: within its representable value class it is sufficient for Individual-Global-Max (IGM) consistency ($\partial Q_{tot}/\partial q_i\ge 0$), though this still does not by itself guarantee recovery of the true optimal joint policy. This qualitative hypothesis may explain the reward regression under severe traffic, but establishes neither convergence to a specific suboptimal policy nor the absence of non-convergence. We do not report a quantitative coordination-breakdown figure here, since no such measurement is tabulated in this study; a dedicated per-station flow analysis, left to future work, would be needed to test this explanation directly (formal non-stationarity argument in Supplement S11).

Behavioral Divergence at Saturation Boundaries. Projecting both policies onto identical physical observation bins (Table 11) exposes key structural differences between decentralized and factorized decision-making:

• Higher Admission Rate at $\mathrm{load}\left[2\right]$: When two dispensers are occupied, tabular IQL admits only $42.6\%$ of requests, versus QMIX’s $76.3\%$. This admission-rate difference does not demonstrate maximized throughput – we have not measured realized plug utilization or completion rates conditional on this admission gap.
• Strict Gating at Saturation ($\mathrm{load}\left[3\right]$): Once total charger saturation occurs, QMIX sharply reduces admission, admitting only $14.6\%$ (rejecting $85.4\%$) compared to IQL’s $26.3\%$ admission rate. Coupled with the spatial-balancing association noted above, this descriptive admission-rate contrast is consistent with tighter queue control at saturation, though we have not directly measured idle-capacity rates to support a stronger claim.
• Higher Queue Acceptance and Stricter Admission at Full Charger Occupancy: QMIX is systematically more conservative than IQL across SoC and distance tiers, but has a higher observed acceptance rate in non-empty-queue bins ($97.7\%$ vs $89.7\%$ at $\mathrm{queue}\left[2\right]$). These marginal projections do not identify the joint load–queue dependence or establish a cross-station coordination mechanism (per-bin accept/reject distribution plotted in Supplement S9).

7.5. Deterministic Greedy Policy Evaluation (Stage 3 Operational KPIs)

To evaluate policy exploitation without exploration noise, converged policy snapshots (Tabular IQL Q-tables and Neural QMIX weights) are deployed in deterministic exploitation mode ($\epsilon =0.0$) using setup_eval_epsilon0_waitzone_campaign_fixaudit_20261015.py. While deterministic execution removes exploration-induced action stochasticity, it serves as an empirical exploitation baseline rather than a theoretical performance upper bound. The evaluation spans 280 real simulation episodes across the 4 non-stationary traffic regimes (5 evaluation episodes per regime × 7 disjoint random seeds1), encompassing $33260$ IQL and $33318$ QMIX controlled EV trips ($66578$ vehicle trips evaluated across regimes). All decisions and trips belong strictly to the controlled EV fleet (veh_MARL_*), ensuring that macro-level metrics are never diluted by background traffic.

Table 12 reports the four core operational Key Performance Indicators: charging success rate (reaching target SoC without operational abort), travel delay (SUMO’s standard timeLoss metric relative to free-flow velocity), mean station charger occupancy, and inter-hub spatial load variance ($\mathrm{bad}\mathrm{hbox}$).

Inter-Seed Dispersion.Table 12 reports a pooled aggregation across all 35 episodes (7 seeds × 5 episodes) per algorithm/regime. Table 13 reports the complementary per-seed view: the mean across the 7 seeds of their respective 5-episode averages (± sample SD, $n-1$ divisor) – a different quantity from the pooled aggregation above (the two agree closely here because episode counts per seed/flux are balanced, but are not generally interchangeable). All 280 retained episodes are used, including the 2 QMIX surge episodes (seeds 2 and 7, episode 10) that originally exceeded the 1200 s watchdog and were deterministically re-executed to completion (footnote, Table 12); recomputing the pooled values from these same 280 episodes reproduced Table 12 to within rounding.

The paired per-seed QMIX−IQL difference in $\mathrm{bad}\mathrm{hbox}$ (same seed, same flux) is negative (QMIX lower) in $2/7$, $7/7$, $4/7$, and $4/7$ seeds for Light, Normal, Peak, and Surge respectively. Normal is the only regime with unanimous per-seed agreement with the pooled-aggregation contrast; Peak and Surge show a near-even split across seeds despite their pooled means both favoring QMIX, and Light – where the pooled means already favor IQL – has only $2/7$ seeds where QMIX is lower, showing that the pooled four-regime summary in Table 12 masks substantial per-seed heterogeneity in 3 of the 4 regimes. Surge shows the largest single-seed dispersion in IQL delay ($\pm 2.72$ s, versus $\le 0.65$ s for the other three regimes), consistent with higher variance in IQL behavior under the most congested curriculum stage.

The empirical evaluation in Table 12 supports two observations:

• Completion Rates under Increasing Traffic: Under pure greedy execution, both decentralized IQL and factorized QMIX achieve closely matched completion rates ($28.7\%--29.7\%$ in Light down to $13.1\%--13.7\%$ in Surge), with minimal variance in trip delay ($25.8--29.5\mathrm{s}$). The Charging + Reserved at Decision Epochs column ($2.84$–$2.94$ out of 3) is reserved_count + charging_count sampled at decision instants (verified in analyze_stage3_greedy_eval_kpis_20261016.py), i.e., charging-plus-reserved occupancy at the moments decisions are made, not a continuously-measured active-plug utilization rate; it should not be read as evidence that chargers are busy $\approx 95$–$98\%$ of wall-clock time. The low completion percentages under high volume are consistent with capacity constraints – network demand exceeding charging supply. The all-vehicle unserved rate and decision-sampled occupancy do not establish this causal mechanism; they are consistent with a structural ceiling also observed with unconstrained heuristics. On the released Tier 0 campaign, NEAREST and WEIGHTED leave the large majority of SUMO-injected vehicles without a successfully-initiated charging session; the exact unserved rate, its pre-fix comparison, and the corrected admission filter’s measured effect are detailed in Supplement S7.
• Spatial Load Balancing Under Congestion, with Substantial Per-Seed Heterogeneity: The training-time reduction in spatial load dispersion only partly persists under deterministic greedy evaluation. The pooled contrast favors QMIX in Normal, Peak, and Surge, but only Normal shows agreement across every seed (per-seed analysis above). These findings do not establish a uniformly robust benefit or isolate the mixing network as its cause (Section 8).

7.6. Offline Value Benchmarking via In-Sample Implicit Q-Learning (IQL)

Following the protocol defined in Section 6.3 (architecture, training, and distinction in scope from the online tabular IQL collection policy), we report Implicit Q-Learning [38] results on the consolidated dataset below.

Results:Table 14 reports leave-one-seed-out (LOSO) results across the 3 learned behavioral tiers and 2 feature sets, following the evaluation budget, feature sets, and train-only preprocessing defined in Section 6.3.

Tier 1 shows the lowest score in both the actor-match-rate (this table) and Behavior Cloning’s balanced-accuracy (Section 6.2) evaluations, despite the two being independently fit models – partial agreement between methods, without an identical overall ranking: Tier 2 exceeds Tier 3 on IQL match rate ($0.905$ vs. $0.891$ physical, $0.923$ vs. $0.922$ context), unlike BC’s own Tier 2/3 ordering. The FQE value ordering (Tier 1 > Tier 2 > Tier 3) tracks each tier’s raw mean logged reward exactly (Section 8’s quality-tier-labels discussion: $+0.115$, $+0.015$, $-0.069$ respectively). The observed FQE ordering is consistent with the tiers’ mean logged rewards, but neither that agreement nor the reward means alone validate the estimates or establish a ranking of policy performance; FQE values should not be compared across tiers as a ranking of learned policy quality. A neural fit that converges with finite losses demonstrates numerical training stability on this dataset; it does not by itself establish Bellman-operator contraction in the formal sense, nor closed-loop deployment performance.

Benchmark Provenance Verification. The Tier 0 admission-filter correction leaves the Tier 1–3 data used by the reported BC and offline-IQL benchmarks unchanged; recorded provenance and column-level comparisons supporting this verification are documented in Supplement S7.

7.7. Susceptibility to Out-of-Distribution (OOD) Maximization Bias

Standard off-policy value iteration algorithms (e.g., standard multi-agent Q-learning or unconstrained QMIX) perform temporal difference updates using the target operator ${\max }_{{\mathbf{a}}^{\prime }}{Q}_{tot}(s_{t+1},{\mathbf{a}}^{\prime })$. In static offline datasets where state-action coverage is structurally bounded by behavioral data tiers, evaluating values on counterfactual joint actions $\left({\mathbf{a}}^{\prime }\right)$ unobserved in the logged buffer is well documented in the offline and conservative RL literature [24] to risk Out-of-Distribution (OOD) Maximization Bias: an unconstrained critic can overestimate the value of unvisited joint actions, which can lead to policy collapse upon deployment. In cooperative EV charging, this would manifest as mixing networks overestimating the value of aggressive, uncoordinated admissions never seen in the logged buffer. We have not measured this failure mode directly on this dataset; it is stated here as a known theoretical risk that can be investigated using this testbed, not as an empirical finding of this paper.

Future offline MARL benchmarks could investigate the following approaches:

• Conservative Multi-Agent Regularization: Penalizing out-of-distribution joint actions via conservative value estimation. While monolithic conservative objectives extend single-agent CQL [24] by penalizing unobserved joint actions, factorized multi-agent architectures such as Counterfactual Conservative Q-Learning (CFCQL-D [25]) calculate decentralized conservative penalties over individual utilities under credit assignment, mitigating exponential scaling while bounding counterfactual extrapolation in prospective offline MARL.
• In-Sample Value Factorization (Multi-Agent IQL): Avoiding out-of-sample maximization entirely by using expectile regression ($L_2^{\tau }$) on the logged support, evaluating value targets strictly within the distribution of observed data.

7.8. Empirical Stability of Off-Policy Value Estimation

Action-Masking Divergence in Physical Off-Policy Evaluation: Beyond OOD maximization bias, offline evaluation in cyber-physical systems faces a separate structural problem: standard off-policy value estimators such as Fitted Q-Evaluation (FQE) are typically fit on an actor’s unconstrained policy, while physical deployment imposes hard admission filters (action masks) absent from that estimation. In a targeted engineering validation (two single-seed pilot episodes under a frozen station-selector with a BC- or IQL-gated actor), such infeasibility was not rare: the hard feasibility constraint was active in 24–25% of real decisions. For the stochastic IQL actor, the deployed, mask-constrained action distribution diverged from the actor’s own raw policy output in 100% of these infeasible cases, with one observed decision assigning 68.8% raw probability to an action the physical constraint forces to 0%. These figures come from a small, non-exhaustive validation rather than a full-corpus statistical estimate. They nonetheless indicate that evaluating constrained deployment policies via standard OPE requires formally incorporating the deployment-time action mask into the offline Bellman operator – an open methodological question that this dataset can help investigate, but that this paper does not resolve.

Empirical Stability and Value-Estimate Drift in Fitted Q-Evaluation (FQE): To evaluate the numerical stability of off-policy value estimation on the master transitions, an empirical sensitivity audit spanning 252 evaluation branches was conducted. These branches span behavioral tiers, frozen actor checkpoints (epochs 100 and 150), 7 cross-validation folds, 3 repetitions, and two optimization learning rates (representing evaluation grid variations rather than 252 independent dataset replications), evaluated under an absolute mean per-state value-change threshold of $0.02$ applied separately to the 200–250 and 250–300 epoch windows. Reducing the learning rate from $\eta =3·{10}^{-4}$ (control) to $\eta =1·{10}^{-4}$ substantially reduced value-estimate drift across evaluation epochs, increasing the proportion of numerically stable evaluation branches from 9 of 126 ($7.1\%$) to 94 of 126 ($74.6\%$). However, while value estimates stabilized reliably on exploratory transitions (Tier 1), FQE runs on the expert regime (Tier 3, $\epsilon =0.0$) exhibited persistent variance. This greater sensitivity may be associated with restricted action support under near-deterministic targets; the causal mechanism remains unverified. Crucially, reduced temporal value-estimate drift reflects numerical optimization stability rather than certified evaluation accuracy or dynamic closed-loop return in SUMO; off-policy value estimates must be interpreted strictly as conditional diagnostics rather than substitutes for measured performance.

8. Limitations

The following limitations concern the released data, the interpretation of the results, and their generalizability.

• Candidate sets are not exported. As stated in Section 3, the released transition contract logs the accept/reject gate on a single, already-selected candidate station per vehicle decision; the alternative candidate stations considered but not selected are not exported in the released tuples. Consequently, the upstream station-selection sub-decision itself cannot be learned or evaluated offline from this release – only the downstream binary gate can be.
• Hypothesized, not proven, non-stationarity explanation. The Discussion in Section 7.4 offers independent-learner non-stationarity as a candidate qualitative explanation for the IQL reward regression under severe traffic; this is a hypothesis grounded in established multi-agent RL theory [40,41,42], not a measurement performed on this dataset. Testing this hypothesis requires the per-station flow analysis identified in the Discussion.
• Shared Load Inputs Do Not Make the Cooperative Bonus a Monotonic Function of Load-Balance Variance (Section 7.4 gives the counterexample). This justifies caution about reward/evaluation-metric independence; the reported improvement in $\mathrm{bad}\mathrm{hbox}$ under QMIX remains a descriptive association sharing the same underlying measurements, neither causally established nor mathematically guaranteed by the reward design.
• Behavior-policy "quality tier" labels do not track measured return. Tier 3 (near-deterministic, $\epsilon =0$, labeled as the most converged behavioral arm) has the lowest mean per-decision reward of the three learned tiers ($-0.069$, versus $+0.115$ for Tier 1 and $+0.015$ for Tier 2, computed directly on the released corpus). No causal mechanism for this ordering (e.g., an interaction between accept rate and network saturation under low exploration) is established in this paper; readers should not infer a monotonic quality ordering from the tier labels alone.
• State-space coverage figures in this paper describe online IQL training logs, not the released offline corpus. The $30.0\%$ ($1,038/3,456$) coverage figure in Table 9 is computed on $210,304$ online IQL training decisions, not on the $157,071$-row released dataset; the released, per-tier coverage figures (verified directly on the released corpus) are $15.7\%$/$44.7\%$/$40.4\%$ for Tiers 1/2/3 respectively. Caution: state_id uses a different codec per tier (Section 4.3’s data dictionary) – IQL’s and QMIX’s IDs are not directly comparable and must never be pooled across tiers without converting through each tier’s own codec first.
• Topology-specific representation of charging heterogeneity. QMIX’s shared local Q-network receives a six-slot one-hot station identity, but no explicit charging-power or efficiency feature. On the fixed topology, this representation allows station-specific behavior to be learned from observed outcomes. However, the experiments do not isolate sensitivity to charging power from other station-specific effects, nor establish transfer to different station configurations. This limits interpretation and generalization claims without invalidating the measurements reported for the evaluated topology.
• FQE accuracy is not established by the reported numerical-stability audit. A comparison with finite-prefix returns from logged behavior trajectories would additionally require matching the evaluated policy, initial-state cohort, and treatment of censored tails. The current findings should therefore not be interpreted as validated errors relative to closed-loop policy returns.

9. Conclusions and Future Work

This paper introduced the IoEV-MARL dataset and benchmark for offline decision-learning in simulated EV charging coordination. By extracting $157071$ empirical Semi-Markov Decision Process (SMDP) transitions across a 700-episode collection curriculum (five behavioral arms, seven disjoint random seeds, four non-stationary traffic regimes), we established a reproducible foundation for offline MARL research. A three-layer schema separates decision-time observations from outcome and evaluation fields. Seven-fold leave-one-seed-out evaluation keeps simulation seeds disjoint between training and testing.

Detailed state-space audits confirm that the curriculum activates critical congestion and buffer-saturation boundaries, with the released corpus’s per-codec coverage differing materially from the pooled online training coverage reported in the body (Section 8) – a distinction that matters for anyone using the released data directly. Multi-seed behavioral benchmarking found a reward regression in decentralized tabular IQL under severe traffic, potentially explained by independent-learner non-stationarity, as discussed above. Centralized monotonic value factorization (QMIX), by contrast, showed positive reward progression and improved spatial load balancing across all regimes. This contrast shows substantial per-seed heterogeneity under deterministic greedy evaluation (Section 7.5), however, and should not be read as a uniformly robust effect. Having benchmarked decision-prediction fidelity via Behavior Cloning and offline Implicit Q-Learning (Table 14), future work will use IoEV-MARL to benchmark regularized and centralized offline MARL algorithms (such as CFCQL-D [25]) to mitigate out-of-distribution maximization bias and evaluate whether learned offline policies can outperform the behavioral data-collection policies in deployment. This will require matched closed-loop SUMO evidence, which was not collected in this paper.