Downloads · 30 days
0
SidraBhatti/xi-ddqn-offroad-energy-management
xi-ddqn-offroad-energy-management is a reinforcement learning model from SidraBhatti. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
An explainable, interpretable Double Deep Q-Network (XI-DDQN) controller for energy management in high-power (350 kW) series-hybrid off-road vehicles, trained on real experimental duty cycles. Addresses fuel optimizat…
Downloads · 30 days
0
Access
Public
Updated Sep 12, 2026
Repo size
243 KB
Likes
0
Public
Click a slice to open those files.
.png243 KB · 97%
From the Hugging Face model README
An explainable, interpretable Double Deep Q-Network (XI-DDQN) controller for energy management in high-power (>350 kW) series-hybrid off-road vehicles, trained on real experimental duty cycles. Addresses fuel optimization under uncertain, variable operating conditions typical of off-highway applications — a configuration with far less RL/EMS research attention than on-road hybrids or lower-power off-road platforms.
Overall XI-DDQN framework: three-head Q-network (fuel, SOC, battery-power
heads) selects battery power actions from state (SOC, P_req), evaluated
against a target network and trained from a replay buffer, controlling a
series-hybrid powertrain (engine–generator, battery, PDU, electric machine)
against a real driving cycle.
Sidra G. Bhatti led the RL/AI methodology: design and implementation of the powertrain constraints-aware, explainable/interpretable DDQN controller, reward shaping, and training pipeline. Full author list (order per the published paper): Hend Abououf, Athar Hanif, Sidra G. Bhatti, Qadeer Ahmed, Scott A. Shearer, Nitish Chandramouli, Jon A. Dickson — OSU Center for Automotive Research (Mobility Systems Lab), with Cummins Inc. co-authors (Chandramouli, Dickson) reflecting the industry partnership.
State & action space. The agent observes a 2D state: battery state of charge (SOC ∈ [0.15, 0.85]) and normalized power demand. It selects from 2000 discretized battery power setpoints (±200 kW); generator power is computed from the power-balance residual, with fuel consumption drawn from a real engine-generator map.
Three-head DDQN. A shared feature trunk (Linear→ReLU ×3, 128-dim
bottleneck) branches into three independent Q-heads — Q_fuel, Q_soc, and
Q_battery — each trained on its own reward component via a weighted-sum
MSE loss. Action selection combines all three heads
(Q_total = Q_fuel + λ_soc·Q_soc + λ_battery·Q_battery), so the controller
jointly reasons about fuel economy, SOC regulation, and battery-power
constraint satisfaction rather than optimizing a single scalarized objective.
Reward shaping.
| Term | Weight | Purpose |
|---|---|---|
| Fuel consumption | 10 | Penalizes generator fuel use directly |
| SOC deviation (quadratic, +in-band bonus) | 5.0 | Keeps battery SOC near the 0.55 target |
| Battery-power constraint violation | 0.001 | Penalizes exceeding hardware/energy limits |
Training hyperparameters.
| Parameter | Value |
|---|---|
| Discount factor (γ) | 0.95 |
| Target network sync | every 20 episodes |
| Exploration (ε) | 0.5 → 0.01, decay 0.9995/episode |
| Optimizer | Adam, lr = 1e-4 |
| Replay buffer size | 200,000 transitions |
| Batch size | 64 |
| Curriculum | multi-phase, sequential SOC operating ranges, exploration reset each phase |
The paper's explainability contribution distills the trained (black-box) DDQN policy into a shallow, human-readable decision tree that predicts the same control action (battery power P_batt) from the same two state inputs (SOC, power demand). This produces:
| Comparison | Outcome |
|---|---|
| vs. conventional (non-hybrid) powertrain | ~23% fuel consumption reduction |
| vs. rule-based control strategy | ~11.7% average reduction on unseen real-world duty cycles |
| vs. dynamic programming (offline optimum) | Captures DP-level optimal control behavior while generalizing to uncertain/unseen conditions |
| Decision-tree distillation fidelity | High fidelity to the DDQN's own action choices — the learned policy is largely explainable by a simple SOC/power-demand rule structure rather than requiring an opaque network at deployment |
Training requires experimental duty-cycle and engine-generator map data (MAT files) that are not publicly released — this repo/page describes the method and results; reproduction requires access to equivalent experimental or sponsor-provided data.
Published: Energy (Elsevier), 2026 — doi.org/10.1016/j.energy.2026.141083 Open-access preprint: SSRN doi.org/10.2139/ssrn.6299023 Code: github.com/OSU-CAR-MSL/Powertrain-Aware-Explainable-Reinforcement-Learning-Energy-Management-in-Off-Road-Vehicles Funding: U.S. Department of Energy; Cummins Inc.
@article{abououf_xiddqn_2026,
author = {Abououf, Hend and Hanif, Athar and Bhatti, Sidra G. and Ahmed, Qadeer and Shearer, Scott A. and Chandramouli, Nitish and Dickson, Jon A.},
title = {Powertrain-aware explainable reinforcement learning energy management in off-road vehicles},
journal = {Energy},
year = {2026},
doi = {10.1016/j.energy.2026.141083}
}