The paper introduces a ground‑truth‑first benchmarking framework for reinforcement learning that uses converse optimality to construct tasks with certified optimal value functions and policies. It unifies sensitivity benchmarking—sampling certified environments from continuous‑time generators—and oracle‑referenced comparison—using native discrete‑time generators to evaluate policies via optimality gap, regret, and policy error. The methodology extends to non‑smooth, non‑affine, and partially observed settings, and the authors provide code and diagnostics on a public project page.
By Gr\'egoire Ouerdane, Sinan Ibrahim, Stefan Streif, Pavel Osinenko
Limiting‑Kernel Q(λ) (LKQL) is an off‑policy value estimator that blends n‑step truncation with a long‑horizon approximation based on the limiting kernel. It maintains the computational efficiency of n‑step methods while improving policy evaluation accuracy, especially for long‑horizon tasks. The authors prove faster convergence of LKQL’s operator under aperiodicity and near‑on‑policy conditions, and demonstrate empirical gains on MuJoCo continuous‑control benchmarks.
By Tolga Ok, Arman Sharifi Kolarijani, Peyman Mohajerin Esfahani, Mohamad Amin Sharifi Kolarijani
The paper introduces Solver-Gradient Guided Reinforcement Learning (SG‑RL), a method that augments standard RL with bounded gradients from a differentiable MPC solver to adapt cost‑function weights online. SG‑RL integrates solver‑gradient guidance into PPO through actor‑update scaling, policy loss, advantage estimation, and value‑function learning, achieving comparable or superior closed‑loop performance while requiring up to 70.6% fewer samples. Experiments on two autonomous racing platforms with intentional model mismatch demonstrate that SG‑RL outperforms both RL and gradient‑based policy learning baselines and generalizes zero‑shot to unseen environments.
By Baha Zarrouki, Arslan Thobani, Jasper Hoffmann, Mattia Piccinini, Rudolf Reiter, Felix Jahncke, S\'ebastien Gros, Davide Scaramuzza, Johannes Betz
The paper introduces Network Feasibility Geometry Reinforcement Learning (NFG‑RL), a method that enforces multi‑layer network constraints—such as interference, power‑rate coupling, flow conservation, service chains, capacity, latency, and reliability—by transporting a proto‑policy through a differentiable feasibility map. By compiling heterogeneous constraints into typed residual blocks and using a variational transport operator, NFG‑RL ensures almost‑sure feasible execution and shapes exploration and gradients to respect active constraints. Experiments on two wireless‑edge surrogate environments show that NFG‑RL boosts feasible utility by 37.5–41.5 %, cuts raw‑action violations by 48.5–60.8 %, and reduces P99 delay by 57.0–75.5 % compared to leading baselines.
By Zuyuan Zhang, Zeyu Fang, Mahdi Imani, Nathaniel D. Bastian, Tian Lan
In value-based reinforcement learning, improving the accuracy of policy evaluation has been shown to improve downstream policy optimization performance. The widely adopted family of approximations rel...
The paper introduces Policy Gradient Penalty (PGP), a single‑loop policy‑space method that enforces convex occupancy‑measure constraints via quadratic‑penalty regularization. PGP constructs pseudo‑rewards to estimate gradients of the penalized objective and uses the classical Policy Gradient Theorem, establishing smoothness and global last‑iterate convergence guarantees for an ε‑optimal constrained entropy value with ε‑bounded constraint violation. The authors validate PGP with ablations on a grid‑world benchmark and demonstrate scalability on two challenging continuous‑control tasks.
By Florian Wolf, Ilyas Fatkhullin, Niao He
arXiv:2610.35748v1 Announce Type: new
Abstract: Model selection and hyperparameter tuning pose a central challenge in off-policy evaluation (OPE) because errors of fitted Q-functions and occupancy-ra...
By Lars van der Laan, Vasilis Syrgkanis, Nathan Kallus
The paper introduces a model‑free continuous‑time reinforcement learning algorithm that learns from a fixed batch of discrete‑time off‑policy data. It builds on the PhiBE‑Q framework to create a PhiBE‑PI policy iteration that uses a single state‑action critic, avoiding explicit drift or diffusion estimation. The authors provide a local end‑to‑end suboptimality guarantee under linear function approximation, analyze error propagation from discretization, function approximation, and finite samples, and reveal a bias‑variance trade‑off where finer time discretization can worsen statistical error.
By Chaorui Yao, Yuming Paul Zhang, Lin Yang, Yuhua Zhu
The paper introduces BUMEX, a reinforcement learning exploration strategy that leverages a set of prior models containing the true transition kernel and reward function. By optimizing over this model set, the method derives upper and lower bounds on the Q‑function to guide exploration, providing theoretical guarantees of convergence to the optimal policy. When the model set follows a bounded‑parameter MDP structure, the optimization becomes convex, enabling finite‑time convergence under mild assumptions and demonstrating accelerated learning in simulations.
By J. S. van Hulst, W. P. M. H. Heemels, D. J. Antunes
The paper introduces a model-based bootstrap framework for uncertainty quantification in offline policy evaluation (OPE) within finite-horizon, time-inhomogeneous Markov decision processes. Unlike traditional bootstrap methods that resample entire episodes, this approach regenerates trajectories from an estimated MDP, enabling use of diverse offline data formats such as complete trajectories, transition-level observations, and trajectory fragments. The authors prove bootstrap distributional consistency, asymptotically valid confidence intervals, and consistent variance estimation, and demonstrate through simulations that the method yields tighter confidence intervals and more accurate variance estimates compared to existing techniques.
By Weiwei Wang, Yuqiang Li, Xianyi Wu, Bingyi Jing
The paper presents the first provably efficient offline reinforcement learning algorithms for risk-sensitive settings using optimized certainty equivalents (OCE). It introduces two formulations—dynamic-OCE, applying OCE at each step, and static-OCE, applying a single OCE to cumulative rewards—and develops pessimistic value-iteration methods with tight sublinear suboptimality bounds in tabular and linear function approximation settings. The authors also establish minimax lower bounds, characterizing the fundamental limits of offline risk-sensitive RL.
By Hanzhen Wang, Boxiang Lyu, Zhongxiang Dai, Siyang Gao, Dake Zhang, Shuang Qiu
arXiv:2604. 26360v2 Announce Type: replace-cross Abstract: Reinforcement learning from human feedback (RLHF) systems face a compounding alignment challenge: not only are learned reward models uncertain about unseen state-action pairs, but the human preference annotations they are trained on are themselves inconsistent, context-dependent, and noisy.
By Disha Singha