RL Theory & Markov Decision Processes
Empirical Robustness and Scalability Evaluation of Provably Efficient RL Algorithms
Generated automatically from the limitations stated in 10 papers (ICML, AISTATS, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
A wide range of theoretical reinforcement learning algorithms offer provable sample-complexity guarantees, yet their empirical validation is almost exclusively restricted to synthetic tabular MDPs with fewer than 30 states. Because these methods have not been evaluated across larger, standardized, or higher-dimensional environments, it remains unknown whether their theoretical efficiency properties hold in practice. Without systematic cross-domain testing, the community lacks evidence on whether provably efficient exploration and representation mechanisms scale beyond toy sandbox problems.
Why it matters
Provides the first systematic empirical baseline determining whether theoretical RL algorithms retain their sample-efficiency guarantees in non-toy settings. Establishes empirical benchmarks that identify practical computational and statistical failure modes in theoretical algorithm design.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Implement a representative set of theoretical RL algorithms (such as reward-free exploration, low-rank MDP algorithms, and optimistic value iteration variants) and evaluate them across parameterized benchmark environments (e.g., scalable bsuite tasks and synthetic MDPs with state spaces varying from $10$ to $10^4$) to measure empirical sample efficiency and computational runtime as dimensions grow.
- 2
Evaluate standard empirical relaxations (e.g., linear and kernel feature representations) of these theoretical methods against standard empirical RL baselines across continuous-state control benchmarks to measure policy performance and sample complexity gaps.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Several theoretical algorithms rely on operations (such as optimal design over entire policy spaces) whose computational complexity is fundamentally exponential in the horizon or state size, making execution on larger benchmarks intractable without altering the underlying algorithms.
Evidence
Each paper's own statement of the limitation, verbatim.
- Reinforcement Learning with Segment FeedbackICML 2025
Sum-feedback algorithms are computationally inefficient (E-optimal design over all policies), so experiments used a tiny MDP with |S|=3, |A|=5, K=1000.
- How to Find the Exact Pareto Front for Multi-Objective MDPs?ICLR 2025
Experiments limited to tiny synthetic MDPs (S≤8, A≤7, D≤4); scalability to realistic state spaces is not demonstrated
- Concurrent Reinforcement Learning with Aggregated States via Randomized Least Squares Value IterationICML 2025
Numerical experiments are on small random MDPs (≤30 states/actions) with deterministic rewards, not large-scale or benchmark domains
- Resilient Constrained Reinforcement LearningAISTATS 2024
Experiments are limited to small tabular MDPs (e.g., 20 states, 5 actions); no function approximation or large-scale domains are studied.
- Model-free Low-Rank Reinforcement Learning via Leveraged Entry-wise Matrix EstimationNeurIPS 2024
Numerical experiments are only on synthetically generated low-rank MDPs (in appendix), with no real-world evaluation
- Provably efficient representation selection in Low-rank Markov Decision Processes: from online to offline RLUAI 2023
Experiments are only on small synthetic MDPs (|S|=20, |A|=3, H=10, d=5); no evaluation on real or deep RL benchmarks
- Near-optimal Conservative Exploration in Reinforcement Learning under Episode-wise ConstraintsICML 2023
Theory and experiments are limited to small synthetic tabular MDPs (S ≤ 5, A ≤ 5, H ≤ 3); scalability to large state spaces is untested.
- Krylov–Bellman boosting: Super-linear policy evaluation in general state spacesAISTATS 2023
Experiments are small sandbox problems (200-state tabular chain, 3-dimensional nonlinear system); no evaluation on large-scale or realistic MDPs
Show all 10 papers
- Finding Safe Zones of Markov Decision Processes PoliciesNeurIPS 2023
Empirical evaluation is limited to a single small grid-world (N=30, horizon 300) with no real-world domains, so practical performance on large MDPs is unverified
- A Tale of Sampling and Estimation in Discounted Reinforcement LearningAISTATS 2023
Empirical validation is limited to a simple illustrative 3-state Markov chain with a hand-picked function f = (1,-1,2); no evaluation on realistic RL tasks
Nearest existing work
- Optimistic Natural Policy Gradient: a Simple Efficient Policy Optimization Framework for Online RLNeurIPS 2023
- Scalable Policy-Based RL Algorithms for POMDPsNeurIPS 2025
- Replicable Reinforcement LearningNeurIPS 2023
- Q-Star Meets Scalable Posterior Sampling: Bridging Theory and Practice via HyperAgentICML 2024
- Sample Efficient Reinforcement Learning In Continuous State Spaces: A Perspective Beyond LinearityICML 2021
- Sample-Efficient Reinforcement Learning Is Feasible for Linearly Realizable MDPs with Limited RevisitingNeurIPS 2021
- Provably efficient reinforcement learning with linear function approximationCOLT 2020
- Stabilizing Q-learning with Linear Architectures for Provable Efficient LearningICML 2022
- Behaviour Suite for Reinforcement LearningICLR 2020
- A Closer Look at Offline RL AgentsNeurIPS 2022
- Optimistically Optimistic Exploration for Provably Efficient Infinite-Horizon Reinforcement and Imitation LearningCOLT 2025
- Optimistically Optimistic Exploration for Provably Efficient Infinite-Horizon Reinforcement and Imitation LearningCOLT 2025
- Mastering Atari Games with Limited DataNeurIPS 2021
- Sample-efficient and Scalable Exploration in Continuous-Time RLICLR 2026
- HyperDQN: A Randomized Exploration Method for Deep Reinforcement LearningICLR 2022
Related open problems
RL Theory & Markov Decision Processes
Sample-Efficient Reinforcement Learning Under Markovian Trajectory Sampling Without a Generative Model
Theoretical analysis in these MDP frameworks relies on a generative model that provides independent transition samples for arbitrary state-action queries. In real-world physical systems and streaming applications, arbitrary state resetting is impossible and data must be gathered along continuous online or episodic trajectories. As a result, existing sample complexity and convergence guarantees fail to hold under realistic Markovian data generation, leaving theoretical bounds disconnected from deployable online settings.
RL Theory & Markov Decision Processes
Reinforcement Learning Algorithms Without Generative Model and Reset Oracles
A wide class of theoretical and model-based RL methods rely on access to a generative model or simulator capable of querying arbitrary state-action pairs and performing instant resets. In physical systems, streaming data settings, and non-resettable environments, arbitrary state querying is impossible, restricting these algorithms to synthetic simulations. Consequently, the theoretical guarantees and policy optimization mechanisms derived under generative oracle assumptions fail completely when deployed in realistic single-trajectory or trajectory-constrained online environments.
RL Theory & Markov Decision Processes
Regret Bounds with Problem-Dependent Constants Removed: Minimax Characterization of the Gap Between Instance-Independent and Instance-Dependent Guarantees
Every one of these bounds is technically instance-independent, but each carries a "constant" — κ, exponential variation budgets, hitting-time factors, (1−γ)^−6, K^n, binomial coefficients — that is itself an instance-dependent quantity scaling exponentially or polynomially-huge on exactly the instances practitioners care about. The result is that published algorithms are provably efficient only on benign instances, and no one knows which of these constants are artifacts of analysis versus intrinsic hardness. Without a theory of which dependence is real, practitioners cannot tell whether a tight-constraint CMDP or a large-state bandit is fundamentally hard or merely badly analyzed.