DARLING tackles non-stationary RL with guarantees, improving dynamic regret.
problem Non-stationary reinforcement learning in unknown change points.
method Detection Augmented Reinforcement Learning (DARLING) for tabular and linear MDPs.
result DARLING matches minimax lower bounds in tabular and linear MDPs.
Efficiently identifies best policies in tabular MDPs with reduced computational cost.
problem Identifying the best policy in tabular MDPs with high computational cost.
method Combines posterior sampling with online learning to achieve asymptotic optimality.
result Achieves optimal sample complexity and posterior contraction rate with O ( S 2 A H ) O(S^2AH) O ( S 2 A H ) per episode. Paper establishes first instance-dependent lower bound for PAC reinforcement learning.
problem Identifying near-optimal policies in tabular MDPs with minimal samples.
method Proposes instance-dependent lower bound for sample complexity.
result Lower bound closely matches PEDEL algorithm's sample complexity.
New algorithm matches best regret bound for tabular Contextual Bandit problems.
problem Learning from observations in uncertain environments.
method A minor variant of a reinforcement learning algorithm for MDPs.
result The algorithm matches the best possible regret bound i l d e O ( S A T ) ilde O (\sqrt{SAT}) i l d e O ( S A T ) for tabular Contextual Bandit problems. RandQL is a new model-free algorithm for MDPs with a novel learning rate randomization approach.
problem Minimizing regret in episodic MDPs with model-free methods.
method Randomized Q-learning with learning rate randomization.
result RandQL achieves optimal regret bounds in both tabular and metric state-action spaces.
New method achieves nearly horizon-free offline reinforcement learning for tabular and linear MDPs.
problem Achieving efficient reinforcement learning without knowing the horizon.
method Simple recursion-based method to bound total variance in offline scenarios.
result Nearly horizon-free sample complexity bounds for offline policy evaluation and optimization.
New framework for reinforcement learning with adversarial preferences in tabular MDPs.
problem Learning from preferences rather than direct rewards in MDPs with adversarial settings.
method Developed PbMDPs framework, established lower bounds, and proposed algorithms for regret minimization.
result Achieved regret bounds of Ω ( ( H 2 S K ) 1 / 3 T 2 / 3 ) Ω( (H^2 S K)^{1/3} T^{2/3} ) Ω (( H 2 S K ) 1/3 T 2/3 ) for PbMDPs with Borda scores. This paper establishes that optimistic algorithms attain gap-dependent and non-asymptotic logarithmic regret for episodic MDPs. In contrast to prior work, our bounds do not suffer a dependence on diameter-like quantities or ergodicity, and smoothly interpolate between the gap dependent logarithmic-regret, and the $\wid…
New model selects robustly in adversarial reinforcement learning with unknown corruption.
problem Adversarial corruption in reinforcement learning with unknown total corruption amount.
method Model selection approach for finite-horizon tabular and linear MDPs.
result First worst-case optimal bound without knowledge of total corruption.
Hybrid RL algorithms improve offline and online RL in linear MDPs.
problem Improving RL performance without single-policy concentrability.
method Developed computationally efficient algorithms for PAC and regret-minimizing RL in linear MDPs.
result Achieved sharper error or regret bounds for linear MDPs.
New RL theory reduces sample complexity for mixing MDPs.
problem Optimal sample complexity for reinforcement learning in mixing MDPs.
method Regeneration-type ideas to analyze mixing times.
result Optimal sample complexity depends on mixing time, not just discount factor.
Optimal strategy for reinforcement learning with delayed observations.
problem Delayed state observation in reinforcement learning.
method Combines augmentation method and upper confidence bound approach.
result Minimax optimal regret bound of i l d e O ( H D max S A K ) ilde{\mathcal{O}}(H \sqrt{D_{\max} SAK}) i l d e O ( H D m a x S A K ) . New RL method learns K-step lookahead Q-functions for fixed-horizon MDPs.
problem Challenges in online reinforcement learning for non-episodic, finite-horizon MDPs.
method Introduces a K-step lookahead Q-function with a time-varying threshold for selecting actions.
result Achieves minimax optimal constant regret for K=1 and O ( max ( ( K − 1 ) , C K − 1 ) S A T log ( T ) ) \mathcal{O}(\max((K-1),C_{K-1})\sqrt{SAT\log(T)}) O ( max (( K − 1 ) , C K − 1 ) S A T log ( T ) ) regret for K ≥ 2. UCBVI-γ algorithm minimizes regret in discounted MDPs.
problem Minimizing regret in discounted MDPs.
method Optimism in the face of uncertainty principle and Bernstein-type bonus.
result UCBVI-γ achieves nearly minimax optimal regret.
New algorithms reduce regret in online MDPs by adapting to data and variance.
problem Adapting to both adversarial and stochastic environments in online MDPs.
method Develops algorithms based on global optimization and policy optimization, using optimistic follow-the-regularized-leader with log-barrier regularization.
result Achieves refined data-dependent and variance-dependent regret bounds.
New method for optimistic planning in MDPs using regularization.
problem Optimistic planning in infinite-horizon discounted MDPs.
method Regularized dynamic programming for approximate value iteration.
result Achieves near-optimal statistical guarantees in learning policies.
Active Reinforcement Learning (ARL) is a twist on RL where the agent observes reward information only if it pays a cost. This subtle change makes exploration substantially more challenging. Powerful principles in RL like optimism, Thompson sampling, and random exploration do not help with ARL. We relate ARL in tabular …
Paper presents a reduction-based framework for conservative bandits and RL with improved lower and upper bounds.
problem Conservative bandits and reinforcement learning problems.
method Reduction technique to calculate necessary and sufficient budget from baseline policy.
result Improved lower and upper bounds for various conservative settings.
Safe exploration in RF-RL doesn't increase sample complexity.
problem Achieving optimal policies with safety constraints in reward-free RL.
method Proposed SWEET framework for tabular and low-rank MDP settings, leveraging truncated value functions.
result Sample complexities match or outperform constraint-free counterparts, proving safety constraints have little impact.
New algorithms approximate state similarity in large MDPs.
problem Computing exact bisimulation metrics in large MDPs is expensive and impractical.
method Developed a new metric tied to behavior policy, and two algorithms for approximating it.
result Presented algorithms that can approximate bisimulation metrics in large, deterministic MDPs.
Efficient RL for linear MDPs with unknown transitions.
problem Long planning horizons and unknown state transitions in linear mixture MDPs.
method Horizon-free algorithm using weighted least squares with variance and uncertainty awareness.
result Achieves optimal regret up to logarithmic factors.
New algorithm tackles non-stationary reinforcement learning with general function approximation.
problem Understanding non-stationary MDPs with function approximation.
method Dynamic Bellman Eluder (DBE) dimension for complexity, sliding window mechanism, confidence set design.
result Upper bound on dynamic regret for proposed SW-OPEA algorithm.
New algorithm for offline RL with linear approx in MDPs and MGs, nearly optimal.
problem Offline RL with linear function approximation in MDPs and MGs.
method Pessimism-based algorithm with uncertainty decomposition via reference function.
result Nearly minimax optimal performance in offline RL for MDPs and MGs.
Policy gradient converges linearly with Hadamard parameterization in tabular settings.
problem Convergence of policy gradient methods under Hadamard parameterization.
method Studied convergence rate and established linear convergence after k 0 k_0 k 0 iterations. result Algorithm converges linearly with rate $O( rac{1}{k})$ and faster locally after k 0 k_0 k 0 . We consider the recently proposed reinforcement learning (RL) framework of Contextual Markov Decision Processes (CMDP), where the agent interacts with a (potentially adversarial) sequence of episodic tabular MDPs. In addition, a context vector determining the MDP parameters is available to the agent at the start of eac…
New RL algorithm maximizes CVaR in low-rank MDPs with provable efficiency.
problem Maximizing CVaR in large state spaces with function approximation.
method Upper Confidence Bound (UCB) bonus-driven algorithm for low-rank MDPs.
result Achieves sample complexity of O(H^7 A^2 d^4 / τ^2 ε^2) for ε-optimal CVaR.
Improved exploration in adversarial MDPs via dilated bonuses.
problem Global exploration in adversarial MDPs with local-search policies.
method Dilated bonuses to facilitate global exploration.
result Improved regret bounds in various MDP settings.
Entropy-regularized NPG methods converge linearly in discounted MDPs.
problem Theoretical limitations of NPG methods in reinforcement learning.
method Entropy regularization in conjunction with NPG methods for discounted MDPs.
result Entropy-regularized NPG methods converge linearly in discounted MDPs.
Paper analyzes convergence of dynamic policy gradient for MDPs, improving performance in finite-time problems.
problem Optimal policies in finite-time MDPs are not stationary and require epoch-specific training.
method Introduces dynamic policy gradient combining dynamic programming and policy gradient, analyzes convergence for softmax parametrisation.
result Dynamic policy gradient training exploits finite-time structure, leading to better convergence bounds.
The paper introduces a method to learn and apply value envelopes for faster online reinforcement learning.
problem Accelerating online reinforcement learning using offline data with theoretical grounding.
method A two-stage framework: offline data for learning value bounds, online algorithms for applying them.
result Substantial regret reductions in empirical tests on tabular MDPs.
Reward-free RL in linear MDPs is as hard as reward-aware RL.
problem Reward-free RL in linear MDPs without access to the reward function during exploration.
method Developed a computationally efficient algorithm with sample complexity O ~ ( d 2 H 5 / ε 2 ) \widetilde{\mathcal{O}}(d^2 H^5/ε^2) O ( d 2 H 5 / ε 2 ) . result Achieved optimal d d d dependence in linear MDPs for reward-free RL, matching the reward-aware RL setting. Study shows policy gradient convergence for entropy-regularized MDPs with neural nets in mean-field regime.
problem Global convergence of policy gradient for entropy-regularized MDPs with neural network approximation.
method Softmax policy with neural network approximation in mean-field regime, gradient flow in 2-Wasserstein metric, exponential convergence under sufficient regularization.
result Gradient flow converges exponentially fast to the unique stationary solution under sufficient regularization.
This paper improves reinforcement learning efficiency for large-scale MDPs.
problem High sample complexity in tabular RL settings with large state and action spaces.
method Model-based approach and Q-learning with linearly parameterized features.
result Provably efficient learning with sample complexity bounds.
A new RL framework allows removing user data without affecting performance.
problem Efficiently removing user data from a reinforcement learning model without affecting performance.
method Formulated a ρ ρ ρ -TV-stable RL algorithm for tabular MDPs that supports exact unlearning. result Achieved a nearly minimax optimal regret bound of Ω ( H S A T + S A H / ρ ) Ω(H\sqrt{\!SAT}\! +\! {SAH}/ρ) Ω ( H S A T + S A H / ρ ) for ρ ρ ρ -TV-stable RL algorithms. Optimistic algorithms and Thompson sampling use info-theory for better reinforcement learning.
problem Designing algorithms that balance exploration and exploitation in reinforcement learning.
method Integrating information-theoretic concepts into optimistic algorithms and Thompson sampling.
result Cumulative regret bound depends on uncertainty and quantifies prior information value.
This paper is concerned with offline reinforcement learning (RL), which learns using pre-collected data without further exploration. Effective offline RL would be able to accommodate distribution shift and limited data coverage. However, prior algorithms or analyses either suffer from suboptimal sample complexities or …
Paper analyzes online reinforcement learning with outcome-based feedback, providing efficient algorithms and fundamental limits.
problem Assigning credit to actions in reinforcement learning with only endpoint rewards.
method Develops a provably sample-efficient algorithm for online reinforcement learning with general function approximation.
result Achieves O ( C m c o v H 3 / ε 2 ) O(C_{
m cov} H^3/ε^2) O ( C m co v H 3 / ε 2 ) sample complexity, characterizing statistical separation between outcome-based and per-step rewards. Model-free Reinforcement Learning (RL) algorithms such as Q-learning [Watkins, Dayan 92] have been widely used in practice and can achieve human level performance in applications such as video games [Mnih et al. 15]. Recently, equipped with the idea of optimism in the face of uncertainty, Q-learning algorithms [Jin, Al…
Transformers can implement reinforcement learning algorithms from data without updates.
problem Training reinforcement learning algorithms from data without parameter updates.
method Design a teacher-mimicking training procedure for transformers to implement policy-improvement methods.
result Gradient flow converges to an optimal parameter manifold corresponding to the desired RL update.
New method reduces regret and communication costs in federated Q-learning.
problem Worst-case regret and communication cost bounds in federated Q-learning.
method Gap-dependent analysis leveraging MDP structures.
result Achieves log T \log T log T -type regret and communication cost bounds. New bounds on IDS for RL show how to balance computation and learning efficiency.
problem Understanding and optimizing information-directed sampling (IDS) for reinforcement learning.
method Developed novel information-theoretic tools to bound information ratio and cumulative information gain.
result Derived prior-free Bayesian regret bounds for IDS in tabular finite-horizon MDPs and improved computational efficiency.
New method quantifies uncertainty in reinforcement learning models.
problem Quantifying uncertainty over expected cumulative rewards in reinforcement learning.
method Proposes a new uncertainty Bellman equation to more accurately estimate value function variance.
result Our method converges to the true posterior variance over values and improves sample-efficiency.
Study non-asymptotic BPI guarantees for online RL.
problem Identify optimal policy in MDP with high confidence.
method Non-asymptotic sample complexity guarantees for NaS algorithm.
result Sample complexity depends on MDP connectivity and curvature.
Study online learning in MDPs with aggregate bandit feedback, achieving low regret in both stochastic and adversarial settings.
problem Online learning in finite-horizon episodic MDPs with aggregate bandit feedback.
method Best-of-both-worlds (BOBW) algorithms using FTRL over occupancy measures, self-bounding techniques, and new loss estimators.
result First BOBW algorithms for episodic tabular MDPs with aggregate bandit feedback achieving O ( log T ) O(\log T) O ( log T ) regret in stochastic and O ( T ) {O}(\sqrt{T}) O ( T ) regret in adversarial settings. Paper presents an efficient algorithm for linear MDP with low switching cost.
problem Large state space reinforcement learning problems with low switching cost.
method First algorithm for linear MDP with low switching cost, achieving near-optimal regret and switching cost.
result Regret bound of $\widetilde{O}\left(\sqrt{d^3H^4K}
ight)$ and near-optimal switching cost of $O\left(d H\log K
ight)$ .
Distributional reinforcement learning (distributional RL) has seen empirical success in complex Markov Decision Processes (MDPs) in the setting of nonlinear function approximation. However, there are many different ways in which one can leverage the distributional approach to reinforcement learning. In this paper, we p…
Improved sample complexity for target Q-learning in finite MDPs with generative oracle.
problem Sample complexity of target Q-learning in finite MDPs with a generative oracle.
method Analyzed target Q-learning algorithm in tabular case with a generative oracle, improved sample complexity.
result Improved sample complexity for target Q-learning in various scenarios.
New algorithm learns optimal policies with minimal memory and time.
problem Learning optimal policies in discounted MDPs with short burn-in time.
method Variance reduction and adaptive policy switching.
result First regret-optimal model-free algorithm with low burn-in time.