New algorithm FLUTE achieves uniform-PAC convergence in RL with linear approx.
problem RL with linear function approximation lacks uniform-PAC guarantees.
method FLUTE algorithm with minimax value function estimator and multi-level partition scheme.
result Uniform-PAC convergence to optimal policy with high probability.
Statistical performance bounds for reinforcement learning (RL) algorithms can be critical for high-stakes applications like healthcare. This paper introduces a new framework for theoretically measuring the performance of such algorithms called Uniform-PAC, which is a strengthening of the classical Probably Approximatel…
New hybrid RL algorithm outperforms model-free and model-based methods.
problem Improving reinforcement learning algorithms for MDPs.
method Combines model-free and model-based learning, with a PAC analysis.
result Outperforms both model-free and model-based methods in most cases.
New reinforcement learning bound improves generalization for sequential data.
problem Challenges in obtaining generalization guarantees for reinforcement learning due to sequential data.
method PAC-Bayesian reinforcement learning with consideration of Markov dependencies and mixing time.
result Demonstrated practical utility through PB-SAC, providing meaningful confidence certificates.
Paper bounds PAC RL sample complexity in deterministic MDPs.
problem Identify ε-optimal policy with high probability.
method Proposes nearly matching upper and lower bounds on sample complexity, introduces deterministic return gap, uses graph-theoretical concepts and maximum-coverage exploration.
result First nearly matching upper and lower bounds on sample complexity for PAC RL in deterministic MDPs.
Bayesian priors offer a compact yet general means of incorporating domain knowledge into many learning tasks. The correctness of the Bayesian analysis and inference, however, largely depends on accuracy and correctness of these priors. PAC-Bayesian methods overcome this problem by providing bounds that hold regardless …
Private RL algorithm with privacy guarantees for personalized medicine decisions.
problem Privacy-preserving reinforcement learning for personalized medicine decisions.
method Developed a private optimism-based RL algorithm using joint differential privacy (JDP).
result Achieved strong PAC and regret bounds with a privacy guarantee.
We study the computational tractability of PAC reinforcement learning with rich observations. We present new provably sample-efficient algorithms for environments with deterministic hidden state dynamics and stochastic rich observations. These methods operate in an oracle model of computation -- accessing policy and va…
New reinforcement learning algorithm achieves instance-optimal sample complexity.
problem Achieving low regret and identifying optimal policies in reinforcement learning.
method A novel planning-based algorithm that explicitly accounts for state visitation distributions.
result The proposed algorithm attains nearly minimax optimal sample complexity, improving over worst-case bounds.
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 PAC-Bayesian approach stabilizes actor-critic learning.
problem Training instability in actor-critic algorithms.
method Employing PAC-Bayesian bound as the critic training objective.
result Significant improvement in online learning performance.
New algorithms achieve uniform-PAC guarantees for RL with bounded eluder dimension.
problem Achieving strong performance guarantees in reinforcement learning.
method Proposes algorithms for nonlinear bandits and model-based episodic RL with a bounded eluder dimension.
result Achieves uniform-PAC sample complexity that matches state-of-the-art regret bounds or sample complexity guarantees.
We propose an algorithm combining calibrated prediction and generalization bounds from learning theory to construct confidence sets for deep neural networks with PAC guarantees---i.e., the confidence set for a given input contains the true label with high probability. We demonstrate how our approach can be used to cons…
Survey of reinforcement learning guarantees with data constraints.
problem Guaranteeing near-optimal policies with limited data in reinforcement learning.
method Coverage-Structure-Objective (CSO) framework to decompose sample complexity results.
result Progress on PAC guarantees for reinforcement learning, covering various models and settings.
ASE safely explores unknown MDPs with unknown dynamics, improving sample efficiency.
problem Balancing exploration and safety in unknown MDPs with stochastic dynamics.
method Exploits analogies between state-action pairs to safely learn near-optimal policies.
result Empirically improves sample efficiency compared to existing methods.
Study risk-sensitive reinforcement learning with optimized certainty equivalents.
problem Risk-sensitive reinforcement learning in finite discounted MDPs.
method Analyzed a simple model-based approach and derived PAC sample complexity bounds.
result Established tight sample complexity bounds for value and policy learning.
New method extends low-rank MDPs to continuous action spaces.
problem Limited applicability of current low-rank MDP methods to continuous action spaces.
method Extending FLAMBE algorithm to continuous action spaces with Hölder smoothness conditions.
result Similar PAC bound achieved for continuous actions with polynomial dependence on smoothness order.
New algorithm reduces bias in off-policy reinforcement learning.
problem Challenges in designing off-policy reinforcement learning algorithms.
method Doubly robust off-policy actor-critic (DR-Off-PAC) with a single timescale structure.
result Establishes the first overall sample complexity analysis for a single time-scale off-policy AC algorithm.
Paper establishes new lower bounds for MDPs with changing transition kernels.
problem Minimizing sample complexity and regret in non-stationary MDPs.
method Developed novel lower bounds and constructed hard MDPs.
result Proved Ω((H3SA/ε2)log(1/δ)) sample complexity lower bound. We present two alternative ways to apply PAC-Bayesian analysis to sequences of dependent random variables. The first is based on a new lemma that enables to bound expectations of convex functions of certain dependent random variables by expectations of the same functions of independent Bernoulli random variables. This …
New complexity measure helps in agnostic reinforcement learning with or without access to MDP dynamics.
problem Understanding the number of rounds needed to learn an ε-suboptimal policy in unknown MDPs.
method Introducing spanning capacity as a new complexity measure and developing POPLER algorithm.
result There is a separation between generative and online access models for agnostic learnability.
We develop model free PAC performance guarantees for multiple concurrent MDPs, extending recent works where a single learner interacts with multiple non-interacting agents in a noise free environment. Our framework allows noisy and resource limited communication between agents, and develops novel PAC guarantees in this…
Study risk-sensitive reinforcement learning with entropic risk measures and generative models.
problem Risk-sensitive reinforcement learning in discounted MDPs with recursive entropic risk measures.
method Introduced Model-Based ERM Q-Value Iteration (MB-RS-QVI) and derived PAC bounds on sample complexity for value and policy learning. result PAC bounds show exponential dependence on ∣β∣/(1−γ), with tight bounds in S and A. This paper explains why distributional reinforcement learning is better than vanilla RL using small-loss bounds.
problem Understanding when and why distributional reinforcement learning (DistRL) is superior to vanilla reinforcement learning (RL).
method The paper uses small-loss bounds to explain the benefits of DistRL, proposing algorithms and proving bounds for different RL settings.
result Distributional reinforcement learning (DistRL) outperforms vanilla RL when optimal costs are small, as shown by small-loss bounds.
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.
We study an exploration method for model-free RL that generalizes the counter-based exploration bonus methods and takes into account long term exploratory value of actions rather than a single step look-ahead. We propose a model-free RL method that modifies Delayed Q-learning and utilizes the long-term exploration bonu…
New bounds use IPMs to improve generalization in machine learning.
problem Improving generalization bounds in machine learning.
method PAC-Bayes bounds with Integral Probability Metrics (IPM).
result Natural interpolation between worst-case and favorable cases.
Recently, there has been significant progress in understanding reinforcement learning in discounted infinite-horizon Markov decision processes (MDPs) by deriving tight sample complexity bounds. However, in many real-world applications, an interactive learning agent operates for a fixed or bounded period of time, for ex…
New algorithm learns optimal policies with just 1 episode, settling horizon-dependence in RL.
problem Understanding the sample complexity of reinforcement learning with horizon length.
method Developed an algorithm using only O(1) episodes to achieve PAC guarantee, leveraging connections between value functions in discounted and finite-horizon MDPs and novel perturbation analysis. result Achieved the same PAC guarantee with only O(1) episodes of environment interactions, completely settling horizon-dependence in RL. CPPO learns policies from partial offline data in MDPs with structural assumptions.
problem Offline Reinforcement Learning with partial coverage assumption.
method Constrained Pessimistic Policy Optimization (CPPO) using a function class and model class constraint.
result CPPO achieves PAC guarantee with partial coverage, learning competitive policies.
We propose a new objective, the counterfactual objective, unifying existing objectives for off-policy policy gradient algorithms in the continuing reinforcement learning (RL) setting. Compared to the commonly used excursion objective, which can be misleading about the performance of the target policy when deployed, our…
We present a set of high-probability inequalities that control the concentration of weighted averages of multiple (possibly uncountably many) simultaneously evolving and interdependent martingales. Our results extend the PAC-Bayesian analysis in learning theory from the i.i.d. setting to martingales opening the way for…
This paper introduces a new metric, ULI, for RL that ensures both cumulative and instantaneous performance.
problem High-stakes applications require RL algorithms to avoid playing bad policies.
method Introduces uniform last-iterate (ULI) guarantee, a stronger metric capturing both cumulative and instantaneous performance.
result ULI directly implies near-optimal cumulative performance across various metrics, but not the other way around.
This paper studies systematic exploration for reinforcement learning with rich observations and function approximation. We introduce a new model called contextual decision processes, that unifies and generalizes most prior settings. Our first contribution is a complexity measure, the Bellman rank, that we show enables …
We consider the problem of knowledge transfer when an agent is facing a series of Reinforcement Learning (RL) tasks. We introduce a novel metric between Markov Decision Processes (MDPs) and establish that close MDPs have close optimal value functions. Formally, the optimal value functions are Lipschitz continuous with …
Regularization improves generalization in Bayesian RL, shown through algorithmic stability.
problem Ensuring good generalization in Bayesian reinforcement learning.
method Algorithmic stability, using regularization and fast convergence rates for mirror descent.
result Regularization makes the optimal policy stable, improving generalization.
New algorithms estimate Q-functions under partial coverage and realizability, improving offline RL guarantees.
problem Offline RL with limited exploration and assumptions about data coverage and Q-function realizability.
method Proposes minimax learning algorithms to estimate soft or vanilla Q-functions with L2-convergence guarantees. result PAC guarantees for offline RL under partial coverage and realizability conditions.
This work studies reinforcement learning in the Sim-to-Real setting, in which an agent is first trained on a number of simulators before being deployed in the real world, with the aim of decreasing the real-world sample complexity requirement. Using a dynamic model known as a rich observation Markov decision process (R…
Study on teaching reinforcement learning with Q-learning, reducing sample complexity.
problem Reducing sample complexity in reinforcement learning.
method Characterized teaching dimension for Q-learning under different teacher control, presented optimal teaching algorithms.
result Minimum number of samples needed for reinforcement learning is characterized.
This paper establishes strong lower bounds for learning in revealing POMDPs.
problem Understanding the fundamental limits of reinforcement learning in revealing partially observable Markov Decision Processes (POMDPs).
method Develops strong PAC and regret lower bounds for learning in revealing POMDPs using multi-step revealing POMDPs as a case study.
result Strong polynomial lower bounds for learning in revealing POMDPs, achieving significantly smaller gaps against current upper bounds.
We propose and study a new model for reinforcement learning with rich observations, generalizing contextual bandits to sequential decision making. These models require an agent to take actions based on observations (features) with the goal of achieving long-term performance competitive with a large set of policies. To …
Study examines sample complexity for RL with safety constraints.
problem Ensuring safety in RL for systems with constraints.
method Analyzes two RL approaches: generative and online.
result Sample complexity increased by a logarithmic factor in constraints.
Paper introduces a meta-critic for accelerating off-policy actor-critic learning.
problem Improving sample efficiency in continuous control tasks.
method Meta-critic that meta-learns an additional loss for the actor.
result Online meta-critic learning leads to improved performance in various continuous control environments.
This paper shows CEM is a special case of TTM, leading to new proofs and improved sample complexity bounds.
problem Improving sample complexity for reinforcement learning algorithms.
method Viewing CEM as an application of TTM, deriving new proofs and bounds.
result Improved sample complexity bounds for CEM under various conditions.
This work improves transferability of rewards inferred from expert demonstrations.
problem Transferability of rewards inferred from expert demonstrations under limited access to the expert's policy.
method Proposed principal angles as a measure of similarity and dissimilarity between transition laws. Established sufficient conditions for transferability under limited access.
result Two key results on sufficient conditions for transferability to any and local changes in transition laws.
We consider a reinforcement learning (RL) setting in which the agent interacts with a sequence of episodic MDPs. At the start of each episode the agent has access to some side-information or context that determines the dynamics of the MDP for that episode. Our setting is motivated by applications in healthcare where ba…
PAC-Bayesian bounds for MLPs with cross entropy loss validated.
problem Generalization bounds for MLPs with cross entropy loss.
method Introduced probabilistic explanations and proved PAC-Bayesian bounds using ELBO.
result MLPs with cross entropy loss inherently guarantee PAC-Bayesian generalization bounds.
Unified derivation of PAC-Bayes and MI bounds for general VC classes with fast rates.
problem Generalization bounds for machine learning models with VC classes.
method Unified derivation of conditional PAC-Bayesian and mutual information bounds, including MAC-Bayesian bounds.
result Nontrivial bounds for general VC classes and faster rates for specific conditions.