New RL algorithm learns good actions from offline data, reducing uncertainty and divergence.
problem Limited applicability of current RL algorithms in real-world settings due to high costs of exploration.
method Proposes an algorithm for batch RL using a fixed offline dataset, with penalties for policy and value constraints.
result Compared favorably to state-of-the-art methods on 32 continuous-action benchmarks.
ABS dynamically adjusts batch size based on policy stability, improving RL performance.
problem Diminishing returns with large batch sizes in RL due to non-stationary data.
method Adaptive Batch Scaling (ABS) with Behavioral Divergence metric.
result Larger batch sizes can improve RL performance, contrary to conventional wisdom.
New batch RL method avoids overly optimistic policies.
problem Challenges in applying RL to large domains with limited data.
method Modified Bellman optimality and evaluation back-up for conservative updates.
result Can find approximately best policies within explored state-action space.
A novel approach learns goal-conditioned policies for locomotion using batch RL.
problem Training goal-conditioned policies for rotation invariant locomotion.
method Data augmentation and Siamese framework for invariance.
result Our approach outperforms existing RL algorithms on 3D locomotion agents.
The paper analyzes risk bounds and Rademacher complexity in batch RL.
problem Estimating/minimizing Bellman error with general value function approximation.
method Characterizes generalization performance using Rademacher complexities of function classes.
result Risk bounds and Rademacher complexities provide insights into batch RL.
Critic-regularized regression improves offline RL performance.
problem Poor performance of off-policy algorithms in offline RL.
method Critic-regularized regression (CRR) for policy learning from fixed datasets.
result CRR outperforms state-of-the-art offline RL algorithms significantly.
STEEL tackles batch RL with singularity, improving policy optimization.
problem Existing RL methods assume absolutely continuous data, but STEEL handles non-overlapping regions.
method Proposes STEEL algorithm using maximum mean discrepancy and distributionally robust optimization.
result First finite-sample regret guarantee for batch RL with singularity.
New RL difficulty shown for discounted settings.
problem Difficulty in reinforcement learning with discounted rewards.
method Adapted Wang et al. (2020) construction to 2-state MDP.
result Learning impossible even with infinite data in discounted setting.
A new algorithm learns policies from batch data in hierarchical RL.
problem Learning policies from fixed batches of data without full exploration.
method Modeling RL as a two-player game with a leader-follower structure, proposing StackelbergLearner.
result StackelbergLearner achieves competitive performance in batch RL and real-world datasets.
A new batch ensemble method reduces regret in stochastic bandits.
problem Efficiently balancing exploration and exploitation in online RL.
method Proposes a batch ensemble scheme to achieve near-optimal regret.
result Proves near-optimal regret with a single parameter independent of variance.
In batch reinforcement learning (RL), one often constrains a learned policy to be close to the behavior (data-generating) policy, e.g., by constraining the learned action distribution to differ from the behavior policy by some maximum degree that is the same at each state. This can cause batch RL to be overly conservat…
This paper introduces sample-averaged Q-learning for better RL performance.
problem Improving reinforcement learning algorithms by managing uncertainty.
method Integrates statistical inference into Q-learning through sample averaging and functional central limit theorem.
result Establishes a unified theoretical foundation for sample-averaged Q-learning.
Most deep reinforcement learning (RL) systems are not able to learn effectively from off-policy data, especially if they cannot explore online in the environment. These are critical shortcomings for applying RL to real-world problems where collecting data is expensive, and models must be tested offline before being dep…
Sparse feature selection improves batch RL efficiency.
problem High-dimensional batch RL with many features.
method Sparse linear function approximation, Lasso, group Lasso, fitted Q-evaluation, fitted Q-iteration.
result Sparse feature selection makes batch RL more sample efficient.
Efficient RL algorithms for linear function approximation with limited adaptivity constraints.
problem Limited adaptivity in reinforcement learning with linear function approximation.
method Proposed two efficient online RL algorithms for episodic linear Markov decision processes under batch learning and rare policy switch models.
result Achieved efficient regret bounds for both batch learning and rare policy switch models, with substantial reduction in adaptivity.
Off-policy reinforcement learning algorithms promise to be applicable in settings where only a fixed data-set (batch) of environment interactions is available and no new experience can be acquired. This property makes these algorithms appealing for real world problems such as robot control. In practice, however, standa…
Breaks the hardness conjecture for batch RL with a novel tournament-based approach.
problem Sample-efficient reinforcement learning from exploratory data.
method BVFT algorithm using pairwise comparison and state-action partition.
result Solves the learning problem in a setting previously thought impossible.
Maximizes Rényi entropy for efficient exploration in reward-free RL.
problem Challenges of exploration in reward-free reinforcement learning.
method Maximizes Rényi entropy over state-action space in exploration phase; uses batch RL for planning phase.
result Effective and sample-efficient exploration leading to superior policies.
This paper considers Safe Policy Improvement (SPI) in Batch Reinforcement Learning (Batch RL): from a fixed dataset and without direct access to the true environment, train a policy that is guaranteed to perform at least as well as the baseline policy used to collect the data. Our approach, called SPI with Baseline Boo…
This work uses RL to optimize batch experiments in SDOE.
problem Maximizing knowledge with limited resources and constraints.
method Reinforcement Learning for batch-sampling in Bayesian SDOE.
result The proposed algorithm optimizes experiment selection for multiple tasks.
Extends Thompson sampling for RL with fewer episodes.
problem Limited episodes in RL settings.
method Batch Bayesian optimization over episodes to learn action bias terms.
result Significantly outperforms standard Thompson sampling.
MOPO optimizes offline RL by penalizing dynamics uncertainty.
problem Learning policies from offline data with distributional shift.
method Modify model-based RL to avoid distributional shift.
result MOPO outperforms model-free and standard model-based RL.
Evolution Strategies (ES) emerged as a scalable alternative to popular Reinforcement Learning (RL) techniques, providing an almost perfect speedup when distributed across hundreds of CPU cores thanks to a reduced communication overhead. Despite providing large improvements in wall-clock time, ES is data inefficient whe…
LF-IBIS learns optimal policies online without explicit likelihood.
problem Bayesian RL challenges due to intractable likelihood functions.
method Combines ABC with IBIS for online belief updates.
result Approximates posterior distributions for policies and parameters.
A simple approach to offline RL without additional complexity.
problem Learning from a fixed dataset of actions with value estimation errors.
method Adding a behavior cloning term to the policy update of an online RL algorithm and normalizing the data.
result Matches the performance of state-of-the-art offline RL algorithms with minimal changes.
A new offline RL framework unifies imitation learning and vanilla offline RL.
problem Learning from expert datasets without active data collection.
method A new offline RL framework that interpolates between imitation learning and vanilla offline RL, using a weak concentrability coefficient and a lower confidence bound algorithm.
result LCB algorithm achieves a faster rate of 1/N for nearly-expert datasets, and is adaptively optimal for the entire data composition range. New batched Langevin Thompson Sampling reduces communication costs for sequential decision making.
problem Efficiently learning unknown reward distributions and transition dynamics in batched settings.
method Langevin Thompson Sampling with logarithmic communication costs.
result Order-optimal regret guarantees for stochastic MABs and RL.
Two new algorithms improve Q* approximation in batch RL with linear error propagation.
problem Improving Q* approximation in batch reinforcement learning.
method Two novel algorithms that estimate Bellman error directly, without quadratic dependence.
result Linear-in-horizon error propagation for batch RL algorithms.
Batch Reinforcement Learning (RL) algorithms attempt to choose a policy from a designer-provided class of policies given a fixed set of training data. Choosing the policy which maximizes an estimate of return often leads to over-fitting when only limited data is available, due to the size of the policy class in relatio…
Paper establishes lower bounds and optimal algorithms for deployment-efficient RL.
problem Deployment efficiency in reinforcement learning.
method Optimization with constraints, lower bounds, algorithms.
result Established optimal algorithms for deployment-efficient RL.
New method ensures consistent inference across different tensor parallel sizes for large language models.
problem Non-deterministic inference in large language models due to inconsistent reduction orders across GPUs.
method Tree-Based Invariant Kernels (TBIK) that align intra- and inter-GPU reduction orders through a unified hierarchical binary tree structure.
result Bit-wise identical results across different tensor parallel sizes for RL training.
We consider the transfer of experience samples (i.e., tuples < s, a, s', r >) in reinforcement learning (RL), collected from a set of source tasks to improve the learning process in a given target task. Most of the related approaches focus on selecting the most relevant source samples for solving the target task, but t…
The distributional perspective on reinforcement learning (RL) has given rise to a series of successful Q-learning algorithms, resulting in state-of-the-art performance in arcade game environments. However, it has not yet been analyzed how these findings from a discrete setting translate to complex practical application…
In an increasing number of domains it has been demonstrated that deep learning models can be trained using relatively large batch sizes without sacrificing data efficiency. However the limits of this massive data parallelism seem to differ from domain to domain, ranging from batches of tens of thousands in ImageNet to …
New benchmarks for offline RL from diverse datasets.
problem Measuring progress in offline RL due to lack of suitable benchmarks.
method Developed benchmarks tailored for offline RL, focusing on diverse dataset properties.
result Revealed deficiencies in existing offline RL algorithms.
Value-function approximation methods that operate in batch mode have foundational importance to reinforcement learning (RL). Finite sample guarantees for these methods often crucially rely on two types of assumptions: (1) mild distribution shift, and (2) representation conditions that are stronger than realizability. H…
State-constrained offline RL expands RL's learning scope.
problem Batch-constrained offline RL limits policies to seen actions.
method Introduces state-constrained offline RL focusing on state distribution.
result Policy can take high-quality out-of-distribution actions.
Learning good representations is a long standing problem in reinforcement learning (RL). One of the conventional ways to achieve this goal in the supervised setting is through regularization of the parameters. Extending some of these ideas to the RL setting has not yielded similar improvements in learning. In this pape…
New algorithm identifies near-optimal policies in adversarial distributed RL settings.
problem Adversarial agents in distributed RL settings that can collude and report arbitrary data.
method Weighted-Clique algorithm for robust mean estimation from batches, combined with novel distributed algorithms.
result Achieves superior robustness guarantees and near-optimal sample complexities in both offline and online settings.
The ability for policies to generalize to new environments is key to the broad application of RL agents. A promising approach to prevent an agent's policy from overfitting to a limited set of training environments is to apply regularization techniques originally developed for supervised learning. However, there are sta…
Reinforcement learning (RL) is about sequential decision making and is traditionally opposed to supervised learning (SL) and unsupervised learning (USL). In RL, given the current state, the agent makes a decision that may influence the next state as opposed to SL (and USL) where, the next state remains the same, regard…
Study uses online bootstrap for RL inference, showing effectiveness.
problem Statistical inference for RL parameters in online settings.
method Online bootstrap method applied to TD and GTD algorithms in RL.
result Method is distributionally consistent for policy evaluation inference.
Paper tackles offline meta-reinforcement learning with a new algorithm.
problem Performing reinforcement learning on limited data from a new task.
method Meta-Actor Critic with Advantage Weighting (MACAW) algorithm.
result Achieves notable gains over prior methods on offline meta-RL benchmarks.
DRIFT uses RL to automate functional software testing efficiently.
problem Efficient and reliable automated software testing.
method DRIFT employs Q-learning with Graph Neural Networks on symbolic UI representations.
result DRIFT can robustly test software functionalities in a fully automated manner.
Many modern commercial sites employ recommender systems to propose relevant content to users. While most systems are focused on maximizing the immediate gain (clicks, purchases or ratings), a better notion of success would be the lifetime value (LTV) of the user-system interaction. The LTV approach considers the future…
In this paper, we investigate the problem of overfitting in deep reinforcement learning. Among the most common benchmarks in RL, it is customary to use the same environments for both training and testing. This practice offers relatively little insight into an agent's ability to generalize. We address this issue by usin…
New algorithm for reward-free RL with linear function approximation, reducing sample complexity.
problem Efficiently learning optimal policies without prior reward information in complex environments.
method Developed an algorithm for reward-free RL in linear Markov decision processes, proving sample complexity bounds.
result Polynomial sample complexity in feature dimension and planning horizon, independent of states and actions.
The paper improves model-based reinforcement learning by using multi-timestep objectives.
problem Compounding errors in one-step dynamics models as trajectory length increases.
method Developed a multi-timestep objective as a weighted sum of losses at various future horizons.
result Exponentially decaying weights significantly improve long-horizon performance.