Stabilizes policy optimization with off-policy data using divergence augmentation.
problem Premature convergence and instability in policy optimization with off-policy data.
method Incorporates Bregman divergence between behavior and current policies to ensure safe policy updates.
result Empirically shows better performance in data-scarce scenarios compared to other algorithms.
To ensure stability of learning, state-of-the-art generalized policy iteration algorithms augment the policy improvement step with a trust region constraint bounding the information loss. The size of the trust region is commonly determined by the Kullback-Leibler (KL) divergence, which not only captures the notion of d…
This work improves policy-based training by proposing an evaluation balance objective for GFlowNets.
problem Reliable estimation of policy divergence under directed acyclic graphs remains challenging.
method Proposes an evaluation balance objective over partial episodes to measure policy divergence and improve policy-based training reliability.
result Evaluation balance strengthens policy-based training reliability and broadens its flexibility.
New algorithm stabilizes RL policy learning through divergence regularization.
problem Stabilize policy learning and improve performance in RL.
method Proximity term constraining discounted state-action visitation distributions to be close to each other.
result Proposed algorithm improves stability and final performance in RL tasks.
New model learns better policies from expert demonstrations with higher efficiency.
problem Learning accurate policies from expert demonstrations with high efficiency.
method Generative adversarial imitation learning (GAIL) model that learns f-divergence automatically. result Learns better policies with higher data efficiency in physics-based control tasks.
Paper analyzes sample complexity for offline f-divergence-regularized contextual bandits.
problem Lack of tight analyses for sample complexity in offline reinforcement learning.
method Novel pessimism-based analysis for reverse KL divergence, establishing ildeO(ε−1) sample complexity. result Achieves ildeO(ε−1) sample complexity for reverse KL divergence, surpassing existing bounds. The paper analyzes the reward improvement of aligned policies in large language models.
problem Optimizing policies in large language models while staying close to a reference policy.
method Information-theoretic analysis and reduction to exponential order statistics.
result Information-theoretic upper bounds on reward improvement are derived.
An optimal feedback controller for a given Markov decision process (MDP) can in principle be synthesized by value or policy iteration. However, if the system dynamics and the reward function are unknown, a learning agent must discover an optimal controller via direct interaction with the environment. Such interactive d…
The paper introduces a new method for estimating optimal policies in dynamic treatment regimes using information geometry.
problem Estimating optimal policies in dynamic treatment regimes.
method Minimum information divergence method based on γ-power divergence. result The γ-power divergence method effectively seeks the optimal policy by vanishing the divergence between policy-equivalent Q-functions. KL-constrained API shows optimization issues and improved with regularization.
problem Optimization issues in KL-constrained API algorithms.
method Comparison of KL divergence as a constraint vs. regularizer, empirical evaluation.
result KL-constrained API is not guaranteed to converge and incurs linear regret.
Reduces policy space complexity for reinforcement learning.
problem Efficiency in exploring vast policy spaces in reinforcement learning.
method Uses Rényi divergence and l1 norm to determine sample size for accurate policy approximation. result Established error bounds for sample size requirements in model-based and model-free settings.
Extends reinforcement learning alignment to scalar rewards, improving math reasoning.
problem Designing reinforcement learning algorithms for general LLM alignment.
method Introduces f-GRPO and f-HAL, estimating f-divergences between reward-aligned and unaligned distributions.
result Improves math reasoning RLVR tasks and mitigates reward hacking.
A key problem in reinforcement learning for control with general function approximators (such as deep neural networks and other nonlinear functions) is that, for many algorithms employed in practice, updates to the policy or Q-function may fail to improve performance---or worse, actually cause the policy performance …
A new method learns robust policies from offline data with latent structures.
problem Conservative policies under unrealistic dynamics shifts.
method d-RRMDP framework with f-divergence regularization and R2PVI algorithm. result R2PVI learns robust policies with superior computational efficiency.
New estimator reduces risk in slate bandits by leveraging Bayes risk criterion.
problem Evaluating slate policies using logged data when policies factorize over slots.
method Developed a new estimator using a control variate approach, showing risk improvement over existing methods.
result The new estimator has lower risk than the pseudoinverse estimator in slate bandit problems.
Paper introduces RCaI, a risk-sensitive control method using Rényi divergence.
problem Risk-sensitive control in reinforcement learning.
method RCaI extends CaI using Rényi divergence variational inference.
result Risk-sensitive optimal policy can be obtained by solving a soft Bellman equation.
TRM improves long-horizon LLM RL by masking divergent sequences.
problem Long-horizon reinforcement learning with LLMs suffers from off-policy mismatch and approximation errors.
method Derives and applies trust region bounds to control divergence, proposing Trust Region Masking.
result First non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.
TRM improves long-horizon reinforcement learning for LLMs by masking divergent sequences.
problem Long-horizon reinforcement learning for LLMs suffers from off-policy mismatch and approximation errors.
method Derives and applies trust region bounds to control divergence, proposing Trust Region Masking.
result First non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.
GRPO optimizes LLMs with verifiable rewards, amplifying policy success.
problem Improving LLMs' reasoning under verifiable binary rewards.
method Introduces GRPO, analyzes variants of reward normalization and regularization.
result GRPO amplifies policy success, converging to a fixed point exceeding the reference.
New RL algorithm learns robust policies without knowing nominal model.
problem Designing robust policies for RL with unknown nominal model.
method Model-based RL algorithm with three uncertainty set forms.
result Precise sample complexity for each uncertainty set.
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.
We propose to improve trust region policy search with normalizing flows policy. We illustrate that when the trust region is constructed by KL divergence constraints, normalizing flows policy generates samples far from the 'center' of the previous policy iterate, which potentially enables better exploration and helps av…
This paper introduces f-DPO, a generalized approach to Direct Preference Optimization using diverse divergence constraints.
problem Aligning large language models with human preferences while mitigating safety risks.
method Incorporates diverse divergence constraints to simplify the relationship between reward and optimal policy, eliminating the need for estimating the normalizing constant.
result Optimizes LLMs to align with human preferences more efficiently and under a broader set of divergence constraints.
Improved inference-time alignment using Best-of-N and smoothing.
problem Reward overoptimization in Best-of-N (BoN) due to poor proxy reward models.
method Introduced Soft Best-of-N (SBoN) and analyzed its performance through KL divergence and regret analysis.
result Smoothing helps SBoN mitigate reward overoptimization, especially when proxy reward quality is low.
Proposes a method to improve treatment policies in data-scarce clinical settings.
problem Improving treatment policies in data-scarce clinical settings with unobserved confounding.
method Uses a causal mechanism to model the underlying generative process and augments counterfactual trajectories with source domain priors.
result Significantly improves treatment policy performance in a simulated sepsis treatment task.
This work proposes robust reinforcement learning methods using both offline and online data.
problem Designing robust policies against parameter uncertainties in high-dimensional systems.
method Proposes RPQ for model-free learning with historical data and HyTQ for hybrid learning with both historical and online data.
result Unified analysis and theoretical guarantees for robust optimal policies in high-dimensional systems.
New method optimizes policies without assuming known link functions between preferences and rewards.
problem Policy alignment with unknown and unrestricted link functions.
method Formulates an f-divergence-constrained reward maximization problem, learning policies directly. result Induces a semiparametric single-index binary choice model for policy alignment.
The success of popular algorithms for deep reinforcement learning, such as policy-gradients and Q-learning, relies heavily on the availability of an informative reward signal at each timestep of the sequential decision-making process. When rewards are only sparsely available during an episode, or a rewarding feedback i…
A new framework for offline RL improves policy flexibility and regularity.
problem Lack of environmental interactions in offline RL leads to poor policy performance.
method Proposes a behavior-regularized implicit policy framework with modified policy-matching methods.
result The framework improves policy effectiveness and robustness beyond static datasets.
This paper optimizes off-policy evaluation in reinforcement learning with function approximation.
problem Estimating cumulative value of a new policy from logged data generated by an unknown policy.
method Regression-based fitted Q iteration method, equivalent to estimating conditional mean embedding of transition operator.
result The method is minimax-optimal, with nearly minimal estimation error.
Composing previously mastered skills to solve novel tasks promises dramatic improvements in the data efficiency of reinforcement learning. Here, we analyze two recent works composing behaviors represented in the form of action-value functions and show that they perform poorly in some situations. As part of this analysi…
New RL method nearly optimally learns policies with generative models.
problem Finding optimal policies in reinforcement learning with generative models.
method Mirror descent value iteration with KL divergence and entropy regularization.
result The method is nearly minimax-optimal for small ε-optimal policies. New methods minimize GFlowNet training divergences for better sampling.
problem Training GFlowNets with KL divergence leads to biased and high-variance estimators.
method Design and implement efficient estimators for four divergence measures.
result Properly minimizing these divergences yields a provably correct and effective training scheme.
A new framework for robust policy learning in MDPs with linear mixture dynamics.
problem Off-dynamics challenge in real-world decision-making problems.
method Linear mixture DRMDP framework, meta algorithm for robust policy learning.
result The new framework provides a more refined representation of uncertainties.
Inspired by the seminal work on Stein Variational Inference and Stein Variational Policy Gradient, we derived a method to generate samples from the posterior variational parameter distribution by \textit{explicitly} minimizing the KL divergence to match the target distribution in an amortize fashion. Consequently, we a…
Linear Q-learning converges to a bounded set without divergence.
problem Proving linear Q-learning does not diverge and converges to a bounded set.
method No modifications to the original linear Q-learning algorithm, no Bellman completeness or near-optimality assumptions, only an ε-softmax behavior policy with adaptive temperature.
result First L2 convergence rate of linear Q-learning iterates to a bounded set. The paper introduces a novel method for stable off-policy learning using value function chaining.
problem Stability issues in off-policy reinforcement learning.
method The approach involves learning on-policy first, then chaining off-policy value estimates.
result The method guarantees convergence and can approximate off-policy TD solutions.
In this paper we consider the basic version of Reinforcement Learning (RL) that involves computing optimal data driven (adaptive) policies for Markovian decision process with unknown transition probabilities. We provide a brief survey of the state of the art of the area and we compare the performance of the classic UCB…
Deep neural networks can estimate Q-values efficiently on low-dimensional state-action spaces.
problem Estimating the performance of a reinforcement learning policy using data from a different policy.
method Deep fitted Q-evaluation method leveraging manifold structure and convolutional neural networks.
result Sharp error bound for fitted Q-evaluation depends on intrinsic dimension and function space mismatch.
Entropy regularization improves policy optimization in reinforcement learning.
problem Improving policy optimization in reinforcement learning.
method Entropy regularization is introduced to soften the greedy policy towards a more diverse softmax policy, leading to a continuously parameterized algorithm that interpolates between policy gradient and Q-learning.
result An intermediate algorithm can improve performance in reinforcement learning.
Sharp policy value estimation for contextual bandits with unobserved confounders.
problem Estimating policy value under unobserved confounders with sensitivity analysis.
method Kernel method to approximate conditional moment constraints, leveraging f-divergence.
result Sharp lower bound of policy value, avoiding coarse relaxation of uncertainty set.
A new method for learning policies from demonstrations without reinforcement.
problem Learning policies from demonstrations without access to reinforcement signals.
method Energy-based distribution matching (EDM) to learn policy parameters and state marginals.
result EDM yields consistent performance gains over existing algorithms for strictly batch imitation learning.
Proposes DR algorithms for distributionally robust off-policy evaluation and learning.
problem Sensitive to environment distribution shifts in offline observational data.
method Doubly robust and distributionally robust approaches for OPE/L.
result Achieves semiparametric efficiency and fast regret rate.
Sparse RSP routing improves graph exploration and classification.
problem Optimal randomized routing and distance measures on weighted graphs.
method Tsallis divergence regularization for sparse RSP.
result Sparse random walk converges to least-cost graph as temperature decreases.
The paper develops methods to estimate value functions in reinforcement learning for infinite horizon problems.
problem Constructing confidence intervals for value functions in infinite horizon reinforcement learning.
method Modeling the Q-function using series/sieve method and recursively updating the policy using SAVE method.
result The proposed CI achieves nominal coverage even when the optimal policy is not unique.
We observe that several existing policy gradient methods (such as vanilla policy gradient, PPO, A2C) may suffer from overly large gradients when the current policy is close to deterministic (even in some very simple environments), leading to an unstable training process. To address this issue, we propose a new method, …
Trust-region methods have yielded state-of-the-art results in policy search. A common approach is to use KL-divergence to bound the region of trust resulting in a natural gradient policy update. We show that the natural gradient and trust region optimization are equivalent if we use the natural parameterization of a st…
A new method stabilizes deep reinforcement learning by using QGraphs to retain replay memory information.
problem Stabilizing model-free off-policy deep reinforcement learning with soft divergence.
method Representing past experiences as a QGraph, selecting a subgraph with favorable structure, and using lower bounds for temporal difference learning.
result QG-DDPG method is less prone to soft divergence and more robust to hyperparameters.