Proposes DRRO to mitigate over-optimization in RLHF from human feedback.
problem Over-optimization due to reward misspecification in RLHF.
method Wasserstein distributionally robust regret optimization (DRRO).
result DRRO mitigates over-optimization more effectively than existing baselines.
Study online learning with off-policy feedback in adversarial bandit problems.
problem Learning with limited direct feedback in sequential decision making.
method Proposed algorithms that adapt pessimistic reward estimators to handle unknown behavior policy.
result Guaranteed regret bounds scaling with policy mismatch, improving performance against well-covered comparators.
Paper tackles constrained bandit problems with a new learning framework.
problem Optimizing a black-box reward function subject to a black-box constraint function over a continuous space.
method Rectified Pessimistic-Optimistic Learning (RPOL) framework, incorporating optimistic and pessimistic GP bandit learning.
result RPOL achieves sublinear regret and minimal cumulative constraint violation.
Proposes Optimistic Pessimistically Initialised Q-Learning (OPIQ) for better exploration in RL.
problem Pessimistic initialisation of Q-values in deep RL leads to poor exploration performance.
method Augments pessimistically initialised Q-values with count-based bonuses to ensure optimism.
result OPIQ outperforms non-optimistic DQN variants in hard exploration tasks.
Dual active learning improves RLHF by selecting optimal conversations and teachers.
problem Efficiently aligning LLMs with human preferences using RLHF from feedback.
method Offline RL for conversation and teacher selection, dual active reward learning, pessimistic RL.
result The proposed algorithm achieves minimal generalized variance and outperforms state-of-the-arts.
A new algorithm for offline RL with trajectory-wise reward reduces bias and variance errors.
problem Offline RL with trajectory-wise reward incurs large bias and variance errors.
method PARTED algorithm that decomposes trajectory return into proxy rewards and performs pessimistic value iteration.
result PARTED achieves provably efficient suboptimality bounds in general MDPs with trajectory-wise reward.
The paper analyzes RLHF with human feedback and provides convergence results for MLE and pessimistic MLE.
problem Improving RLHF with human feedback from pairwise or K K K -wise comparisons. method Theoretical framework for RLHF with convergence analysis of MLE and pessimistic MLE.
result MLE fails but pessimistic MLE provides improved policies under certain coverage assumptions.
Semi-pessimistic RL tackles distributional shift and data scarcity in offline RL.
problem Distributional shift and scarcity of labeled data in offline RL.
method Proposes a semi-pessimistic RL method that simplifies learning by seeking a lower bound of the reward function.
result Demonstrates clear competitiveness and improved policy learning with vast unlabeled data.
Pessimistic estimator improves multi-objective policy optimization.
problem Optimizing multi-objective policies from existing data.
method Pessimistic estimator based on inverse propensity scores (IPS).
result Pessimistic estimator outperforms naive IPS estimator in theory and experiments.
POLAR optimizes treatment strategies in dynamic settings with statistical guarantees.
problem Optimizing sequential decisions in dynamic treatment regimes with robustness and statistical guarantees.
method Pessimistic model-based approach estimating transition dynamics and incorporating uncertainty penalties.
result Offers statistical and computational guarantees, including finite-sample bounds on policy suboptimality.
Paper tackles RLHF with DCPPO method, proving near-optimal suboptimality.
problem Challenges in offline RLHF with limited human feedback and bounded rationality.
method DCPPO method involving three stages: MLE, reward function recovery, and pessimistic value iteration.
result DCPPO's suboptimality almost matches classical pessimistic offline RL in terms of distribution shift and dimension.
New algorithm for reinforcement learning in uncertain environments with unknown thresholds.
problem Safety in reinforcement learning in unknown and uncertain environments.
method Growing-Window estimator sampling and Stochastic Pessimistic-Optimistic Thresholding (SPOT) algorithm.
result Achieves sublinear regret and constraint violation of i l d e O ( T ) ilde{\mathcal{O}}(\sqrt{T}) i l d e O ( T ) . New offline RL method handles average-reward MDPs with single-policy coverage.
problem Challenges in offline reinforcement learning due to distribution shift and non-uniform coverage.
method Develops an algorithm based on pessimistic discounted value iteration with quantile clipping.
result First fully single-policy sample complexity bound for average-reward offline RL.
It is well known that quantile regression model minimizes the portfolio extreme risk, whenever the attention is placed on the estimation of the response variable left quantiles. We show that, by considering the entire conditional distribution of the dependent variable, it is possible to optimize different risk and perf…
MOReL learns offline RL policies using pessimistic MDPs.
problem Offline RL's data efficiency and velocity.
method Two-step process: learn P-MDP and near-optimal policy in it.
result MOReL is minimax optimal and matches state-of-the-art results.
Improved DPO framework penalizes preference uncertainty to avoid overoptimization.
problem Aligning LLMs to human preferences is challenging due to varied, context-dependent, and ambiguous preferences.
method Developed a pessimistic framework for DPO by introducing preference uncertainty penalization schemes.
result Improved overall performance and better completions on high-uncertainty responses compared to vanilla DPO.
Algorithm learns optimal coordination for strategic agents in uncertain settings.
problem Optimizing rewards for strategic agents with private types and actions.
method Combines delaying mechanism, reward angle estimation, and LinUCB algorithm.
result Near optimal regret bound of O ~ ( T ) \tilde{O}(\sqrt{T}) O ~ ( T ) for learning optimal policy. Pessimistic Q-learning improves sample efficiency in offline reinforcement learning.
problem Insufficient coverage and sample scarcity in offline reinforcement learning datasets.
method Pessimistic Q-learning algorithm for offline reinforcement learning, focusing on variance reduction.
result Near-optimal sample complexity achieved with the proposed algorithm.
Study proposes a new risk measure for optimal portfolio allocation.
problem Challenges in estimating optimal portfolios based on pessimistic risk.
method Introduces uniform pessimistic risk and computational algorithm.
result Demonstrates the usefulness of the proposed risk and portfolio model with real data analysis.
Extends linear MDP to handle nonlinear rewards.
problem Restrictive linear MDP assumption limits real-world applicability.
method Proposes Generalized Linear MDP (GLMDP) with GLMs for rewards.
result Develops offline RL algorithms achieving suboptimality guarantees.
Combines experimental and historical data for robust policy evaluation.
problem Policy evaluation with mixed data sources, especially experimental vs historical.
method Linear integration of estimators from experimental and historical data, optimized for MSE minimization.
result Proposed estimators outperform traditional methods in ridesharing company data.
An algorithm for maximizing rewards under linear cost constraints.
problem Maximizing rewards while adhering to cost constraints in a linear bandit problem.
method Proposes an upper-confidence bound algorithm called optimistic pessimistic linear bandit (OPLB) for constrained contextual linear bandits.
result Proves an O ~ ( d T τ − c 0 ) \widetilde{\mathcal{O}}(\frac{d\sqrt{T}}{τ-c_0}) O ( τ − c 0 d T ) bound on regret for the proposed algorithm. Unified framework for risk-aware policy learning in contextual bandits.
problem Optimizing decision rules in high-stakes domains with adverse outcomes.
method Distributional framework for Lipschitz-continuous risk functionals, with novel empirical concentration inequalities.
result Data-dependent suboptimality bounds with an i l d e O ( 1 / n ) ilde{\mathcal{O}}(1/\sqrt{n}) i l d e O ( 1/ n ) rate, matching risk-neutral offline policy optimization. Improves decision complexity in hybrid environments.
problem Complexity in hybrid decision-making problems.
method General extension of DEC framework, model aggregation approach.
result Improved regret bounds for linear Q*/V* MDPs.
Paper develops neural network approximation for pessimistic offline RL with theoretical guarantees.
problem Challenges in offline reinforcement learning with deep neural networks and data dependence.
method Establishes estimation error for pessimistic offline RL using neural network approximation with C \mathcal{C} C -mixing data. result Explicit efficiency of deep adversarial offline RL frameworks demonstrated with two converging error components.
New method learns optimal policies in presence of unmeasured confounders.
problem Optimal policy learning with unobserved confounders.
method Causal-assisted policy learning methods using instrumental variables and negative controls.
result Policies are i l d e O ( n − 1 / 2 ) ilde{\mathscr{O}}(n^{-1/2}) i l d e O ( n − 1/2 ) quantile-optimal under mild coverage assumptions. This work improves RLHF performance guarantees using greedy sampling.
problem Learning the KL-regularized target with preference feedback.
method General preference model, greedy sampling.
result Order-wise improvements over existing RLHF performance guarantees.
Paper optimizes GAIL for online and offline learning with linear approximations.
problem Imitation learning from expert demonstrations with linear function approximations.
method Proposes optimistic and pessimistic algorithms for online and offline settings.
result Proves optimality and efficiency of proposed algorithms.
Study optimal product assortment using historical data, proving item coverage suffices.
problem Offline assortment optimization under MNL model with limited historical data.
method Pessimistic Rank-Breaking (PRB) algorithm combining rank-breaking and pessimistic estimation.
result Optimal item coverage is both sufficient and necessary for efficient offline learning.
PESCAL uses mediators to learn from confounded offline data.
problem Learning from confounded observational data in reinforcement learning.
method PESCAL uses mediator variables and the pessimistic principle to address confounding bias and distributional shift.
result It is sufficient to learn a lower bound of the mediator distribution function to mitigate distributional shift.
Pessimistic Minimax Value Iteration finds efficient NE policies from offline data.
problem Finding an approximate Nash equilibrium in offline Markov games with non-uniform coverage.
method Pessimistic Minimax Value Iteration (PMVI) constructs pessimistic value function estimates and solves NEs.
result Established a nearly minimax optimal result for offline Markov games with function approximation.
New algorithm reduces best-in-class regret in contextual bandits.
problem Compete with the best policy in a class without model restrictions.
method Proposes an algorithm that updates policies by minimizing a pessimistic objective, including a clipped inverse-propensity estimate and variance penalty.
result Achieves fast best-in-class regret rates, including polylogarithmic rates in the parametric case.
We study the problem of offline policy optimization in stochastic contextual bandit problems, where the goal is to learn a near-optimal policy based on a dataset of decision data collected by a suboptimal behavior policy. Rather than making any structural assumptions on the reward function, we assume access to a given …
Trajectory-level supervision allows efficient offline reinforcement learning.
problem Offline reinforcement learning
method Developing a statistical theory for offline policy optimization from trajectory-level labels
result Proving a high-probability guarantee of order O ~ ( H 2 C s a ( π ⋆ ) / n ) \widetilde O(H^2\sqrt{C_{sa}(\pi^\star)/n}) O ( H 2 C s a ( π ⋆ ) / n ) A new approach estimates propagators for trading risky assets.
problem Estimating price impact kernel from static data for optimal trading.
method Nonparametric estimation of propagator using offline reinforcement learning.
result Pessimistic trading strategy optimises execution costs under uncertainty.
This work improves policy evaluation and selection using logarithmic smoothing for pessimistic off-policy estimation.
problem Offline evaluation and selection of policies from past data.
method Develops novel concentration bounds and a logarithmically smoothed estimator (LS) for improved policy selection and learning.
result The logarithmically smoothed estimator (LS) provides tighter bounds and better policy selection and learning.
Paper tackles robust decision-making from multiple sites with shared structure.
problem Learning robust sequential decisions from heterogeneous multi-site datasets.
method Group-Robust MDPs with d-rectangular uncertainty sets, feature-wise worst-case aggregation, and cluster-level pooling.
result Proves suboptimality bound for robust planning policy under robust partial coverage assumption.
Paper proposes a method to optimize policies for diverse individuals using heterogeneous data.
problem Learning optimal policies for a heterogeneous population from pre-collected data.
method Individualized offline policy optimization framework for heterogeneous MDPs.
result The proposed P4L algorithm achieves a fast rate of average regret.
Papers learn from data to make decisions without interacting, improving on previous methods.
problem Achieving optimal decision-making from offline data with non-linear function approximation.
method Pessimistic Nonlinear Least-Square Value Iteration (PNLSVI) with three innovative components.
result Achieves minimax optimal instance-dependent regret for non-linear function approximation.
Study tackles RLHF with diverse human feedback, showing limitations and proposing a meta-learning approach.
problem Traditional RLHF fails to balance diverse human preferences.
method Integrates meta-learning and multiple social welfare functions to optimize diverse preferences.
result Establishes sample complexity bounds for optimizing diverse social welfare functions.
Study tight offline learning bounds for linear MDPs using variance information.
problem Understanding statistical limits with linear function representations in offline reinforcement learning.
method Variance-aware pessimistic value iteration (VAPVI) that reweights Bellman residuals based on estimated variances.
result Improved offline learning bounds expressed in terms of system quantities.
PQR estimates reward functions from actions and states without assuming state-only rewards.
problem Estimating reward functions from actions and states without state-only assumptions.
method Deep learning approach that sequentially estimates policy, Q-function, and reward.
result PQR uniquely recovers true reward with known transitions and bounds error with unknown transitions.
Q-Distribution Guided Q-Learning corrects overestimation of uncertain OOD actions in offline RL.
problem Overestimation of Q-values for out-of-distribution actions in offline reinforcement learning.
method QDQ applies a pessimistic adjustment to Q-values in uncertain OOD regions based on a consistency model.
result QDQ improves performance on the D4RL benchmark and achieves significant improvements across many tasks.
Estimates rewards from a demonstrator's learning process.
problem Estimating rewards from a demonstrator's behavior.
method Leveraging the demonstrator's exploration phase for reward estimation.
result Consistent reward estimation possible without identifiability issues.
A new algorithm CAP learns optimal policies from observational data with confounding bias and missing observations.
problem Offline contextual bandit with confounding bias and missing observations.
method CAP policy learning, forming reward function as solution of integral equation system, building confidence set, and greedily taking action with pessimism.
result Developed an upper bound to the suboptimality of CAP for the offline contextual bandit problem.
TT-DAC-PS: A deterministic actor-critic approach for optimal trade execution
problem Optimal execution of large stock sell programs
method Twin-Target Deterministic Actor-Critic with Policy Smoothing
result Reduces mean implementation shortfall percentage
This study tackles adversarial corruption in model-based reinforcement learning.
problem Adversarial corruption in model-based reinforcement learning.
method Maximum likelihood estimation (MLE) approach for learning transition model in both online and offline settings.
result Proves a regret of i l d e O ( T + C ) ilde{\mathcal{O}}(\sqrt{T} + C) i l d e O ( T + C ) for CR-OMLE and a suboptimality of O ( C / n ) \mathcal{O}(C/n) O ( C / n ) for CR-PMLE. Adaptive smooth non-stationary bandits achieve optimal regret rates without knowing parameters.
problem Smooth non-stationary bandits with Hölder class rewards.
method Established optimal dynamic regret rate and adaptive algorithm.
result Optimal dynamic regret can be attained adaptively without knowing Hölder exponent and coefficient.