A new framework improves reinforcement learning algorithms with policy guarantees.
problem Designing efficient and stable reinforcement learning algorithms.
method A general framework (FMA-PG) based on functional mirror ascent that constructs surrogate functions enabling policy improvement guarantees.
result The proposed framework enables policy improvement guarantees that hold regardless of policy parameterization, and recovers important heuristics.
Develops methods to create consistent surrogate models for agent-based simulators.
problem High computational costs and misjudgment of interventions in agent-based models.
method Causal abstractions to learn interventionally consistent surrogate models.
result Surrogates trained for interventional consistency closely mimic the agent-based model's behavior under interventions.
When learning from a batch of logged bandit feedback, the discrepancy between the policy to be learned and the off-policy training data imposes statistical and computational challenges. Unlike classical supervised learning and online learning settings, in batch contextual bandit learning, one only has access to a colle…
PPOS improves PPO by smoothing the surrogate objective function.
problem Performance instability and optimization inefficiency in PPO.
method Use of a functional clipping method instead of a flat clipping method.
result PPOS conducts more accurate updates and outperforms other PPO variants.
New algorithm learns optimal resource allocation in wireless systems without models.
problem Learning optimal resource allocation in wireless systems without system models.
method Developed a model-free primal-dual algorithm using smoothed surrogates of constrained problems.
result The algorithm can make the gap between optimal values and dual values arbitrarily small.
Adapts GRPO for off-policy RL, improving reward.
problem Improving training stability and efficiency in RL.
method Adapts GRPO to off-policy setting, uses clipped surrogate objectives.
result Off-policy GRPO outperforms on-policy GRPO in empirical tests.
The paper studies consistency of surrogate loss procedures under constrained classifiers.
problem Consistency of surrogate loss approaches under constrained classifiers without correct specification.
method The paper develops theoretical results and hinge loss based procedures for a constrained classification problem.
result Hinge losses are the only surrogate losses that preserve consistency in second-best scenarios.
Robust optimization and statistical robustness improve robot navigation policies.
problem Efficiently finding optimal robot navigation policies in uncertain environments.
method Combining robust optimization and statistical robustness with improved Bayesian optimization techniques.
result Safe and repeatable robot navigation policies are achieved with improved robust optimization methods.
Revisits PPO design choices, exposing failure modes and proposing alternatives.
problem Failure modes of standard PPO in new environments.
method Revisits standard PPO design choices, exposes failure modes, and proposes alternative approaches.
result Alternative design choices prevent failure modes in new environments.
We study the safe reinforcement learning problem with nonlinear function approximation, where policy optimization is formulated as a constrained optimization problem with both the objective and the constraint being nonconvex functions. For such a problem, we construct a sequence of surrogate convex constrained optimiza…
Recent policy optimization approaches have achieved substantial empirical success by constructing surrogate optimization objectives. The Approximate Policy Iteration objective (Schulman et al., 2015a; Kakade and Langford, 2002) has become a standard optimization target for reinforcement learning problems. Using this ob…
Efficient policy learning from observational data using weighted classification reductions.
problem Efficient policy evaluation does not necessarily lead to efficient estimation of policy parameters.
method Proposed an estimation approach based on generalized method of moments, efficient for policy parameters.
result Demonstrated empirical efficiency and regret benefits of a proposed method.
RRPI improves offline RL by optimizing policies against worst-case dynamics.
problem Offline RL's performance degrades under distribution shift and transition uncertainty.
method Formulates offline RL as robust policy optimization, treating transition kernel as decision variable.
result RRPI achieves strong average performance on D4RL benchmarks, outperforming recent baselines.
Seq2Seq models speed up epidemic model predictions.
problem Complex epidemic models are computationally expensive.
method Used deep seq2seq models as surrogates for complex models.
result Surrogates predict scenarios up to several thousand times faster.
New method improves solving combinatorial optimization problems with smoothed policies.
problem Solving combinatorial optimization problems repeatedly with varying instances.
method Smoothed policies with controlled random perturbations to linear oracle, leading to differentiable surrogate risk.
result Generalization bound decomposes excess risk into bias, estimation, and optimization components.
In this paper, we point out a fundamental property of the objective in reinforcement learning, with which we can reformulate the policy gradient objective into a perceptron-like loss function, removing the need to distinguish between on and off policy training. Namely, we posit that it is sufficient to only update a po…
Multi-step greedy policies have been extensively used in model-based reinforcement learning (RL), both when a model of the environment is available (e.g.,~in the game of Go) and when it is learned. In this paper, we explore their benefits in model-free RL, when employed using multi-step dynamic programming algorithms: …
CJE calibrates cheap LLM judges against an oracle, achieving high accuracy at a fraction of the cost.
problem Inexpensive LLM judges can produce biased rankings, leading to unreliable outcomes.
method CJE uses a small oracle to calibrate cheap scores, then evaluates at scale with valid uncertainty.
result CJE achieves 99% pairwise ranking accuracy at 14x lower cost compared to a 16x oracle/judge cost ratio.
Contextual policy search (CPS) is a class of multi-task reinforcement learning algorithms that is particularly useful for robotic applications. A recent state-of-the-art method is Contextual Covariance Matrix Adaptation Evolution Strategies (C-CMA-ES). It is based on the standard black-box optimization algorithm CMA-ES…
Managers of US National Forests must decide what policy to apply for dealing with lightning-caused wildfires. Conflicts among stakeholders (e.g., timber companies, home owners, and wildlife biologists) have often led to spirited political debates and even violent eco-terrorism. One way to transform these conflicts into…
PPO's gradients are heavy-tailed, affecting learning; a robust estimator improves performance.
problem Heavy-tailedness of PPO gradients causing learning issues.
method Characterized heavy-tailed gradients, identified likelihood ratios and advantages as sources, proposed GMOM as a robust estimator.
result GMOM improves PPO performance without clipping tricks.
We use surrogate losses to obtain several new regret bounds and new algorithms for contextual bandit learning. Using the ramp loss, we derive new margin-based regret bounds in terms of standard sequential complexity measures of a benchmark class of real-valued regression functions. Using the hinge loss, we derive an ef…
The paper interprets policy-gradient algorithms using continuation theory.
problem Optimizing nonconvex functions in reinforcement learning.
method Formulates policy optimization as optimization by continuation, interprets policy-gradient algorithms as implicitly optimizing deterministic policies.
result Exploration in policy-gradient algorithms is seen as computing a continuation of the return of the policy.
Policy optimization is a core component of reinforcement learning (RL), and most existing RL methods directly optimize parameters of a policy based on maximizing the expected total reward, or its surrogate. Though often achieving encouraging empirical success, its underlying mathematical principle on {\em policy-distri…
The study analyzes how neural reward models learn features for policy optimization in a Gaussian single-index model.
problem Reward modeling in policy optimization and its impact on downstream value.
method Two-stage neural reward model: first learns hidden direction, then fits readout layer.
result For any feature-learning temperature above a dimension-free threshold, a constant fraction of neurons recover the hidden direction.
Estimating the long-term effects of treatments is of interest in many fields. A common challenge in estimating such treatment effects is that long-term outcomes are unobserved in the time frame needed to make policy decisions. One approach to overcome this missing data problem is to analyze treatments effects on an int…
Accelerates policy optimization in RL with optimistic and adaptive updates.
problem Improving policy optimization methods in reinforcement learning.
method Integrates foresight into policy improvement step via optimistic and adaptive updates.
result Designs an optimistic policy gradient algorithm, adaptive via meta-gradient learning.
New PG losses improve decision optimization in misspecified models.
problem Improving decision optimization in models that are not perfectly specified.
method Introducing Perturbation Gradient (PG) losses to connect decision loss with directional derivatives and optimizing using gradient techniques.
result PG losses yield best-in-class policies asymptotically, even in misspecified settings.
Estimates long-term effects of new treatments using historical and short-term data.
problem Estimating long-term effects of novel treatments with limited historical data.
method Surrogate indices, dynamic treatment effect estimation, and double machine learning combined in a unified pipeline.
result Consistent and asymptotically normal estimates of long-term effects under Markovian assumption.
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.
Policy optimization is an effective reinforcement learning approach to solve continuous control tasks. Recent achievements have shown that alternating online and offline optimization is a successful choice for efficient trajectory reuse. However, deciding when to stop optimizing and collect new trajectories is non-triv…
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.
Bayesian approach uses Gaussian process for reinforcement learning.
problem Robotic locomotion environments
method Bayesian actor-critic, model-free reinforcement learning with Gaussian process for exploration and policy optimization.
result Gaussian process method outperforms current algorithms in robotic locomotion environments.
Efficient RL method uses SSL for small labeled data in health outcomes.
problem Lack of precise health outcome data for reinforcement learning.
method Semi-supervised learning approach for Q-learning and value estimation.
result Method efficiently estimates Q-function and value function, robust to mis-specification.
This work extends ME-RL using diffusion models to sample optimal policies.
problem Sampling from the optimal policy trajectory distribution in ME-RL.
method Introducing Diffusion-Augmented Markov Decision Processes (DA-MDPs) to minimize reverse KL divergence.
result DA-MDPs enable seamless integration into various ME-RL methods and outperform baselines.
A method to improve surrogate model accuracy using multiple fidelity models.
problem Efficiently combining models of varying accuracy and computational cost.
method Multifidelity Gaussian process models and leave-one-out cross-validation.
result Reduced LOO-CV error at the highest fidelity through adaptive learning.
POLAR learns efficient data acquisition policies using pretrained belief representations.
problem Challenges in learning effective policies for adaptive data acquisition.
method POLAR decouples representation learning from policy learning by leveraging pretrained predictive foundation models as belief-state encoders.
result POLAR outperforms state-of-the-art methods across diverse tasks while requiring fewer training samples.
Proposes a new sampling policy for ranking and selection problems.
problem Improving ranking and selection in adaptive sampling policies.
method Annealed entropic allocation, using soft-min weights and saddlepoint corrections.
result Consistently competitive performance in various settings.
What is a good exploration strategy for an agent that interacts with an environment in the absence of external rewards? Ideally, we would like to get a policy driving towards a uniform state-action visitation (highly exploring) in a minimum number of steps (fast mixing), in order to ease efficient learning of any goal-…
Study finds dividend payout policy positively impacts firm profitability.
problem Determining the optimal dividend payout ratio and its effect on financial performance.
method Panel data analysis of 60 Indian listed firms over 10 years, using ROA as a proxy for profitability.
result Positive and significant relationship between dividend payout policy and firm performance.
We study how the behavior of deep policy gradient algorithms reflects the conceptual framework motivating their development. To this end, we propose a fine-grained analysis of state-of-the-art methods based on key elements of this framework: gradient estimation, value prediction, and optimization landscapes. Our result…
Very recently proximal policy optimization (PPO) algorithms have been proposed as first-order optimization methods for effective reinforcement learning. While PPO is inspired by the same learning theory that justifies trust region policy optimization (TRPO), PPO substantially simplifies algorithm design and improves da…
TAMO optimizes multiple objectives in-context using transformers.
problem Balancing competing objectives in expensive, black-box problems.
method Fully amortized, transformer-based policy for multi-objective optimization.
result Significant speedup in proposal time with improved Pareto quality.
New algorithm uses machine learning to predict rewards for decision-making problems.
problem Sequential decision-making under uncertainty with scarce online data.
method Machine Learning-Assisted Upper Confidence Bound (MLA-UCB) algorithm.
result Proves to improve cumulative regret even with biased surrogate rewards.
Comment on entropy learning for dynamic treatment regimes.
problem Evaluating dynamic treatment regimes using entropy loss.
method Optimization-based alternative to IPW estimate.
result Suggests optimization-based approach for evaluation.
Trust region policy optimization (TRPO) is a popular and empirically successful policy search algorithm in Reinforcement Learning (RL) in which a surrogate problem, that restricts consecutive policies to be 'close' to one another, is iteratively solved. Nevertheless, TRPO has been considered a heuristic algorithm inspi…
Most policy search algorithms require thousands of training episodes to find an effective policy, which is often infeasible with a physical robot. This survey article focuses on the extreme other end of the spectrum: how can a robot adapt with only a handful of trials (a dozen) and a few minutes? By analogy with the wo…
The Exploration-Exploitation tradeoff arises in Reinforcement Learning when one cannot tell if a policy is optimal. Then, there is a constant need to explore new actions instead of exploiting past experience. In practice, it is common to resolve the tradeoff by using a fixed exploration mechanism, such as ε-greedy ex…