NeurIPS 2024 论文列表

We study the evaluation of a policy under best- and worst-case perturbations to a Markov decision process (MDP), using transition observations from the original MDP, whether they are generated under the same or a different policy. This is an important problem when there is the possibility of a shift between historical and future environments, \emph{e.g.} due to unmeasured confounding, distributional shift, or an adversarial environment. We propose a perturbation model that allows changes in the transition kernel densities up to a given multiplicative factor or its reciprocal, extending the classic marginal sensitivity model (MSM) for single time-step decision-making to infinite-horizon RL. We characterize the sharp bounds on policy value under this model -- \emph{i.e.}, the tightest possible bounds based on transition observations from the original MDP -- and we study the estimation of these bounds from such transition observations. We develop an estimator with several important guarantees: it is semiparametrically efficient, and remains so even when certain necessary nuisance functions, such as worst-case Q-functions, are estimated at slow, nonparametric rates. Our estimator is also asymptotically normal, enabling straightforward statistical inference using Wald confidence intervals. Moreover, when certain nuisances are estimated inconsistently, the estimator still provides valid, albeit possibly not sharp, bounds on the policy value. We validate these properties in numerical simulations. The combination of accounting for environment shifts from train to test (robustness), being insensitive to nuisance-function estimation (orthogonality), and addressing the challenge of learning from finite samples (inference) together leads to credible and reliable policy evaluation.

Large Language Models (LLMs) have demonstrated proficiency in utilizing various tools by coding, yet they face limitations in handling intricate logic and precise control. In embodied tasks, high-level planning is amenable to direct coding, while low-level actions often necessitate task-specific refinement, such as Reinforcement Learning (RL). To seamlessly integrate both modalities, we introduce a two-level hierarchical framework, RL-GPT, comprising a slow agent and a fast agent. The slow agent analyzes actions suitable for coding, while the fast agent executes coding tasks. This decomposition effectively focuses each agent on specific tasks, proving highly efficient within our pipeline. Our approach outperforms traditional RL methods and existing GPT agents, demonstrating superior efficiency. In the Minecraft game, it rapidly obtains diamonds within a single day on an RTX3090. Additionally, it achieves SOTA performance across all designated MineDojo tasks.

Reward specification is one of the most tricky problems in Reinforcement Learning, which usually requires tedious hand engineering in practice. One promising approach to tackle this challenge is to adopt existing expert video demonstrations for policy learning. Some recent work investigates how to learn robot policies from only a single/few expert video demonstrations. For example, reward labeling via Optimal Transport (OT) has been shown to be an effective strategy to generate a proxy reward by measuring the alignment between the robot trajectory and the expert demonstrations. However, previous work mostly overlooks that the OT reward is invariant to temporal order information, which could bring extra noise to the reward signal. To address this issue, in this paper, we introduce the Temporal Optimal Transport (TemporalOT) reward to incorporate temporal order information for learning a more accurate OT-based proxy reward. Extensive experiments on the Meta-world benchmark tasks validate the efficacy of the proposed method. Our code is available at: https://github.com/fuyw/TemporalOT.

This paper studies Generative Flow Networks (GFlowNets), which learn to sample objects proportionally to a given reward function through the trajectory of state transitions. In this work, we observe that GFlowNets tend to under-exploit the high-reward objects due to training on insufficient number of trajectories, which may lead to a large gap between the estimated flow and the (known) reward value. In response to this challenge, we propose a pessimistic backward policy for GFlowNets (PBP-GFN), which maximizes the observed flow to align closely with the true reward for the object. We extensively evaluate PBP-GFN across eight benchmarks, including hyper-grid environment, bag generation, structured set generation, molecular generation, and four RNA sequence generation tasks. In particular, PBP-GFN enhances the discovery of high-reward objects, maintains the diversity of the objects, and consistently outperforms existing methods.

Computation graphs are Directed Acyclic Graphs (DAGs) where the nodes correspond to mathematical operations and are used widely as abstractions in optimizations of neural networks. The device placement problem aims to identify optimal allocations of those nodes to a set of (potentially heterogeneous) devices. Existing approaches rely on two types of architectures known as grouper-placer and encoder-placer, respectively. In this work, we bridge the gap between encoder-placer and grouper-placer techniques and propose a novel framework for the task of device placement, relying on smaller computation graphs extracted from the OpenVINO toolkit. The framework consists of five steps, including graph coarsening, node representation learning and policy optimization. It facilitates end-to-end training and takes into account the DAG nature of the computation graphs. We also propose a model variant, inspired by graph parsing networks and complex network analysis, enabling graph representation learning and jointed, personalized graph partitioning, using an unspecified number of groups. To train the entire framework, we use reinforcement learning using the execution time of the placement as a reward. We demonstrate the flexibility and effectiveness of our approach through multiple experiments with three benchmark models, namely Inception-V3, ResNet, and BERT. The robustness of the proposed framework is also highlighted through an ablation study. The suggested placements improve the inference speed for the benchmark models by up to $58.2\%$ over CPU execution and by up to $60.24\%$ compared to other commonly used baselines.

User intentions are typically formalized as evaluation rewards to be maximized when fine-tuning language models (LMs). Existing alignment methods, such as Direct Preference Optimization (DPO), are mainly tailored for pairwise preference data where rewards are implicitly defined rather than explicitly given. In this paper, we introduce a general framework for LM alignment, leveraging Noise Contrastive Estimation (NCE) to bridge the gap in handling reward datasets explicitly annotated with scalar evaluations. Our framework comprises two parallel algorithms, NCA and InfoNCA, both enabling the direct extraction of an LM policy from reward data as well as preference data. Notably, we show that the DPO loss is a special case of our proposed InfoNCA objective under pairwise preference settings, thereby integrating and extending current alignment theories. By comparing NCA and InfoNCA, we demonstrate that the well-observed decreasing-likelihood trend of DPO/InfoNCA is caused by their focus on adjusting relative likelihood across different responses. In contrast, NCA optimizes the absolute likelihood for each response, thereby effectively preventing the chosen likelihood from decreasing. We evaluate our methods in both reward and preference settings with Mistral-8$\times$7B and 7B models. Experiments suggest that InfoNCA/NCA surpasses various preference baselines when reward datasets are available. We also find NCA significantly outperforms DPO in complex reasoning tasks like math and coding.

We present a novel method, Contextual goal-Oriented Data Augmentation (CODA), which uses commonly available unlabeled trajectories and context-goal pairs to solve Contextual Goal-Oriented (CGO) problems. By carefully constructing an action-augmented MDP that is equivalent to the original MDP, CODA creates a fully labeled transition dataset under training contexts without additional approximation error. We conduct a novel theoretical analysis to demonstrate CODA's capability to solve CGO problems in the offline data setup. Empirical results also showcase the effectiveness of CODA, which outperforms other baseline methods across various context-goal relationships of CGO problem. This approach offers a promising direction to solving CGO problems using offline datasets.

The recent rapid advancement of machine learning has been driven by increasingly powerful models with the growing availability of training data and computational resources. However, real-time decision-making tasks with limited time and sparse learning signals remain challenging. One way of improving the learning speed and performance of these agents is to leverage human guidance. In this work, we introduce GUIDE, a framework for real-time human-guided reinforcement learning by enabling continuous human feedback and grounding such feedback into dense rewards to accelerate policy learning. Additionally, our method features a simulated feedback module that learns and replicates human feedback patterns in an online fashion, effectively reducing the need for human input while allowing continual training. We demonstrate the performance of our framework on challenging tasks with sparse rewards and visual observations. Our human study involving 50 subjects offers strong quantitative and qualitative evidence of the effectiveness of our approach. With only 10 minutes of human feedback, our algorithm achieves up to 30\% increase in success rate compared to its RL baseline.

LLMs can now act as autonomous agents that interact with digital environments and complete specific objectives (e.g., arranging an online meeting). However, accuracy is still far from satisfactory, partly due to a lack of large-scale, direct demonstrations for digital tasks. Obtaining supervised data from humans is costly, and automatic data collection through exploration or reinforcement learning relies on complex environmental and content setup, resulting in datasets that lack comprehensive coverage of various scenarios. On the other hand, there is abundant knowledge that may indirectly assist task completion, such as online tutorials that were created for human consumption. In this work, we present Synatra, an approach that effectively transforms this indirect knowledge into direct supervision at scale. We define different types of indirect knowledge, and carefully study the available sources to obtain it, methods to encode the structure of direct demonstrations, and finally methods to transform indirect knowledge into direct demonstrations. We use 100k such synthetically-created demonstrations to finetune a 7B CodeLlama, and demonstrate that the resulting agent surpasses all comparably sized models on three web-based task benchmarks Mind2Web, MiniWoB++ and WebArena, as well as surpassing GPT-3.5 on WebArena and Mind2Web. In addition, while synthetic demonstrations prove to be only 3% the cost of human demonstrations (at $0.031 each), we show that the synthetic demonstrations can be more effective than an identical number of human demonstrations collected from limited domains.

Visual text design plays a critical role in conveying themes, emotions, and atmospheres in multimodal formats such as film posters and album covers. Translating these visual and textual elements across languages extends the concept of translation beyond mere text, requiring the adaptation of aesthetic and stylistic features. To address this, we introduce a novel task of Multimodal Style Translation (MuST-Bench), a benchmark designed to evaluate the ability of visual text generation models to perform translation across different writing systems while preserving design intent. Our initial experiments on MuST-Bench reveal that existing visual text generation models struggle with the proposed task due to the inadequacy of textual descriptions in conveying visual design. In response, we introduce SIGIL, a framework for multimodal style translation that eliminates the need for style descriptions. SIGIL enhances image generation models through three innovations: glyph latent for multilingual settings, pre-trained VAEs for stable style guidance, and an OCR model with reinforcement learning feedback for optimizing readable character generation. SIGIL outperforms existing baselines by achieving superior style consistency and legibility while maintaining visual fidelity, setting itself apart from traditional description-based approaches. We release MuST-Bench publicly for broader use and exploration https://huggingface.co/datasets/yejinc/MuST-Bench.

Existing Goal-Conditioned Reinforcement Learning (GCRL) algorithms are built upon Hindsight Experience Replay (HER), which densifies rewards through hindsight replay and leverages historical goal-achieving information to construct a learning curriculum. However, when the task is characterized by a non-Markovian reward (NMR), whose computation depends on multiple steps of states and actions, HER can no longer densify rewards by treating a single encountered state as the hindsight goal. The lack of informative rewards hinders policy learning, resulting in rolling out failed trajectories. Consequently, the replay buffer is overwhelmed with failed trajectories, impeding the establishment of an applicable curriculum. To circumvent these limitations, we deviate from existing HER-based methods and propose an on-policy GCRL framework, GCPO, which is applicable to both multi-goal Markovian reward (MR) and NMR problems. GCPO consists of (1) Pre-training from Demonstrations, which pre-trains the policy to possess an initial goal-achieving capability, thereby diminishing the difficulty of subsequent online learning. (2) Online Self-Curriculum Learning, which first estimates the policy's goal-achieving capability based on historical evaluation information and then selects progressively challenging goals for learning based on its current capability. We evaluate GCPO on a challenging multi-goal long-horizon task: fixed-wing UAV velocity vector control. Experimental results demonstrate that GCPO is capable of effectively addressing both multi-goal MR and NMR problems.

Deep reinforcement learning (RL) is a powerful approach to complex decision-making. However, one issue that limits its practical application is its brittleness, sometimes failing to train in the presence of small changes in the environment. Motivated by the success of zero-shot transfer—where pre-trained models perform well on related tasks—we consider the problem of selecting a good set of training tasks to maximize generalization performance across a range of tasks. Given the high cost of training, it is critical to select training tasks strategically, but not well understood how to do so. We hence introduce Model-Based Transfer Learning (MBTL), which layers on top of existing RL methods to effectively solve contextual RL problems. MBTL models the generalization performance in two parts: 1) the performance set point, modeled using Gaussian processes, and 2) performance loss (generalization gap), modeled as a linear function of contextual similarity. MBTL combines these two pieces of information within a Bayesian optimization (BO) framework to strategically select training tasks. We show theoretically that the method exhibits sublinear regret in the number of training tasks and discuss conditions to further tighten regret bounds. We experimentally validate our methods using urban traffic and standard continuous control benchmarks. The experimental results suggest that MBTL can achieve up to 43x improved sample efficiency compared with canonical independent training and multi-task training. Further experiments demonstrate the efficacy of BO and the insensitivity to the underlying RL algorithm and hyperparameters. This work lays the foundations for investigating explicit modeling of generalization, thereby enabling principled yet effective methods for contextual RL. Code is available at https://github.com/jhoon-cho/MBTL/.

Reinforcement learning (RL) in large or infinite state spaces is notoriously challenging, both theoretically (where worst-case sample and computational complexities must scale with state space cardinality) and experimentally (where function approximation and policy gradient techniques often scale poorly and suffer from instability and high variance). One line of research attempting to address these difficulties makes the natural assumption that we are given a collection of base or *constituent* policies (possibly heuristic) upon which we would like to improve in a scalable manner. In this work we aim to compete with the *max-following policy*, which at each state follows the action of whichever constituent policy has the highest value. The max-following policy is always at least as good as the best constituent policy, and may be considerably better. Our main result is an efficient algorithm that learns to compete with the max-following policy, given only access to the constituent policies (but not their value functions). In contrast to prior work in similar settings, our theoretical results require only the minimal assumption of an ERM oracle for value function approximation for the constituent policies (and not the global optimal policy or the max-following policy itself) on samplable distributions. We illustrate our algorithm's experimental effectiveness and behavior on several robotic simulation testbeds.

Multi-objective reinforcement learning (MORL) is an excellent framework for multi-objective sequential decision-making. MORL employs a utility function to aggregate multiple objectives into one that expresses a user's preferences. However, MORL still misses two crucial theoretical analyses of the properties of utility functions: (1) a characterisation of the utility functions for which an associated optimal policy exists, and (2) a characterisation of the types of preferences that can be expressed as utility functions. As a result, we formally characterise the families of preferences and utility functions that MORL should focus on: those for which an optimal policy is guaranteed to exist. We expect our theoretical results to promote the development of novel MORL algorithms that exploit our theoretical findings.

Recommending out-of-vocabulary (OOV) items is a challenging problem since the in-vocabulary (IV) items have well-trained behavioral embeddings but the OOV items only have content features. Current OOV recommendation models often generate 'makeshift' embeddings for OOV items from content features and then jointly recommend with the `makeshift' OOV item embeddings and the behavioral IV item embeddings. However, merely using the 'makeshift' embedding will result in suboptimal recommendation performance due to the substantial gap between the content feature and the behavioral embeddings. To bridge the gap, we propose a novel **User Sequence IMagination (USIM)** fine-tuning framework, which first imagines the user sequences and then refines the generated OOV embeddings with the user behavioral embeddings. Specifically, we frame the user sequence imagination as a reinforcement learning problem and develop a recommendation-focused reward function to evaluate to what extent a user can help recommend the OOV items. Besides, we propose an embedding-driven transition function to model the embedding transition after imaging a user. USIM has been deployed on a prominent e-commerce platform for months, offering recommendations for millions of OOV items and billions of users. Extensive experiments demonstrate that USIM outperforms traditional generative models in OOV item recommendation performance across traditional collaborative filtering and GNN-based collaborative filtering models.

We study online control for continuous-time linear systems with finite sampling rates, where the objective is to design an online procedure that learns under non-stochastic noise and performs comparably to a fixed optimal linear controller. We present a novel two-level online algorithm, by integrating a higher-level learning strategy and a lower-level feedback control strategy. This method offers a practical and robust solution for online control, which achieves sublinear regret. Our work provides the first nonasymptotic results for controlling continuous-time linear systems with finite number of interactions with the system. Moreover, we examine how to train an agent in domain randomization environments from a non-stochastic control perspective. By applying our method to the SAC (Soft Actor-Critic) algorithm, we achieved improved results in multiple reinforcement learning tasks within domain randomization environments. Our work provides new insights into non-asymptotic analyses of controlling continuous-time systems. Furthermore, our work brings practical intuition into controller learning under non-stochastic environments.

In order to mitigate the sample complexity of real-world reinforcement learning, common practice is to first train a policy in a simulator where samples are cheap, and then deploy this policy in the real world, with the hope that it generalizes effectively. Such \emph{direct sim2real} transfer is not guaranteed to succeed, however, and in cases where it fails, it is unclear how to best utilize the simulator. In this work, we show that in many regimes, while direct sim2real transfer may fail, we can utilize the simulator to learn a set of \emph{exploratory} policies which enable efficient exploration in the real world. In particular, in the setting of low-rank MDPs, we show that coupling these exploratory policies with simple, practical approaches---least-squares regression oracles and naive randomized exploration---yields a polynomial sample complexity in the real world, an exponential improvement over direct sim2real transfer, or learning without access to a simulator. To the best of our knowledge, this is the first evidence that simulation transfer yields a provable gain in reinforcement learning in settings where direct sim2real transfer fails. We validate our theoretical results on several realistic robotic simulators and a real-world robotic sim2real task, demonstrating that transferring exploratory policies can yield substantial gains in practice as well.

We propose a new algorithm for model-based distributional reinforcement learning (RL), and prove that it is minimax-optimal for approximating return distributions in the generative model regime (up to logarithmic factors), the first result of this kind for any distributional RL algorithm. Our analysis also provides new theoretical perspectives on categorical approaches to distributional RL, as well as introducing a new distributional Bellman equation, the stochastic categorical CDF Bellman equation, which we expect to be of independent interest. Finally, we provide an experimental study comparing a variety of model-based distributional RL algorithms, with several key takeaways for practitioners.

Reinforcement Learning from Human Feedback (RLHF) has recently surged in popularity, particularly for aligning large language models and other AI systems with human intentions. At its core, RLHF can be viewed as a specialized instance of Preference-based Reinforcement Learning (PbRL), where the preferences specifically originate from human judgments rather than arbitrary evaluators. Despite this connection, most existing approaches in both RLHF and PbRL primarily focus on optimizing a mean reward objective, neglecting scenarios that necessitate risk-awareness, such as AI safety, healthcare, and autonomous driving. These scenarios often operate under a one-episode-reward setting, which makes conventional risk-sensitive objectives inapplicable. To address this, we explore and prove the applicability of two risk-aware objectives to PbRL : nested and static quantile risk objectives. We also introduce Risk-AwarePbRL (RA-PbRL), an algorithm designed to optimize both nested and static objectives. Additionally, we provide a theoretical analysis of the regret upper bounds, demonstrating that they are sublinear with respect to the number of episodes, and present empirical results to support our findings. Our code is available in https://github.com/aguilarjose11/PbRLNeurips.