ICML 2023 论文列表

General function approximation is a powerful tool to handle large state and action spaces in a broad range of reinforcement learning (RL) scenarios. However, theoretical understanding of non-stationary MDPs with general function approximation is still limited. In this paper, we make the first such an attempt. We first propose a new complexity metric called dynamic Bellman Eluder (DBE) dimension for non-stationary MDPs, which subsumes majority of existing tractable RL problems in static MDPs as well as non-stationary MDPs. Based on the proposed complexity metric, we propose a novel confidence-set based model-free algorithm called SW-OPEA, which features a sliding window mechanism and a new confidence set design for non-stationary MDPs. We then establish an upper bound on the dynamic regret for the proposed algorithm, and show that SW-OPEA is provably efficient as long as the variation budget is not significantly large. We further demonstrate via examples of non-stationary linear and tabular MDPs that our algorithm performs better in small variation budget scenario than the existing UCB-type algorithms. To the best of our knowledge, this is the first dynamic regret analysis in non-stationary MDPs with general function approximation.

We provide a theoretical framework for Reinforcement Learning with Human Feedback (RLHF). We show that when the underlying true reward is linear, under both Bradley-Terry-Luce (BTL) model (pairwise comparison) and Plackett-Luce (PL) model ($K$-wise comparison), MLE converges under certain semi-norm for the family of linear reward. On the other hand, when training a policy based on the learned reward model, we show that MLE fails while a pessimistic MLE provides policies with good performance under certain coverage assumption. We also show that under the PL model, both the true MLE and a different MLE which splits the $K$-wise comparison into pairwise comparisons converge, while the true MLE is asymptotically more efficient. Our results validate the empirical success of the existing RLHF algorithms, and provide new insights for algorithm design. Our analysis can also be applied for the problem of online RLHF and inverse reinforcement learning.

Reinforcement learning (RL) is able to solve complex sequential decision-making tasks but is currently limited by sample efficiency and required computation. To improve sample efficiency, recent work focuses on model-based RL which interleaves model learning with planning. Recent methods further utilize policy learning, value estimation, and, self-supervised learning as auxiliary objectives. In this paper we show that, surprisingly, a simple representation learning approach relying only on a latent dynamics model trained by latent temporal consistency is sufficient for high-performance RL. This applies when using pure planning with a dynamics model conditioned on the representation, but, also when utilizing the representation as policy and value function features in model-free RL. In experiments, our approach learns an accurate dynamics model to solve challenging high-dimensional locomotion tasks with online planners while being 4.1$\times$ faster to train compared to ensemble-based methods. With model-free RL without planning, especially on high-dimensional tasks, such as the Deepmind Control Suite Humanoid and Dog tasks, our approach outperforms model-free methods by a large margin and matches model-based methods' sample efficiency while training 2.4$\times$ faster.

As previous representations for reinforcement learning cannot effectively incorporate a human-intuitive understanding of the 3D environment, they usually suffer from sub-optimal performances. In this paper, we present Semantic-aware Neural Radiance Fields for Reinforcement Learning (SNeRL), which jointly optimizes semantic-aware neural radiance fields (NeRF) with a convolutional encoder to learn 3D-aware neural implicit representation from multi-view images. We introduce 3D semantic and distilled feature fields in parallel to the RGB radiance fields in NeRF to learn semantic and object-centric representation for reinforcement learning. SNeRL outperforms not only previous pixel-based representations but also recent 3D-aware representations both in model-free and model-based reinforcement learning.

Unsupervised object-centric representation (OCR) learning has recently drawn attention as a new paradigm of visual representation. This is because of its *potential* of being an effective pre-training technique for various downstream tasks in terms of sample efficiency, systematic generalization, and reasoning. Although image-based reinforcement learning (RL) is one of the most important and thus frequently mentioned such downstream tasks, the benefit in RL has surprisingly not been investigated systematically thus far. Instead, most of the evaluations have focused on rather indirect metrics such as segmentation quality and object property prediction accuracy. In this paper, we investigate the effectiveness of OCR pre-training for image-based reinforcement learning via empirical experiments. For systematic evaluation, we introduce a simple object-centric visual RL benchmark and conduct experiments to answer questions such as "Does OCR pre-training improve performance on object-centric tasks?" and "Can OCR pre-training help with out-of-distribution generalization?". Our results provide empirical evidence for valuable insights into the effectiveness of OCR pre-training for RL and the potential limitations of its use in certain scenarios. Additionally, this study also examines the critical aspects of incorporating OCR pre-training in RL, including performance in a visually complex environment and the appropriate pooling layer to aggregate the object representations.

We consider solving sequential decision-making problems in the scenario where the agent has access to two supervision sources: $\textit{reward signal}$ and a $\textit{teacher}$ that can be queried to obtain a $\textit{good}$ action for any state encountered by the agent. Learning solely from rewards, or reinforcement learning, is data inefficient and may not learn high-reward policies in challenging scenarios involving sparse rewards or partial observability. On the other hand, learning from a teacher may sometimes be infeasible. For instance, the actions provided by a teacher with privileged information may be unlearnable by an agent with limited information (i.e., partial observability). In other scenarios, the teacher might be sub-optimal, and imitating their actions can limit the agent's performance. To overcome these challenges, prior work proposed to jointly optimize imitation and reinforcement learning objectives but relied on heuristics and problem-specific hyper-parameter tuning to balance the two objectives. We introduce Teacher Guided Reinforcement Learning (TGRL), a principled approach to dynamically balance following the teacher's guidance and leveraging RL. TGRL outperforms strong baselines across diverse domains without hyperparameter tuning.

We study reinforcement learning (RL) with linear function approximation. For episodic time-inhomogeneous linear Markov decision processes (linear MDPs) whose transition probability can be parameterized as a linear function of a given feature mapping, we propose the first computationally efficient algorithm that achieves the nearly minimax optimal regret $\tilde O(d\sqrt{H^3K})$, where $d$ is the dimension of the feature mapping, $H$ is the planning horizon, and $K$ is the number of episodes. Our algorithm is based on a weighted linear regression scheme with a carefully designed weight, which depends on a new variance estimator that (1) directly estimates the variance of the *optimal* value function, (2) monotonically decreases with respect to the number of episodes to ensure a better estimation accuracy, and (3) uses a rare-switching policy to update the value function estimator to control the complexity of the estimated value function class. Our work provides a complete answer to optimal RL with linear MDPs, and the developed algorithm and theoretical tools may be of independent interest.

In online reinforcement learning (RL), instead of employing standard structural assumptions on Markov decision processes (MDPs), using a certain coverage condition (original from offline RL) is enough to ensure sample-efficient guarantees (Xie et al. 2023). In this work, we focus on this new direction by digging more possible and general coverage conditions, and study the potential and the utility of them in efficient online RL. We identify more concepts, including the $L^p$ variant of concentrability, the density ratio realizability, and trade-off on the partial/rest coverage condition, that can be also beneficial to sample-efficient online RL, achieving improved regret bound. Furthermore, if exploratory offline data are used, under our coverage conditions, both statistically and computationally efficient guarantees can be achieved for online RL. Besides, even though the MDP structure is given, e.g., linear MDP, we elucidate that, good coverage conditions are still beneficial to obtain faster regret bound beyond $\widetilde{\mathcal{O}}(\sqrt{T})$ and even a logarithmic order regret. These results provide a good justification for the usage of general coverage conditions in efficient online RL.

In Reinforcement Learning (RL), an agent acts in an unknown environment to maximize the expected cumulative discounted sum of an external reward signal, i.e., the expected return. In practice, in many tasks of interest, such as policy optimization, the agent usually spends its interaction budget by collecting episodes of *fixed length* within a simulator (i.e., Monte Carlo simulation). However, given the discounted nature of the RL objective, this data collection strategy might not be the best option. Indeed, the rewards taken in early simulation steps weigh exponentially more than future rewards. Taking a cue from this intuition, in this paper, we design an a-priori budget allocation strategy that leads to the collection of trajectories of different lengths, i.e., *truncated*. The proposed approach provably minimizes the width of the confidence intervals around the empirical estimates of the expected return of a policy. After discussing the theoretical properties of our method, we make use of our trajectory truncation mechanism to extend Policy Optimization via Importance Sampling (POIS, Metelli et al., 2018) algorithm. Finally, we conduct a numerical comparison between our algorithm and POIS: the results are consistent with our theory and show that an appropriate truncation of the trajectories can succeed in improving performance.

Although distributional reinforcement learning (DRL) has been widely examined in the past few years, very few studies investigate the validity of the obtained Q-function estimator in the distributional setting. To fully understand how the approximation errors of the Q-function affect the whole training process, we do some error analysis and theoretically show how to reduce both the bias and the variance of the error terms. With this new understanding, we construct a new estimator Quantiled Expansion Mean (QEM) and introduce a new DRL algorithm (QEMRL) from the statistical perspective. We extensively evaluate our QEMRL algorithm on a variety of Atari and Mujoco benchmark tasks and demonstrate that QEMRL achieves significant improvement over baseline algorithms in terms of sample efficiency and convergence performance.

Offline reinforcement learning suffers from the out-of-distribution issue and extrapolation error. Most policy constraint methods regularize the density of the trained policy towards the behavior policy, which is too restrictive in most cases. We propose Supported Trust Region optimization (STR) which performs trust region policy optimization with the policy constrained within the support of the behavior policy, enjoying the less restrictive support constraint. We show that, when assuming no approximation and sampling error, STR guarantees strict policy improvement until convergence to the optimal support-constrained policy in the dataset. Further with both errors incorporated, STR still guarantees safe policy improvement for each step. Empirical results validate the theory of STR and demonstrate its state-of-the-art performance on MuJoCo locomotion domains and much more challenging AntMaze domains.

Reinforcement learning (RL) has made significant advancements in artificial intelligence. However, its real-world applications are limited due to differences between simulated environments and the actual world. Consequently, it is crucial to systematically analyze how each component of the RL system can affect the final model performance. In this study, we propose an adversarial learning framework for distributional reinforcement learning, which adopts the concept of influence measure from the statistics community. This framework enables us to detect performance loss caused by either the internal policy structure or the external state observation. The proposed influence measure is based on information geometry and has desirable properties of invariance. We demonstrate that the influence measure is useful for three diagnostic tasks: identifying fragile states in trajectories, determining the instability of the policy architecture, and pinpointing anomalously sensitive policy parameters.

Many existing reinforcement learning (RL) methods employ stochastic gradient iteration on the back end, whose stability hinges upon a hypothesis that the data-generating process mixes exponentially fast with a rate parameter that appears in the step-size selection. Unfortunately, this assumption is violated for large state spaces or settings with sparse rewards, and the mixing time is unknown, making the step size inoperable. In this work, we propose an RL methodology attuned to the mixing time by employing a multi-level Monte Carlo estimator for the critic, the actor, and the average reward embedded within an actor-critic (AC) algorithm. This method, which we call **M**ulti-level **A**ctor-**C**ritic (MAC), is developed specifically for infinite-horizon average-reward settings and neither relies on oracle knowledge of the mixing time in its parameter selection nor assumes its exponential decay; it is therefore readily applicable to applications with slower mixing times. Nonetheless, it achieves a convergence rate comparable to SOTA actor-critic algorithms. We experimentally show that these alleviated restrictions on the technical conditions required for stability translate to superior performance in practice for RL problems with sparse rewards.

We study reinforcement learning with linear function approximation and adversarially changing cost functions, a setup that has mostly been considered under simplifying assumptions such as full information feedback or exploratory conditions. We present a computationally efficient policy optimization algorithm for the challenging general setting of unknown dynamics and bandit feedback, featuring a combination of mirror-descent and least squares policy evaluation in an auxiliary MDP used to compute exploration bonuses. Our algorithm obtains an $\widetilde O(K^{6/7})$ regret bound, improving significantly over previous state-of-the-art of $\widetilde O (K^{14/15})$ in this setting. In addition, we present a version of the same algorithm under the assumption a simulator of the environment is available to the learner (but otherwise no exploratory assumptions are made), and prove it obtains state-of-the-art regret of $\widetilde O (K^{2/3})$.

One of the fundamental quests of AI is to produce agents that coordinate well with humans. This problem is challenging, especially in domains that lack high quality human behavioral data, because multi-agent reinforcement learning (RL) often converges to different equilibria from the ones that humans prefer. We propose a novel framework, instructRL, that enables humans to specify what kind of strategies they expect from their AI partners through natural language instructions. We use pretrained large language models to generate a prior policy conditioned on the human instruction and use the prior to regularize the RL objective. This leads to the RL agent converging to equilibria that are aligned with human preferences. We show that instructRL converges to human-like policies that satisfy the given instructions in a proof-of-concept environment as well as the challenging Hanabi benchmark. Finally, we show that knowing the language instruction significantly boosts human-AI coordination performance in human evaluations in Hanabi.

Discount regularization, using a shorter planning horizon when calculating the optimal policy, is a popular choice to restrict planning to a less complex set of policies when estimating an MDP from sparse or noisy data (Jiang et al., 2015). It is commonly understood that discount regularization functions by de-emphasizing or ignoring delayed effects. In this paper, we reveal an alternate view of discount regularization that exposes unintended consequences. We demonstrate that planning under a lower discount factor produces an identical optimal policy to planning using any prior on the transition matrix that has the same distribution for all states and actions. In fact, it functions like a prior with stronger regularization on state-action pairs with more transition data. This leads to poor performance when the transition matrix is estimated from data sets with uneven amounts of data across state-action pairs. Our equivalence theorem leads to an explicit formula to set regularization parameters locally for individual state-action pairs rather than globally. We demonstrate the failures of discount regularization and how we remedy them using our state-action-specific method across simple empirical examples as well as a medical cancer simulator.

Safe reinforcement learning (RL) trains a constraint satisfaction policy by interacting with the environment. We aim to tackle a more challenging problem: learning a safe policy from an offline dataset. We study the offline safe RL problem from a novel multi-objective optimization perspective and propose the $\epsilon$-reducible concept to characterize problem difficulties. The inherent trade-offs between safety and task performance inspire us to propose the constrained decision transformer (CDT) approach, which can dynamically adjust the trade-offs during deployment. Extensive experiments show the advantages of the proposed method in learning an adaptive, safe, robust, and high-reward policy. CDT outperforms its variants and strong offline safe RL baselines by a large margin with the same hyperparameters across all tasks, while keeping the zero-shot adaptation capability to different constraint thresholds, making our approach more suitable for real-world RL under constraints.

Off-policy learning from multistep returns is crucial for sample-efficient reinforcement learning, but counteracting off-policy bias without exacerbating variance is challenging. Classically, off-policy bias is corrected in a per-decision manner: past temporal-difference errors are re-weighted by the instantaneous Importance Sampling (IS) ratio after each action via eligibility traces. Many off-policy algorithms rely on this mechanism, along with differing protocols for cutting the IS ratios (traces) to combat the variance of the IS estimator. Unfortunately, once a trace has been cut, the effect cannot be easily reversed. This has led to the development of credit-assignment strategies that account for multiple past experiences at a time. These trajectory-aware methods have not been extensively analyzed, and their theoretical justification remains uncertain. In this paper, we propose a multistep operator that unifies per-decision and trajectory-aware methods. We prove convergence conditions for our operator in the tabular setting, establishing the first guarantees for several existing methods as well as many new ones. Finally, we introduce Recency-Bounded Importance Sampling (RBIS), which leverages trajectory awareness to perform robustly across $\lambda$-values in an off-policy control task.

Natural agents can effectively learn from multiple data sources that differ in size, quality, and types of measurements. We study this heterogeneity in the context of offline reinforcement learning (RL) by introducing a new, practically motivated semi-supervised setting. Here, an agent has access to two sets of trajectories: labelled trajectories containing state, action and reward triplets at every timestep, along with unlabelled trajectories that contain only state and reward information. For this setting, we develop and study a simple meta-algorithmic pipeline that learns an inverse dynamics model on the labelled data to obtain proxy-labels for the unlabelled data, followed by the use of any offline RL algorithm on the true and proxy-labelled trajectories. Empirically, we find this simple pipeline to be highly successful --- on several D4RL benchmarks (Fu et al., 2020), certain offline RL algorithms can match the performance of variants trained on a fully labelled dataset even when we label only 10% of trajectories which are highly suboptimal. To strengthen our understanding, we perform a large-scale controlled empirical study investigating the interplay of data-centric properties of the labelled and unlabelled datasets, with algorithmic design choices (e.g., choice of inverse dynamics, offline RL algorithm) to identify general trends and best practices for training RL agents on semi-supervised offline datasets.