NeurIPS 2023 Papers

Robots often rely on a repertoire of previously-learned motion policies for performing tasks of diverse complexities. When facing unseen task conditions or when new task requirements arise, robots must adapt their motion policies accordingly. In this context, policy optimization is the \emph{de facto} paradigm to adapt robot policies as a function of task-specific objectives. Most commonly-used motion policies carry particular structures that are often overlooked in policy optimization algorithms. We instead propose to leverage the structure of probabilistic policies by casting the policy optimization as an optimal transport problem. Specifically, we focus on robot motion policies that build on Gaussian mixture models (GMMs) and formulate the policy optimization as a Wassertein gradient flow over the GMMs space. This naturally allows us to constrain the policy updates via the $L^2$-Wasserstein distance between GMMs to enhance the stability of the policy optimization process. Furthermore, we leverage the geometry of the Bures-Wasserstein manifold to optimize the Gaussian distributions of the GMM policy via Riemannian optimization. We evaluate our approach on common robotic settings: Reaching motions, collision-avoidance behaviors, and multi-goal tasks. Our results show that our method outperforms common policy optimization baselines in terms of task success rate and low-variance solutions.

An accurate environment dynamics model is crucial for various downstream tasks in sequential decision-making, such as counterfactual prediction, off-policy evaluation, and offline reinforcement learning. Currently, these models were learned through empirical risk minimization (ERM) by step-wise fitting of historical transition data. This way was previously believed unreliable over long-horizon rollouts because of the compounding errors, which can lead to uncontrollable inaccuracies in predictions. In this paper, we find that the challenge extends beyond just long-term prediction errors: we reveal that even when planning with one step, learned dynamics models can also perform poorly due to the selection bias of behavior policies during data collection. This issue will significantly mislead the policy optimization process even in identifying single-step optimal actions, further leading to a greater risk in sequential decision-making scenarios. To tackle this problem, we introduce a novel model-learning objective called adversarial weighted empirical risk minimization (AWRM). AWRM incorporates an adversarial policy that exploits the model to generate a data distribution that weakens the model's prediction accuracy, and subsequently, the model is learned under this adversarial data distribution. We implement a practical algorithm, GALILEO, for AWRM and evaluate it on two synthetic tasks, three continuous-control tasks, and \textit{a real-world application}. The experiments demonstrate that GALILEO can accurately predict counterfactual actions and improve various downstream tasks, including offline policy evaluation and improvement, as well as online decision-making.

Recent works in learning-integrated optimization have shown promise in settings where the optimization problem is only partially observed or where general-purpose optimizers perform poorly without expert tuning. By learning an optimizer $\mathbf{g}$ to tackle these challenging problems with $f$ as the objective, the optimization process can be substantially accelerated by leveraging past experience. The optimizer can be trained with supervision from known optimal solutions or implicitly by optimizing the compound function $f\circ \mathbf{g}$. The implicit approach may not require optimal solutions as labels and is capable of handling problem uncertainty; however, it is slow to train and deploy due to frequent calls to optimizer $\mathbf{g}$ during both training and testing. The training is further challenged by sparse gradients of $\mathbf{g}$, especially for combinatorial solvers. To address these challenges, we propose using a smooth and learnable **Landscape Surrogate** $\mathcal{M}$ as a replacement for $f\circ \mathbf{g}$. This surrogate, learnable by neural networks, can be computed faster than the solver $\mathbf{g}$, provides dense and smooth gradients during training, can generalize to unseen optimization problems, and is efficiently learned via alternating optimization. We test our approach on both synthetic problems, including shortest path and multidimensional knapsack, and real-world problems such as portfolio optimization, achieving comparable or superior objective values compared to state-of-the-art baselines while reducing the number of calls to $\mathbf{g}$. Notably, our approach outperforms existing methods for computationally expensive high-dimensional problems.

In reinforcement learning (RL), simplicity is typically quantified on an action-by-action basis -- but this timescale ignores temporal regularities, like repetitions, often present in sequential strategies. We therefore propose an RL algorithm that learns to solve tasks with sequences of actions that are compressible. We explore two possible sources of simple action sequences: Sequences that can be learned by autoregressive models, and sequences that are compressible with off-the-shelf data compression algorithms. Distilling these preferences into sequence priors, we derive a novel information-theoretic objective that incentivizes agents to learn policies that maximize rewards while conforming to these priors. We show that the resulting RL algorithm leads to faster learning, and attains higher returns than state-of-the-art model-free approaches in a series of continuous control tasks from the DeepMind Control Suite. These priors also produce a powerful information-regularized agent that is robust to noisy observations and can perform open-loop control.

The advancement of continual reinforcement learning (RL) has been facing various obstacles, including standardized metrics and evaluation protocols, demanding computational requirements, and a lack of widely accepted standard benchmarks. In response to these challenges, we present COOM ($\textbf{C}$ontinual D$\textbf{OOM}$), a continual RL benchmark tailored for embodied pixel-based RL. COOM presents a meticulously crafted suite of task sequences set within visually distinct 3D environments, serving as a robust evaluation framework to assess crucial aspects of continual RL, such as catastrophic forgetting, knowledge transfer, and sample-efficient learning. Following an in-depth empirical evaluation of popular continual learning (CL) methods, we pinpoint their limitations, provide valuable insight into the benchmark and highlight unique algorithmic challenges. This makes our work the first to benchmark image-based CRL in 3D environments with embodied perception. The primary objective of the COOM benchmark is to offer the research community a valuable and cost-effective challenge. It seeks to deepen our comprehension of the capabilities and limitations of current and forthcoming CL methods in an RL setting. The code and environments are open-sourced and accessible on GitHub.

False arrhythmia alarms in intensive care units (ICUs) are a continuing problem despite considerable effort from industrial and academic algorithm developers. Of all life-threatening arrhythmias, ventricular tachycardia (VT) stands out as the most challenging arrhythmia to detect reliably. We introduce a new annotated VT alarm database, VTaC (Ventricular Tachycardia annotated alarms from ICUs) consisting of over 5,000 waveform recordings with VT alarms triggered by bedside monitors in the ICUs. Each VT alarm in the dataset has been labeled by at least two independent human expert annotators. The dataset encompasses data collected from ICUs in three major US hospitals and includes data from three leading bedside monitor manufacturers, providing a diverse and representative collection of alarm waveform data. Each waveform recording comprises at least two electrocardiogram (ECG) leads and one or more pulsatile waveforms, such as photoplethysmogram (PPG or PLETH) and arterial blood pressure (ABP) waveforms. We demonstrate the utility of this new benchmark dataset for the task of false arrhythmia alarm reduction, and present performance of multiple machine learning approaches, including conventional supervised machine learning, deep learning, contrastive learning and generative approaches for the task of VT false alarm reduction.

Generalizing policies across different domains with dynamics mismatch poses a significant challenge in reinforcement learning. For example, a robot learns the policy in a simulator, but when it is deployed in the real world, the dynamics of the environment may be different. Given the source and target domain with dynamics mismatch, we consider the online dynamics adaptation problem, in which case the agent can access sufficient source domain data while online interactions with the target domain are limited. Existing research has attempted to solve the problem from the dynamics discrepancy perspective. In this work, we reveal the limitations of these methods and explore the problem from the value difference perspective via a novel insight on the value consistency across domains. Specifically, we present the Value-Guided Data Filtering (VGDF) algorithm, which selectively shares transitions from the source domain based on the proximity of paired value targets across the two domains. Empirical results on various environments with kinematic and morphology shifts demonstrate that our method achieves superior performance compared to prior approaches.

Offline reinforcement learning (RL) is suitable for safety-critical domains where online exploration is not feasible. In such domains, decision-making should take into consideration the risk of catastrophic outcomes. In other words, decision-making should be *risk-averse*. An additional challenge of offline RL is avoiding *distributional shift*, i.e. ensuring that state-action pairs visited by the policy remain near those in the dataset. Previous offline RL algorithms that consider risk combine offline RL techniques (to avoid distributional shift), with risk-sensitive RL algorithms (to achieve risk-aversion). In this work, we propose risk-aversion as a mechanism to jointly address *both* of these issues. We propose a model-based approach, and use an ensemble of models to estimate epistemic uncertainty, in addition to aleatoric uncertainty. We train a policy that is risk-averse, and avoids high uncertainty actions. Risk-aversion to epistemic uncertainty prevents distributional shift, as areas not covered by the dataset have high epistemic uncertainty. Risk-aversion to aleatoric uncertainty discourages actions that are risky due to environment stochasticity. Thus, by considering epistemic uncertainty via a model ensemble and introducing risk-aversion, our algorithm (1R2R) avoids distributional shift in addition to achieving risk-aversion to aleatoric risk. Our experiments show that 1R2R achieves strong performance on deterministic benchmarks, and outperforms existing approaches for risk-sensitive objectives in stochastic domains.

Conversational Recommender Systems (CRS) actively elicit user preferences to generate adaptive recommendations. Mainstream reinforcement learning-based CRS solutions heavily rely on handcrafted reward functions, which may not be aligned with user intent in CRS tasks. Therefore, the design of task-specific rewards is critical to facilitate CRS policy learning, which remains largely under-explored in the literature. In this work, we propose a novel approach to address this challenge by learning intrinsic rewards from interactions with users. Specifically, we formulate intrinsic reward learning as a multi-objective bi-level optimization problem. The inner level optimizes the CRS policy augmented by the learned intrinsic rewards, while the outer level drives the intrinsic rewards to optimize two CRS-specific objectives: maximizing the success rate and minimizing the number of turns to reach a successful recommendation}in conversations. To evaluate the effectiveness of our approach, we conduct extensive experiments on three public CRS benchmarks. The results show that our algorithm significantly improves CRS performance by exploiting informative learned intrinsic rewards.

Goal-conditioned reinforcement learning (GCRL) refers to learning general-purpose skills that aim to reach diverse goals. In particular, offline GCRL only requires purely pre-collected datasets to perform training tasks without additional interactions with the environment. Although offline GCRL has become increasingly prevalent and many previous works have demonstrated its empirical success, the theoretical understanding of efficient offline GCRL algorithms is not well established, especially when the state space is huge and the offline dataset only covers the policy we aim to learn. In this paper, we provide a rigorous theoretical analysis of an existing empirically successful offline GCRL algorithm. We prove that under slight modification, this algorithm enjoys an $\tilde{O}(\text{poly}(1/\epsilon))$ sample complexity (where $\epsilon$ is the desired suboptimality of the learned policy) with general function approximation thanks to the property of (semi-)strong convexity of the objective functions. We only require nearly minimal assumptions on the dataset (single-policy concentrability) and the function class (realizability). Moreover, this algorithm consists of two uninterleaved optimization steps, which we refer to as $V$-learning and policy learning, and is computationally stable since it does not involve minimax optimization. We also empirically validate our theory by showing that the modified algorithm outperforms the previous algorithm in various real-world environments. To the best of our knowledge, this is the first algorithm that is both provably efficient with general function approximation and single-policy concentrability, and empirically successful without requiring solving minimax optimization problems.

In safety-critical robotic tasks, potential failures must be reduced, and multiple constraints must be met, such as avoiding collisions, limiting energy consumption, and maintaining balance. Thus, applying safe reinforcement learning (RL) in such robotic tasks requires to handle multiple constraints and use risk-averse constraints rather than risk-neutral constraints. To this end, we propose a trust region-based safe RL algorithm for multiple constraints called a safe distributional actor-critic (SDAC). Our main contributions are as follows: 1) introducing a gradient integration method to manage infeasibility issues in multi-constrained problems, ensuring theoretical convergence, and 2) developing a TD($\lambda$) target distribution to estimate risk-averse constraints with low biases. We evaluate SDAC through extensive experiments involving multi- and single-constrained robotic tasks. While maintaining high scores, SDAC shows 1.93 times fewer steps to satisfy all constraints in multi-constrained tasks and 1.78 times fewer constraint violations in single-constrained tasks compared to safe RL baselines. Code is available at: https://github.com/rllab-snu/Safe-Distributional-Actor-Critic.

Learning from active human involvement enables the human subject to actively intervene and demonstrate to the AI agent during training. The interaction and corrective feedback from human brings safety and AI alignment to the learning process. In this work, we propose a new reward-free active human involvement method called Proxy Value Propagation for policy optimization. Our key insight is that a proxy value function can be designed to express human intents, wherein state- action pairs in the human demonstration are labeled with high values, while those agents’ actions that are intervened receive low values. Through the TD-learning framework, labeled values of demonstrated state-action pairs are further propagated to other unlabeled data generated from agents’ exploration. The proxy value function thus induces a policy that faithfully emulates human behaviors. Human- in-the-loop experiments show the generality and efficiency of our method. With minimal modification to existing reinforcement learning algorithms, our method can learn to solve continuous and discrete control tasks with various human control devices, including the challenging task of driving in Grand Theft Auto V. Demo video and code are available at: https://metadriverse.github.io/pvp.

We introduce IMP-MARL, an open-source suite of multi-agent reinforcement learning (MARL) environments for large-scale Infrastructure Management Planning (IMP), offering a platform for benchmarking the scalability of cooperative MARL methods in real-world engineering applications. In IMP, a multi-component engineering system is subject to a risk of failure due to its components' damage condition. Specifically, each agent plans inspections and repairs for a specific system component, aiming to minimise maintenance costs while cooperating to minimise system failure risk. With IMP-MARL, we release several environments including one related to offshore wind structural systems, in an effort to meet today's needs to improve management strategies to support sustainable and reliable energy systems. Supported by IMP practical engineering environments featuring up to 100 agents, we conduct a benchmark campaign, where the scalability and performance of state-of-the-art cooperative MARL methods are compared against expert-based heuristic policies. The results reveal that centralised training with decentralised execution methods scale better with the number of agents than fully centralised or decentralised RL approaches, while also outperforming expert-based heuristic policies in most IMP environments. Based on our findings, we additionally outline remaining cooperation and scalability challenges that future MARL methods should still address. Through IMP-MARL, we encourage the implementation of new environments and the further development of MARL methods.

In this paper, we present Neural k-Opt (NeuOpt), a novel learning-to-search (L2S) solver for routing problems. It learns to perform flexible k-opt exchanges based on a tailored action factorization method and a customized recurrent dual-stream decoder. As a pioneering work to circumvent the pure feasibility masking scheme and enable the autonomous exploration of both feasible and infeasible regions, we then propose the Guided Infeasible Region Exploration (GIRE) scheme, which supplements the NeuOpt policy network with feasibility-related features and leverages reward shaping to steer reinforcement learning more effectively. Additionally, we equip NeuOpt with Dynamic Data Augmentation (D2A) for more diverse searches during inference. Extensive experiments on the Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) demonstrate that our NeuOpt not only significantly outstrips existing (masking-based) L2S solvers, but also showcases superiority over the learning-to-construct (L2C) and learning-to-predict (L2P) solvers. Notably, we offer fresh perspectives on how neural solvers can handle VRP constraints. Our code is available: https://github.com/yining043/NeuOpt.

Alleviating overestimation bias is a critical challenge for deep reinforcement learning to achieve successful performance on more complex tasks or offline datasets containing out-of-distribution data. In order to overcome overestimation bias, ensemble methods for Q-learning have been investigated to exploit the diversity of multiple Q-functions. Since network initialization has been the predominant approach to promote diversity in Q-functions, heuristically designed diversity injection methods have been studied in the literature. However, previous studies have not attempted to approach guaranteed independence over an ensemble from a theoretical perspective. By introducing a novel regularization loss for Q-ensemble independence based on random matrix theory, we propose spiked Wishart Q-ensemble independence regularization (SPQR) for reinforcement learning. Specifically, we modify the intractable hypothesis testing criterion for the Q-ensemble independence into a tractable KL divergence between the spectral distribution of the Q-ensemble and the target Wigner's semicircle distribution. We implement SPQR in several online and offline ensemble Q-learning algorithms. In the experiments, SPQR outperforms the baseline algorithms in both online and offline RL benchmarks.

Despite the promising performance of deep reinforcement learning (DRL) agents in many challenging scenarios, the black-box nature of these agents greatly limits their applications in critical domains. Prior research has proposed several explanation techniques to understand the deep learning-based policies in RL. Most existing methods explain why an agent takes individual actions rather than pinpointing the critical steps to its final reward. To fill this gap, we propose StateMask, a novel method to identify the states most critical to the agent's final reward. The high-level idea of StateMask is to learn a mask net that blinds a target agent and forces it to take random actions at some steps without compromising the agent's performance. Through careful design, we can theoretically ensure that the masked agent performs similarly to the original agent. We evaluate StateMask in various popular RL environments and show its superiority over existing explainers in explanation fidelity. We also show that StateMask has better utilities, such as launching adversarial attacks and patching policy errors.

Reinforcement Learning (RL) algorithms typically utilize learning and/or planning techniques to derive effective policies. The integration of both approaches has proven to be highly successful in addressing complex sequential decision-making challenges, as evidenced by algorithms such as AlphaZero and MuZero, which consolidate the planning process into a parametric search-policy. AIXI, the most potent theoretical universal agent, leverages planning through comprehensive search as its primary means to find an optimal policy. Here we define an alternative universal agent, which we call Self-AIXI, that on the contrary to AIXI, maximally exploits learning to obtain good policies. It does so by self-predicting its own stream of action data, which is generated, similarly to other TD(0) agents, by taking an action maximization step over the current on-policy (universal mixture-policy) Q-value estimates. We prove that Self-AIXI converges to AIXI, and inherits a series of properties like maximal Legg-Hutter intelligence and the self-optimizing property.

Multi-objective reinforcement learning (MORL) has been proposed to learn control policies over multiple competing objectives with each possible preference over returns. However, current MORL algorithms fail to account for distributional preferences over the multi-variate returns, which are particularly important in real-world scenarios such as autonomous driving. To address this issue, we extend the concept of Pareto-optimality in MORL into distributional Pareto-optimality, which captures the optimality of return distributions, rather than the expectations. Our proposed method, called Distributional Pareto-Optimal Multi-Objective Reinforcement Learning~(DPMORL), is capable of learning distributional Pareto-optimal policies that balance multiple objectives while considering the return uncertainty. We evaluated our method on several benchmark problems and demonstrated its effectiveness in discovering distributional Pareto-optimal policies and satisfying diverse distributional preferences compared to existing MORL methods.

Modern reinforcement learning (RL) has shown empirical success in various real world settings with complex models and large state-action spaces. The existing analytical results, however, typically focus on settings with a small number of state-actions or simple models such as linearly modeled state-action value functions. To derive RL policies that efficiently handle large state-action spaces with more general value functions, some recent works have considered nonlinear function approximation using kernel ridge regression. We propose $\pi$-KRVI, an optimistic modification of least-squares value iteration, when the action-value function is represented by an RKHS. We prove the first order-optimal regret guarantees under a general setting. Our results show a significant polynomial in the number of episodes improvement over the state of the art. In particular, with highly non-smooth kernels (such as Neural Tangent kernel or some Matérn kernels) the existing results lead to trivial (superlinear in the number of episodes) regret bounds. We show a sublinear regret bound that is order optimal in the cases where a lower bound on regret is known (which includes the kernels mentioned above).