NeurIPS 2023 论文列表

We introduce a principled method for performing zero-shot transfer in reinforcement learning (RL) by exploiting approximate models of the environment. Zero-shot transfer in RL has been investigated by leveraging methods rooted in generalized policy improvement (GPI) and successor features (SFs). Although computationally efficient, these methods are model-free: they analyze a library of policies---each solving a particular task---and identify which action the agent should take. We investigate the more general setting where, in addition to a library of policies, the agent has access to an approximate environment model. Even though model-based RL algorithms can identify near-optimal policies, they are typically computationally intensive. We introduce $h$-GPI, a multi-step extension of GPI that interpolates between these extremes---standard model-free GPI and fully model-based planning---as a function of a parameter, $h$, regulating the amount of time the agent has to reason. We prove that $h$-GPI's performance lower bound is strictly better than GPI's, and show that $h$-GPI generally outperforms GPI as $h$ increases. Furthermore, we prove that as $h$ increases, $h$-GPI's performance becomes arbitrarily less susceptible to sub-optimality in the agent's policy library. Finally, we introduce novel bounds characterizing the gains achievable by $h$-GPI as a function of approximation errors in both the agent's policy library and its (possibly learned) model. These bounds strictly generalize those known in the literature. We evaluate $h$-GPI on challenging tabular and continuous-state problems under value function approximation and show that it consistently outperforms GPI and state-of-the-art competing methods under various levels of approximation errors.

Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data. In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43\% of cases; this statistic is as high as 58\% when compared to Bard and 65\% versus DaVinci003, which was trained with human feedback. Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

We investigate the problem of corruption robustness in offline reinforcement learning (RL) with general function approximation, where an adversary can corrupt each sample in the offline dataset, and the corruption level $\zeta\geq0$ quantifies the cumulative corruption amount over $n$ episodes and $H$ steps. Our goal is to find a policy that is robust to such corruption and minimizes the suboptimality gap with respect to the optimal policy for the uncorrupted Markov decision processes (MDPs). Drawing inspiration from the uncertainty-weighting technique from the robust online RL setting \citep{he2022nearly,ye2022corruptionrobust}, we design a new uncertainty weight iteration procedure to efficiently compute on batched samples and propose a corruption-robust algorithm for offline RL. Notably, under the assumption of single policy coverage and the knowledge of $\zeta$, our proposed algorithm achieves a suboptimality bound that is worsened by an additive factor of $\mathcal O(\zeta \cdot (\text CC(\lambda,\hat{\mathcal F},\mathcal Z_n^H))^{1/2} (C(\hat{\mathcal F},\mu))^{-1/2} n^{-1})$ due to the corruption. Here $\text CC(\lambda,\hat{\mathcal F},\mathcal Z_n^H)$ is the coverage coefficient that depends on the regularization parameter $\lambda$, the confidence set $\hat{\mathcal F}$, and the dataset $\mathcal Z_n^H$, and $C(\hat{\mathcal F},\mu)$ is a coefficient that depends on $\hat{\mathcal F}$ and the underlying data distribution $\mu$. When specialized to linear MDPs, the corruption-dependent error term reduces to $\mathcal O(\zeta d n^{-1})$ with $d$ being the dimension of the feature map, which matches the existing lower bound for corrupted linear MDPs. This suggests that our analysis is tight in terms of the corruption-dependent term.

While Graph Neural Networks (GNNs) recently became powerful tools in graph learning tasks, considerable efforts have been spent on improving GNNs' structural encoding ability. A particular line of work proposed subgraph GNNs that use subgraph information to improve GNNs' expressivity and achieved great success. However, such effectivity sacrifices the efficiency of GNNs by enumerating all possible subgraphs. In this paper, we analyze the necessity of complete subgraph enumeration and show that a model can achieve a comparable level of expressivity by considering a small subset of the subgraphs. We then formulate the identification of the optimal subset as a combinatorial optimization problem and propose Magnetic Graph Neural Network (MAG-GNN), a reinforcement learning (RL) boosted GNN, to solve the problem. Starting with a candidate subgraph set, MAG-GNN employs an RL agent to iteratively update the subgraphs to locate the most expressive set for prediction. This reduces the exponential complexity of subgraph enumeration to the constant complexity of a subgraph search algorithm while keeping good expressivity. We conduct extensive experiments on many datasets, showing that MAG-GNN achieves competitive performance to state-of-the-art methods and even outperforms many subgraph GNNs. We also demonstrate that MAG-GNN effectively reduces the running time of subgraph GNNs.

The options framework yields empirical success in long-horizon planning problems of reinforcement learning (RL). Recent works show that options help improve the sample efficiency in online RL. However, these results are no longer applicable to scenarios where exploring the environment online is risky, e.g., automated driving and healthcare. In this paper, we provide the first analysis of the sample complexity for offline RL with options, where the agent learns from a dataset without further interaction with the environment. We derive a novel information-theoretic lower bound, which generalizes the one for offline learning with actions. We propose the PEssimistic Value Iteration for Learning with Options (PEVIO) algorithm and establish near-optimal suboptimality bounds for two popular data-collection procedures, where the first one collects state-option transitions and the second one collects state-action transitions. We show that compared to offline RL with actions, using options not only enjoys a faster finite-time convergence rate (to the optimal value) but also attains a better performance when either the options are carefully designed or the offline data is limited. Based on these results, we analyze the pros and cons of the data-collection procedures.

A major challenge of reinforcement learning (RL) in real-world applications is the variation between environments, tasks or clients. Meta-RL (MRL) addresses this issue by learning a meta-policy that adapts to new tasks. Standard MRL methods optimize the average return over tasks, but often suffer from poor results in tasks of high risk or difficulty. This limits system reliability since test tasks are not known in advance. In this work, we define a robust MRL objective with a controlled robustness level. Optimization of analogous robust objectives in RL is known to lead to both **biased gradients** and **data inefficiency**. We prove that the gradient bias disappears in our proposed MRL framework. The data inefficiency is addressed via the novel Robust Meta RL algorithm (RoML). RoML is a meta-algorithm that generates a robust version of any given MRL algorithm, by identifying and over-sampling harder tasks throughout training. We demonstrate that RoML achieves robust returns on multiple navigation and continuous control benchmarks.

Extensive experiments have gradually revealed the potential performance bottleneck of modeling Combinatorial Optimization (CO) solving as neural solution prediction tasks. The neural networks, in their pursuit of minimizing the average objective score across the distribution of historical problem instances, diverge from the core target of CO of seeking optimal solutions for every test instance. This calls for an effective search on each problem instance, while the model should serve to provide supporting knowledge that benefits the search. To this end, we propose T2T (Training to Testing) framework that first leverages the generative modeling to estimate the high-quality solution distribution for each instance during training, and then conducts a gradient-based search within the solution space during testing. The proposed neural search paradigm consistently leverages generative modeling, specifically diffusion, for graduated solution improvement. It disrupts the local structure of the given solution by introducing noise and reconstructs a lower-cost solution guided by the optimization objective. Experimental results on Traveling Salesman Problem (TSP) and Maximal Independent Set (MIS) show the significant superiority of T2T, demonstrating an average performance gain of 49.15% for TSP solving and 17.27% for MIS solving compared to the previous state-of-the-art.

Meta-reinforcement learning (meta-RL) techniques have demonstrated remarkable success in generalizing deep reinforcement learning across a range of tasks. Nevertheless, these methods often struggle to generalize beyond tasks with parametric variations. To overcome this challenge, we propose Subtask Decomposition and Virtual Training (SDVT), a novel meta-RL approach that decomposes each non-parametric task into a collection of elementary subtasks and parameterizes the task based on its decomposition. We employ a Gaussian mixture VAE to meta-learn the decomposition process, enabling the agent to reuse policies acquired from common subtasks. Additionally, we propose a virtual training procedure, specifically designed for non-parametric task variability, which generates hypothetical subtask compositions, thereby enhancing generalization to previously unseen subtask compositions. Our method significantly improves performance on the Meta-World ML-10 and ML-45 benchmarks, surpassing current state-of-the-art techniques.

In Reinforcement Learning, agents learn policies by exploring and interacting with the environment. Due to the curse of dimensionality, learning policies that map high-dimensional sensory input to motor output is particularly challenging. During training, state of the art methods (SAC, PPO, etc.) explore the environment by perturbing the actuation with independent Gaussian noise. While this unstructured exploration has proven successful in numerous tasks, it can be suboptimal for overactuated systems. When multiple actuators, such as motors or muscles, drive behavior, uncorrelated perturbations risk diminishing each other's effect, or modifying the behavior in a task-irrelevant way. While solutions to introduce time correlation across action perturbations exist, introducing correlation across actuators has been largely ignored. Here, we propose LATent TIme-Correlated Exploration (Lattice), a method to inject temporally-correlated noise into the latent state of the policy network, which can be seamlessly integrated with on- and off-policy algorithms. We demonstrate that the noisy actions generated by perturbing the network's activations can be modeled as a multivariate Gaussian distribution with a full covariance matrix. In the PyBullet locomotion tasks, Lattice-SAC achieves state of the art results, and reaches 18\% higher reward than unstructured exploration in the Humanoid environment. In the musculoskeletal control environments of MyoSuite, Lattice-PPO achieves higher reward in most reaching and object manipulation tasks, while also finding more energy-efficient policies with reductions of 20-60\%. Overall, we demonstrate the effectiveness of structured action noise in time and actuator space for complex motor control tasks. The code is available at: https://github.com/amathislab/lattice.

We propose a novel formulation for dictionary learning that minimizes the determinant of the dictionary matrix, also known as its volume, subject to the constraint that each row of the sparse coefficient matrix has unit $\ell_1$ norm. The main motivation for the proposed formulation is that it provides global identifiability guarantee of the groundtruth dictionary and sparse coefficient matrices, up to the inherent and inconsequential permutation and scaling ambiguity, if a set of vectors obtained from the coefficient matrix lies inside the $\ell_\infty$ norm ball but contains the $\ell_2$ norm ball in their convex hull. Unlike existing work on identifiability of dictionary learning, our result is global, meaning that a globally optimal solution to our proposed formulation has to be a permuted and rescaled version of the groundtruth factors. Another major improvement in our result is that there is no additional assumption on the dictionary matrix other than it is nonsingular, unlike most other work that require the atoms of the dictionary to be mutually incoherent. We also provide a probabilistic analysis and show that if the sparse coefficient matrix is generated from the widely adopted Bernoulli-Gaussian model, then it is globally identifiable if the sample size is bigger than a constant times $k\log k$, where $k$ is the number atoms in the dictionary, with overwhelming probability. The bound is essentially the same as those local identifiability results, but we show that it is also global. Finally, we propose algorithms to solve the new proposed formulation, specifically one based on the linearized-ADMM with efficient per-iteration updates. The proposed algorithms exhibit surprisingly effective performance in correctly and efficiently recovering the dictionary, as demonstrated in the numerical experiments.

Learning to solve tasks from a sparse reward signal is a major challenge for standard reinforcement learning (RL) algorithms. However, in the real world, agents rarely need to solve sparse reward tasks entirely from scratch. More often, we might possess prior experience to draw on that provides considerable guidance about which actions and outcomes are possible in the world, which we can use to explore more effectively for new tasks. In this work, we study how prior data without reward labels may be used to guide and accelerate exploration for an agent solving a new sparse reward task. We propose a simple approach that learns a reward model from online experience, labels the unlabeled prior data with optimistic rewards, and then uses it concurrently alongside the online data for downstream policy and critic optimization. This general formula leads to rapid exploration in several challenging sparse-reward domains where tabula rasa exploration is insufficient, including the AntMaze domain, Adroit hand manipulation domain, and a visual simulated robotic manipulation domain. Our results highlight the ease of incorporating unlabeled prior data into existing online RL algorithms, and the (perhaps surprising) effectiveness of doing so.

In this paper, we introduce Randomized Q-learning (RandQL), a novel randomized model-free algorithm for regret minimization in episodic Markov Decision Processes (MDPs). To the best of our knowledge, RandQL is the first tractable model-free posterior sampling-based algorithm. We analyze the performance of RandQL in both tabular and non-tabular metric space settings. In tabular MDPs, RandQL achieves a regret bound of order $\widetilde{\mathcal{O}}(\sqrt{H^{5}SAT})$, where $H$ is the planning horizon, $S$ is the number of states, $A$ is the number of actions, and $T$ is the number of episodes. For a metric state-action space, RandQL enjoys a regret bound of order $\widetilde{\mathcal{O}}(H^{5/2} T^{(d_z+1)/(d_z+2)})$, where $d_z$ denotes the zooming dimension. Notably, RandQL achieves optimistic exploration without using bonuses, relying instead on a novel idea of learning rate randomization. Our empirical study shows that RandQL outperforms existing approaches on baseline exploration environments.

A major challenge in reinforcement learning is to develop practical, sample-efficient algorithms for exploration in high-dimensional domains where generalization and function approximation is required. Low-Rank Markov Decision Processes---where transition probabilities admit a low-rank factorization based on an unknown feature embedding---offer a simple, yet expressive framework for RL with function approximation, yet existing algorithms either (1) are computationally intractable, or (2) require restrictive statistical assumptions such as latent variable structure or access to model-based function approximation. In this work, we propose the first provably sample-efficient algorithm for exploration in Low-Rank MDPs that is both computationally efficient and model-free, allowing for general function approximation while requiring no structural assumptions beyond a reachability condition that we show is substantially weaker than that assumed in prior work. Our algorithm, SpanRL, uses the notion of a barycentric spanner for the feature embedding as an efficiently computable basis for exploration, performing efficient spanner computation by interleaving representation learning and policy optimization subroutines. Our analysis---which is appealingly simple and modular---carefully combines several techniques, including a new approach to error-tolerant barycentric spanner computation, and a new analysis of a certain minimax representation learning objective found in prior work.

This paper investigates group distributionally robust optimization (GDRO), with the purpose to learn a model that performs well over $m$ different distributions. First, we formulate GDRO as a stochastic convex-concave saddle-point problem, and demonstrate that stochastic mirror descent (SMD), using $m$ samples in each iteration, achieves an $O(m (\log m)/\epsilon^2)$ sample complexity for finding an $\epsilon$-optimal solution, which matches the $\Omega(m/\epsilon^2)$ lower bound up to a logarithmic factor. Then, we make use of techniques from online learning to reduce the number of samples required in each round from $m$ to $1$, keeping the same sample complexity. Specifically, we cast GDRO as a two-players game where one player simply performs SMD and the other executes an online algorithm for non-oblivious multi-armed bandits. Next, we consider a more practical scenario where the number of samples that can be drawn from each distribution is different, and propose a novel formulation of weighted GDRO, which allows us to derive distribution-dependent convergence rates. Denote by $n_i$ the sample budget for the $i$-th distribution, and assume $n_1 \geq n_2 \geq \cdots \geq n_m$. In the first approach, we incorporate non-uniform sampling into SMD such that the sample budget is satisfied in expectation, and prove that the excess risk of the $i$-th distribution decreases at an $O(\sqrt{n_1 \log m}/n_i)$ rate. In the second approach, we use mini-batches to meet the budget exactly and also reduce the variance in stochastic gradients, and then leverage stochastic mirror-prox algorithm, which can exploit small variances, to optimize a carefully designed weighted GDRO problem. Under appropriate conditions, it attains an $O((\log m)/\sqrt{n_i})$ convergence rate, which almost matches the optimal $O(\sqrt{1/n_i})$ rate of only learning from the $i$-th distribution with $n_i$ samples.

Despite the success of cooperative multi-agent reinforcement learning algorithms, most of them focus on a single team composition, which prevents them from being used in more realistic scenarios where dynamic team composition is possible. While some studies attempt to solve this problem via multi-task learning in a fixed set of team compositions, there is still a risk of overfitting to the training set, which may lead to catastrophic performance when facing dramatically varying team compositions during execution. To address this problem, we propose to use mutual information (MI) as an augmented reward to prevent individual policies from relying too much on team-related information and encourage agents to learn policies that are robust in different team compositions. Optimizing this MI-augmented objective in an off-policy manner can be intractable due to the existence of dynamic marginal distribution. To alleviate this problem, we first propose a multi-agent policy iteration algorithm with a fixed marginal distribution and prove its convergence and optimality. Then, we propose to employ the Blahut–Arimoto algorithm and an imaginary team composition distribution for optimization with approximate marginal distribution as the practical implementation. Empirically, our method demonstrates strong zero-shot generalization to dynamic team compositions in complex cooperative tasks.

Aligning the world model with the environment for the agent’s specific task is crucial in model-based reinforcement learning. While value-equivalent models may achieve better task awareness than maximum-likelihood models, they sacrifice a large amount of semantic information and face implementation issues. To combine the benefits of both types of models, we propose Task-aware Environment Modeling Pipeline with bi-level Optimization (TEMPO), a bi-level model learning framework that introduces an additional level of optimization on top of a maximum-likelihood model by incorporating a meta weighter network that weights each training sample. The meta weighter in the upper level learns to generate novel sample weights by minimizing a proposed task-aware model loss. The model in the lower level focuses on important samples while maintaining rich semantic information in state representations. We evaluate TEMPO on a variety of continuous and discrete control tasks from the DeepMind Control Suite and Atari video games. Our results demonstrate that TEMPO achieves state-of-the-art performance regarding asymptotic performance, training stability, and convergence speed.

Bayesian Optimization (BO) is a powerful method for tackling expensive black-box optimization problems. As a sequential model-based optimization strategy, BO iteratively explores promising solutions until a predetermined budget, either iterations or time, is exhausted. The decision on when to terminate BO significantly influences both the quality of solutions and its computational efficiency. In this paper, we propose a simple, yet theoretically grounded, two-step method for automatically terminating BO. Our core concept is to proactively identify if the search is within a convex region by examining previously observed samples. BO is halted once the local regret within this convex region falls below a predetermined threshold. To enhance numerical stability, we propose an approximation method for calculating the termination indicator by solving a bilevel optimization problem. We conduct extensive empirical studies on diverse benchmark problems, including synthetic functions, reinforcement learning, and hyperparameter optimization. Experimental results demonstrate that our proposed method saves up to $\approx 80\%$ computational budget yet is with an order of magnitude smaller performance degradation, comparing against the other peer methods. In addition, our proposed termination method is robust in terms of the setting of its termination criterion.

Recent work has demonstrated the effectiveness of formulating decision making as supervised learning on offline-collected trajectories. Powerful sequence models, such as GPT or BERT, are often employed to encode the trajectories. However, the benefits of performing sequence modeling on trajectory data remain unclear. In this work, we investigate whether sequence modeling has the ability to condense trajectories into useful representations that enhance policy learning. We adopt a two-stage framework that first leverages sequence models to encode trajectory-level representations, and then learns a goal-conditioned policy employing the encoded representations as its input. This formulation allows us to consider many existing supervised offline RL methods as specific instances of our framework. Within this framework, we introduce Goal-Conditioned Predictive Coding (GCPC), a sequence modeling objective that yields powerful trajectory representations and leads to performant policies. Through extensive empirical evaluations on AntMaze, FrankaKitchen and Locomotion environments, we observe that sequence modeling can have a significant impact on challenging decision making tasks. Furthermore, we demonstrate that GCPC learns a goal-conditioned latent representation encoding the future trajectory, which enables competitive performance on all three benchmarks.

In many real-world scenarios, Reinforcement Learning (RL) algorithms are trained on data with dynamics shift, i.e., with different underlying environment dynamics. A majority of current methods address such issue by training context encoders to identify environment parameters. Data with dynamics shift are separated according to their environment parameters to train the corresponding policy. However, these methods can be sample inefficient as data are used \textit{ad hoc}, and policies trained for one dynamics cannot benefit from data collected in all other environments with different dynamics. In this paper, we find that in many environments with similar structures and different dynamics, optimal policies have similar stationary state distributions. We exploit such property and learn the stationary state distribution from data with dynamics shift for efficient data reuse. Such distribution is used to regularize the policy trained in a new environment, leading to the SRPO (\textbf{S}tate \textbf{R}egularized \textbf{P}olicy \textbf{O}ptimization) algorithm. To conduct theoretical analyses, the intuition of similar environment structures is characterized by the notion of homomorphous MDPs. We then demonstrate a lower-bound performance guarantee on policies regularized by the stationary state distribution. In practice, SRPO can be an add-on module to context-based algorithms in both online and offline RL settings. Experimental results show that SRPO can make several context-based algorithms far more data efficient and significantly improve their overall performance.