NeurIPS 2023 论文列表

Imitation learning uses data for training policies to solve complex tasks. However, when the training data is collected from human demonstrators, it often leads to multimodal distributions because of the variability in human actions. Most imitation learning methods rely on a maximum likelihood (ML) objective to learn a parameterized policy, but this can result in suboptimal or unsafe behavior due to the mode-averaging property of the ML objective. In this work, we propose Information Maximizing Curriculum, a curriculum-based approach that assigns a weight to each data point and encourages the model to specialize in the data it can represent, effectively mitigating the mode-averaging problem by allowing the model to ignore data from modes it cannot represent. To cover all modes and thus, enable versatile behavior, we extend our approach to a mixture of experts (MoE) policy, where each mixture component selects its own subset of the training data for learning. A novel, maximum entropy-based objective is proposed to achieve full coverage of the dataset, thereby enabling the policy to encompass all modes within the data distribution. We demonstrate the effectiveness of our approach on complex simulated control tasks using versatile human demonstrations, achieving superior performance compared to state-of-the-art methods.

Simulation is an essential tool to develop and benchmark autonomous vehicle planning software in a safe and cost-effective manner. However, realistic simulation requires accurate modeling of multi-agent interactive behaviors to be trustworthy, behaviors which can be highly nuanced and complex. To address these challenges, we introduce Waymax, a new data-driven simulator for autonomous driving in multi-agent scenes, designed for large-scale simulation and testing. Waymax uses publicly-released, real-world driving data (e.g., the Waymo Open Motion Dataset) to initialize or play back a diverse set of multi-agent simulated scenarios. It runs entirely on hardware accelerators such as TPUs/GPUs and supports in-graph simulation for training, making it suitable for modern large-scale, distributed machine learning workflows. To support online training and evaluation, Waymax includes several learned and hard-coded behavior models that allow for realistic interaction within simulation. To supplement Waymax, we benchmark a suite of popular imitation and reinforcement learning algorithms with ablation studies on different design decisions, where we highlight the effectiveness of routes as guidance for planning agents and the ability of RL to overfit against simulated agents.

A common setting in multitask reinforcement learning (RL) demands that an agent rapidly adapt to various stationary reward functions randomly sampled from a fixed distribution. In such situations, the successor representation (SR) is a popular framework which supports rapid policy evaluation by decoupling a policy's expected discounted, cumulative state occupancies from a specific reward function. However, in the natural world, sequential tasks are rarely independent, and instead reflect shifting priorities based on the availability and subjective perception of rewarding stimuli. Reflecting this disjunction, in this paper we study the phenomenon of diminishing marginal utility and introduce a novel state representation, the $\lambda$ representation ($\lambda$R) which, surprisingly, is required for policy evaluation in this setting and which generalizes the SR as well as several other state representations from the literature. We establish the $\lambda$R's formal properties and examine its normative advantages in the context of machine learning, as well as its usefulness for studying natural behaviors, particularly foraging.

We first raise and tackle a ``time synchronization'' issue between the agent and the environment in non-stationary reinforcement learning (RL), a crucial factor hindering its real-world applications. In reality, environmental changes occur over wall-clock time ($t$) rather than episode progress ($k$), where wall-clock time signifies the actual elapsed time within the fixed duration $t \in [0, T]$. In existing works, at episode $k$, the agent rolls a trajectory and trains a policy before transitioning to episode $k+1$. In the context of the time-desynchronized environment, however, the agent at time $t_{k}$ allocates $\Delta t$ for trajectory generation and training, subsequently moves to the next episode at $t_{k+1}=t_{k}+\Delta t$. Despite a fixed total number of episodes ($K$), the agent accumulates different trajectories influenced by the choice of interaction times ($t_1,t_2,...,t_K$), significantly impacting the suboptimality gap of the policy. We propose a Proactively Synchronizing Tempo ($\texttt{ProST}$) framework that computes a suboptimal sequence {$t_1,t_2,...,t_K$} (= { $t_{1:K}$}) by minimizing an upper bound on its performance measure, i.e., the dynamic regret. Our main contribution is that we show that a suboptimal {$t_{1:K}$} trades-off between the policy training time (agent tempo) and how fast the environment changes (environment tempo). Theoretically, this work develops a suboptimal {$t_{1:K}$} as a function of the degree of the environment's non-stationarity while also achieving a sublinear dynamic regret. Our experimental evaluation on various high-dimensional non-stationary environments shows that the $\texttt{ProST}$ framework achieves a higher online return at suboptimal {$t_{1:K}$} than the existing methods.

Value decomposition methods have gained popularity in the field of cooperative multi-agent reinforcement learning. However, almost all existing methods follow the principle of Individual Global Max (IGM) or its variants, which limits their problem-solving capabilities. To address this, we propose a dual self-awareness value decomposition framework, inspired by the notion of dual self-awareness in psychology, that entirely rejects the IGM premise. Each agent consists of an ego policy for action selection and an alter ego value function to solve the credit assignment problem. The value function factorization can ignore the IGM assumption by utilizing an explicit search procedure. On the basis of the above, we also suggest a novel anti-ego exploration mechanism to avoid the algorithm becoming stuck in a local optimum. As the first fully IGM-free value decomposition method, our proposed framework achieves desirable performance in various cooperative tasks.

While machine learning can myopically reinforce social inequalities, it may also be used to dynamically seek equitable outcomes. In this paper, we formalize long-term fairness as an online reinforcement learning problem for a policy affecting human populations. This formulation accommodates dynamical control objectives, such as achieving equitable population states, that cannot be incorporated into static formulations of fairness. We demonstrate that algorithmic solutions to the proposed fairness problem can adapt to unknown dynamics and, by sacrificing short-term incentives, drive the policy-population system towards more desirable equilibria. For the proposed setting, we develop an algorithm that adapts recent work in online learning and prove that this algorithm achieves simultaneous probabilistic bounds on cumulative loss and cumulative violations of fairness. In the classification setting subject to group fairness, we compare our proposed algorithm to several baselines, including the repeated retraining of myopic or distributionally robust classifiers, and to a deep reinforcement learning algorithm that lacks fairness guarantees. Our experiments model human populations according to evolutionary game theory and integrate real-world datasets.

The divergence of the Q-value estimation has been a prominent issue offline reinforcement learning (offline RL), where the agent has no access to real dynamics. Traditional beliefs attribute this instability to querying out-of-distribution actions when bootstrapping value targets. Though this issue can be alleviated with policy constraints or conservative Q estimation, a theoretical understanding of the underlying mechanism causing the divergence has been absent. In this work, we aim to thoroughly comprehend this mechanism and attain an improved solution. We first identify a fundamental pattern, \emph{self-excitation}, as the primary cause of Q-value estimation divergence in offline RL. Then, we propose a novel Self-Excite Eigenvalue Measure (SEEM) metric based on Neural Tangent Kernel (NTK) to measure the evolving property of Q-network at training, which provides an intriguing explanation of the emergence of divergence. For the first time, our theory can reliably decide whether the training will diverge at an early stage, and even predict the order of the growth for the estimated Q-value, the model's norm, and the crashing step when an SGD optimizer is used. The experiments demonstrate perfect alignment with this theoretic analysis. Building on our insights, we propose to resolve divergence from a novel perspective, namely improving the model's architecture for better extrapolating behavior. Through extensive empirical studies, we identify LayerNorm as a good solution to effectively avoid divergence without introducing detrimental bias, leading to superior performance. Experimental results prove that it can still work in some most challenging settings, i.e. using only 1$\%$ transitions of the dataset, where all previous methods fail. Moreover, it can be easily plugged into modern offline RL methods and achieve SOTA results on many challenging tasks. We also give unique insights into its effectiveness.

Pandora’s Box is a central problem in decision making under uncertainty that can model various real life scenarios. In this problem we are given n boxes, each with a fixed opening cost, and an unknown value drawn from a known distribution, only revealed if we pay the opening cost. Our goal is to find a strategy for opening boxes to minimize the sum of the value selected and the opening cost paid. In this work we revisit Pandora’s Box when the value distributions are correlated, first studied in [CGT+20]. We show that the optimal algorithm for the independent case, given by Weitzman’s rule, directly works for the correlated case. In fact, our algorithm results in significantly improved approximation guarantees compared to the previous work, while also being substantially simpler. We also show how to implement the rule given only sample access to the correlated distribution of values. Specifically, we find that a number of samples that is polynomial in the number of boxes is sufficient for the algorithm to work.

We present the largest and most comprehensive empirical study of pre-trained visual representations (PVRs) or visual ‘foundation models’ for Embodied AI. First, we curate CortexBench, consisting of 17 different tasks spanning locomotion, navigation, dexterous, and mobile manipulation. Next, we systematically evaluate existing PVRs and find that none are universally dominant. To study the effect of pre-training data size and diversity, we combine over 4,000 hours of egocentric videos from 7 different sources (over 4.3M images) and ImageNet to train different-sized vision transformers using Masked Auto-Encoding (MAE) on slices of this data. Contrary to inferences from prior work, we find that scaling dataset size and diversity does not improve performance universally (but does so on average). Our largest model, named VC-1, outperforms all prior PVRs on average but does not universally dominate either. Next, we show that task- or domain-specific adaptation of VC-1 leads to substantial gains, with VC-1 (adapted) achieving competitive or superior performance than the best known results on all of the benchmarks in CortexBench. Finally, we present real-world hardware experiments, in which VC-1 and VC-1 (adapted) outperform the strongest pre-existing PVR. Overall, this paper presents no new techniques but a rigorous systematic evaluation, a broad set of findings about PVRs (that in some cases, refute those made in narrow domains in prior work), and open-sourced code and models (that required over 10,000 GPU-hours to train) for the benefit of the research community.

The efficient frontier (EF) is a fundamental resource allocation problem where one has to find an optimal portfolio maximizing a reward at a given level of risk. This optimal solution is traditionally found by solving a convex optimization problem. In this paper, we introduce NeuralEF: a fast neural approximation framework that robustly forecasts the result of the EF convex optimizations problems with respect to heterogeneous linear constraints and variable number of optimization inputs. By reformulating an optimization problem as a sequence to sequence problem, we show that NeuralEF is a viable solution to accelerate large-scale simulation while handling discontinuous behavior.

Finetuning language models with reinforcement learning (RL), e.g. from human feedback (HF), is a prominent method for alignment. But optimizing against a reward model can improve on reward while degrading performance in other areas, a phenomenon known as reward hacking, alignment tax, or language drift. First, we argue that commonly-used test metrics are insufficient and instead measure how different algorithms tradeoff between reward and drift. The standard method modified the reward with a Kullback-Lieber (KL) penalty between the online and initial model. We propose Elastic Reset, a new algorithm that achieves higher reward with less drift without explicitly modifying the training objective. We periodically reset the online model to an exponentially moving average (EMA) of itself, then reset the EMA model to the initial model. Through the use of an EMA, our model recovers quickly after resets and achieves higher reward with less drift in the same number of steps. We demonstrate that fine-tuning language models with Elastic Reset leads to state-of-the-art performance on a small scale pivot-translation benchmark, outperforms all baselines in a medium-scale RLHF-like IMDB mock sentiment task and leads to a more performant and more aligned technical QA chatbot with LLaMA-7B. Code available https://github.com/mnoukhov/elastic-reset

A key theme in the past decade has been that when large neural networks and large datasets combine they can produce remarkable results. In deep reinforcement learning (RL), this paradigm is commonly made possible through experience replay, whereby a dataset of past experiences is used to train a policy or value function. However, unlike in supervised or self-supervised learning, an RL agent has to collect its own data, which is often limited. Thus, it is challenging to reap the benefits of deep learning, and even small neural networks can overfit at the start of training. In this work, we leverage the tremendous recent progress in generative modeling and propose Synthetic Experience Replay (SynthER), a diffusion-based approach to flexibly upsample an agent's collected experience. We show that SynthER is an effective method for training RL agents across offline and online settings, in both proprioceptive and pixel-based environments. In offline settings, we observe drastic improvements when upsampling small offline datasets and see that additional synthetic data also allows us to effectively train larger networks. Furthermore, SynthER enables online agents to train with a much higher update-to-data ratio than before, leading to a significant increase in sample efficiency, without any algorithmic changes. We believe that synthetic training data could open the door to realizing the full potential of deep learning for replay-based RL algorithms from limited data. Finally, we open-source our code at https://github.com/conglu1997/SynthER.

We study the problem of agnostic PAC reinforcement learning (RL): given a policy class $\Pi$, how many rounds of interaction with an unknown MDP (with a potentially large state and action space) are required to learn an $\epsilon$-suboptimal policy with respect to \(\Pi\)? Towards that end, we introduce a new complexity measure, called the \emph{spanning capacity}, that depends solely on the set \(\Pi\) and is independent of the MDP dynamics. With a generative model, we show that the spanning capacity characterizes PAC learnability for every policy class $\Pi$. However, for online RL, the situation is more subtle. We show there exists a policy class $\Pi$ with a bounded spanning capacity that requires a superpolynomial number of samples to learn. This reveals a surprising separation for agnostic learnability between generative access and online access models (as well as between deterministic/stochastic MDPs under online access). On the positive side, we identify an additional \emph{sunflower} structure which in conjunction with bounded spanning capacity enables statistically efficient online RL via a new algorithm called POPLER, which takes inspiration from classical importance sampling methods as well as recent developments for reachable-state identification and policy evaluation in reward-free exploration.

In offline imitation learning (IL), an agent aims to learn an optimal expert behavior policy without additional online environment interactions. However, in many real-world scenarios, such as robotics manipulation, the offline dataset is collected from suboptimal behaviors without rewards. Due to the scarce expert data, the agents usually suffer from simply memorizing poor trajectories and are vulnerable to the variations in the environments, lacking the capability of generalizing to new environments.To automatically generate high-quality expert data and improve the generalization ability of the agent, we propose a framework named \underline{O}ffline \underline{I}mitation \underline{L}earning with \underline{C}ounterfactual data \underline{A}ugmentation (OILCA) by doing counterfactual inference. In particular, we leverage identifiable variational autoencoder to generate \textit{counterfactual} samples for expert data augmentation. We theoretically analyze the influence of the generated expert data and the improvement of generalization. Moreover, we conduct extensive experiments to demonstrate that our approach significantly outperforms various baselines on both \textsc{DeepMind Control Suite} benchmark for in-distribution performance and \textsc{CausalWorld} benchmark for out-of-distribution generalization.

We propose a novel model-based offline Reinforcement Learning (RL) framework, called Adversarial Model for Offline Reinforcement Learning (ARMOR), which can robustly learn policies to improve upon an arbitrary reference policy regardless of data coverage. ARMOR is designed to optimize policies for the worst-case performance relative to the reference policy through adversarially training a Markov decision process model. In theory, we prove that ARMOR, with a well-tuned hyperparameter, can compete with the best policy within data coverage when the reference policy is supported by the data. At the same time, ARMOR is robust to hyperparameter choices: the policy learned by ARMOR, with any admissible hyperparameter, would never degrade the performance of the reference policy, even when the reference policy is not covered by the dataset. To validate these properties in practice, we design a scalable implementation of ARMOR, which by adversarial training, can optimize policies without using model ensembles in contrast to typical model-based methods. We show that ARMOR achieves competent performance with both state-of-the-art offline model-free and model-based RL algorithms and can robustly improve the reference policy over various hyperparameter choices.

Actor-critic (AC) methods are widely used in reinforcement learning (RL), and benefit from the flexibility of using any policy gradient method as the actor and value-based method as the critic. The critic is usually trained by minimizing the TD error, an objective that is potentially decorrelated with the true goal of achieving a high reward with the actor. We address this mismatch by designing a joint objective for training the actor and critic in a decision-aware fashion. We use the proposed objective to design a generic, AC algorithm that can easily handle any function approximation. We explicitly characterize the conditions under which the resulting algorithm guarantees monotonic policy improvement, regardless of the choice of the policy and critic parameterization. Instantiating the generic algorithm results in an actor that involves maximizing a sequence of surrogate functions (similar to TRPO, PPO), and a critic that involves minimizing a closely connected objective. Using simple bandit examples, we provably establish the benefit of the proposed critic objective over the standard squared error. Finally, we empirically demonstrate the benefit of our decision-aware actor-critic framework on simple RL problems.

Extracting a stable and compact representation of the environment is crucial for efficient reinforcement learning in high-dimensional, noisy, and non-stationary environments. Different categories of information coexist in such environments -- how to effectively extract and disentangle the information remains a challenging problem. In this paper, we propose IFactor, a general framework to model four distinct categories of latent state variables that capture various aspects of information within the RL system, based on their interactions with actions and rewards. Our analysis establishes block-wise identifiability of these latent variables, which not only provides a stable and compact representation but also discloses that all reward-relevant factors are significant for policy learning. We further present a practical approach to learning the world model with identifiable blocks, ensuring the removal of redundancies but retaining minimal and sufficient information for policy optimization. Experiments in synthetic worlds demonstrate that our method accurately identifies the ground-truth latent variables, substantiating our theoretical findings. Moreover, experiments in variants of the DeepMind Control Suite and RoboDesk showcase the superior performance of our approach over baselines.

Offline multi-agent reinforcement learning is challenging due to the coupling effect of both distribution shift issue common in offline setting and the high dimension issue common in multi-agent setting, making the action out-of-distribution (OOD) and value overestimation phenomenon excessively severe. To mitigate this problem, we propose a novel multi-agent offline RL algorithm, named CounterFactual Conservative Q-Learning (CFCQL) to conduct conservative value estimation. Rather than regarding all the agents as a high dimensional single one and directly applying single agent conservative methods to it, CFCQL calculates conservative regularization for each agent separately in a counterfactual way and then linearly combines them to realize an overall conservative value estimation. We prove that it still enjoys the underestimation property and the performance guarantee as those single agent conservative methods do, but the induced regularization and safe policy improvement bound are independent of the agent number, which is therefore theoretically superior to the direct treatment referred to above, especially when the agent number is large. We further conduct experiments on four environments including both discrete and continuous action settings on both existing and our man-made datasets, demonstrating that CFCQL outperforms existing methods on most datasets and even with a remarkable margin on some of them.