NeurIPS 2023 论文列表

This study reveals the inherent tolerance of contrastive learning (CL) towards sampling bias, wherein negative samples may encompass similar semantics (\eg labels). However, existing theories fall short in providing explanations for this phenomenon. We bridge this research gap by analyzing CL through the lens of distributionally robust optimization (DRO), yielding several key insights: (1) CL essentially conducts DRO over the negative sampling distribution, thus enabling robust performance across a variety of potential distributions and demonstrating robustness to sampling bias; (2) The design of the temperature $\tau$ is not merely heuristic but acts as a Lagrange Coefficient, regulating the size of the potential distribution set; (3) A theoretical connection is established between DRO and mutual information, thus presenting fresh evidence for ``InfoNCE as an estimate of MI'' and a new estimation approach for $\phi$-divergence-based generalized mutual information. We also identify CL's potential shortcomings, including over-conservatism and sensitivity to outliers, and introduce a novel Adjusted InfoNCE loss (ADNCE) to mitigate these issues. It refines potential distribution, improving performance and accelerating convergence. Extensive experiments on various domains (image, sentence, and graph) validate the effectiveness of the proposal.

The availability of challenging benchmarks has played a key role in the recent progress of machine learning. In cooperative multi-agent reinforcement learning, the StarCraft Multi-Agent Challenge (SMAC) has become a popular testbed for centralised training with decentralised execution. However, after years of sustained improvement on SMAC, algorithms now achieve near-perfect performance. In this work, we conduct new analysis demonstrating that SMAC lacks the stochasticity and partial observability to require complex *closed-loop* policies. In particular, we show that an *open-loop* policy conditioned only on the timestep can achieve non-trivial win rates for many SMAC scenarios. To address this limitation, we introduce SMACv2, a new version of the benchmark where scenarios are procedurally generated and require agents to generalise to previously unseen settings (from the same distribution) during evaluation. We also introduce the extended partial observability challenge (EPO), which augments SMACv2 to ensure meaningful partial observability. We show that these changes ensure the benchmark requires the use of *closed-loop* policies. We evaluate state-of-the-art algorithms on SMACv2 and show that it presents significant challenges not present in the original benchmark. Our analysis illustrates that SMACv2 addresses the discovered deficiencies of SMAC and can help benchmark the next generation of MARL methods. Videos of training are available on our [website](https://sites.google.com/view/smacv2).

Markov decision processes (MDPs) often suffer from the sensitivity issue under model ambiguity. In recent years, robust MDPs have emerged as an effective framework to overcome this challenge. Distributionally robust MDPs extend the robust MDP framework by incorporating distributional information of the uncertain model parameters to alleviate the conservative nature of robust MDPs. This paper proposes a computationally efficient solution framework for solving distributionally robust MDPs with Wasserstein ambiguity sets. By exploiting the specific problem structure, the proposed framework decomposes the optimization problems associated with distributionally robust Bellman updates into smaller subproblems, which can be solved efficiently. The overall complexity of the proposed algorithm is quasi-linear in both the numbers of states and actions when the distance metric of the Wasserstein distance is chosen to be $L_1$, $L_2$, or $L_{\infty}$ norm, and so the computational cost of distributional robustness is substantially reduced. Our numerical experiments demonstrate that the proposed algorithms outperform other state-of-the-art solution methods.

Learning from label proportions (LLP) is a generalization of supervised learning in which the training data is available as sets or bags of feature-vectors (instances) along with the average instance-label of each bag. The goal is to train a good instance classifier. While most previous works on LLP have focused on training models on such training data, computational learnability of LLP was only recently explored by Saket (2021, 2022) who showed worst case intractability of properly learning linear threshold functions (LTFs) from label proportions. However, their work did not rule out efficient algorithms for this problem for natural distributions. In this work we show that it is indeed possible to efficiently learn LTFs using LTFs when given access to random bags of some label proportion in which feature-vectors are, conditioned on their labels, independently sampled from a Gaussian distribution $N(µ, Σ)$. Our work shows that a certain matrix – formed using covariances of the differences of feature-vectors sampled from the bags with and without replacement – necessarily has its principal component, after a transformation, in the direction of the normal vector of the LTF. Our algorithm estimates the means and covariance matrices using subgaussian concentration bounds which we show can be applied to efficiently sample bags for approximating the normal direction. Using this in conjunction with novel generalization error bounds in the bag setting, we show that a low error hypothesis LTF can be identified. For some special cases of the $N(0, I)$ distribution we provide a simpler mean estimation based algorithm. We include an experimental evaluation of our learning algorithms along with a comparison with those of Saket (2021, 2022) and random LTFs, demonstrating the effectiveness of our techniques.

Recently, neural heuristics based on deep reinforcement learning have exhibited promise in solving multi-objective combinatorial optimization problems (MOCOPs). However, they are still struggling to achieve high learning efficiency and solution quality. To tackle this issue, we propose an efficient meta neural heuristic (EMNH), in which a meta-model is first trained and then fine-tuned with a few steps to solve corresponding single-objective subproblems. Specifically, for the training process, a (partial) architecture-shared multi-task model is leveraged to achieve parallel learning for the meta-model, so as to speed up the training; meanwhile, a scaled symmetric sampling method with respect to the weight vectors is designed to stabilize the training. For the fine-tuning process, an efficient hierarchical method is proposed to systematically tackle all the subproblems. Experimental results on the multi-objective traveling salesman problem (MOTSP), multi-objective capacitated vehicle routing problem (MOCVRP), and multi-objective knapsack problem (MOKP) show that, EMNH is able to outperform the state-of-the-art neural heuristics in terms of solution quality and learning efficiency, and yield competitive solutions to the strong traditional heuristics while consuming much shorter time.

We explore the methodology and theory of reward-directed generation via conditional diffusion models. Directed generation aims to generate samples with desired properties as measured by a reward function, which has broad applications in generative AI, reinforcement learning, and computational biology. We consider the common learning scenario where the dataset consists of majorly unlabeled data and a small set of data with noisy reward labels. Our approach leverages a learned reward function on the smaller data set as a pseudolabeler to label the unlabelled data. After pseudo-labelling, a conditional diffusion model (CDM) is trained on the data and samples are generated by setting a target value $a$ as the condition in CDM. From a theoretical standpoint, we show that this directed generator can effectively learn and sample from the reward-conditioned data distribution: 1. our model is capable of recovering the data's latent subspace representation. 2. the model generates samples moving closer to the user-specified target. The improvement in rewards of samples is influenced by a interplay between the strength of the reward signal, the distribution shift, and the cost of off-support extrapolation. We provide empirical results to validate our theory and highlight the relationship between the strength of extrapolation and the quality of generated samples.

Reward-free data is abundant and contains rich prior knowledge of human behaviors, but it is not well exploited by offline reinforcement learning (RL) algorithms. In this paper, we propose UBER, an unsupervised approach to extract useful behaviors from offline reward-free datasets via diversified rewards. UBER assigns different pseudo-rewards sampled from a given prior distribution to different agents to extract a diverse set of behaviors, and reuse them as candidate policies to facilitate the learning of new tasks. Perhaps surprisingly, we show that rewards generated from random neural networks are sufficient to extract diverse and useful behaviors, some even close to expert ones. We provide both empirical and theoretical evidences to justify the use of random priors for the reward function. Experiments on multiple benchmarks showcase UBER's ability to learn effective and diverse behavior sets that enhance sample efficiency for online RL, outperforming existing baselines. By reducing reliance on human supervision, UBER broadens the applicability of RL to real-world scenarios with abundant reward-free data.

Inverse Constraint Reinforcement Learning (ICRL) aims to recover the underlying constraints respected by expert agents in a data-driven manner. Existing ICRL algorithms typically assume that the demonstration data is generated by a single type of expert. However, in practice, demonstrations often comprise a mixture of trajectories collected from various expert agents respecting different constraints, making it challenging to explain expert behaviors with a unified constraint function. To tackle this issue, we propose a Multi-Modal Inverse Constrained Reinforcement Learning (MMICRL) algorithm for simultaneously estimating multiple constraints corresponding to different types of experts. MMICRL constructs a flow-based density estimator that enables unsupervised expert identification from demonstrations, so as to infer the agent-specific constraints. Following these constraints, MMICRL imitates expert policies with a novel multi-modal constrained policy optimization objective that minimizes the agent-conditioned policy entropy and maximizes the unconditioned one. To enhance robustness, we incorporate this objective into the contrastive learning framework. This approach enables imitation policies to capture the diversity of behaviors among expert agents. Extensive experiments in both discrete and continuous environments show that MMICRL outperforms other baselines in terms of constraint recovery and control performance.

Machine learning based decision-support tools in criminal justice systems are subjects of intense discussions and academic research. There are important open questions about the utility and fairness of such tools. Academic researchers often rely on a few small datasets that are not sufficient to empirically study various real-world aspects of these questions. In this paper, we contribute WCLD, a curated large dataset of 1.5 million criminal cases from circuit courts in the U.S. state of Wisconsin. We used reliable public data from 1970 to 2020 to curate attributes like prior criminal counts and recidivism outcomes. The dataset contains large number of samples from five racial groups, in addition to information like sex and age (at judgment and first offense). Other attributes in this dataset include neighborhood characteristics obtained from census data, detailed types of offense, charge severity, case decisions, sentence lengths, year of filing etc. We also provide pseudo-identifiers for judge, county and zipcode. The dataset will not only enable researchers to more rigorously study algorithmic fairness in the context of criminal justice, but also relate algorithmic challenges with various systemic issues. We also discuss in detail the process of constructing the dataset and provide a datasheet. The WCLD dataset is available at https://clezdata.github.io/wcld/.

Large language models (LLMs) such as ChatGPT have seen widespread adoption due to their ability to follow user instructions well. Developing these LLMs involves a complex yet poorly understood workflow requiring training with human feedback. Replicating and understanding this instruction-following process faces three major challenges: the high cost of data collection, the lack of trustworthy evaluation, and the absence of reference method implementations. We address these bottlenecks with AlpacaFarm, a simulator that enables research and development for learning from feedback at a low cost. First, we design LLM based simulator for human feedback that is 45x cheaper than crowdworkers and displays high agreement with humans. Second, we identify an evaluation dataset representative of real-world instructions and propose an automatic evaluation procedure. Third, we contribute reference implementations for several methods (PPO, best-of-n, expert iteration, among others) that learn from pairwise feedback. Finally, as an end-to-end validation of AlpacaFarm, we train and evaluate eleven models on 10k pairs of human feedback and show that rankings of models trained in AlpacaFarm match rankings of models trained on human data. As a demonstration of the research possible in AlpacaFarm, we find that methods that use a reward model can substantially improve over supervised fine-tuning and that our reference PPO implementation leads to a +10% win-rate improvement against Davinci003.

In this paper we deal with the state distributional shift problem commonly encountered in offline reinforcement learning during test, where the agent tends to take unreliable actions at out-of-sample (unseen) states. Our idea is to encourage the agent to follow the so called state recovery principle when taking actions, i.e., besides long-term return, the immediate consequences of the current action should also be taken into account and those capable of recovering the state distribution of the behavior policy are preferred. For this purpose, an inverse dynamics model is learned and employed to guide the state recovery behavior of the new policy. Theoretically, we show that the proposed method helps aligning the transited state distribution of the new policy with the offline dataset at out-of-sample states, without the need of explicitly predicting the transited state distribution, which is usually difficult in high-dimensional and complicated environments. The effectiveness and feasibility of the proposed method is demonstrated with the state-of-the-art performance on the general offline RL benchmarks.

We propose A-Crab (Actor-Critic Regularized by Average Bellman error), a new practical algorithm for offline reinforcement learning (RL) in complex environments with insufficient data coverage. Our algorithm combines the marginalized importance sampling framework with the actor-critic paradigm, where the critic returns evaluations of the actor (policy) that are pessimistic relative to the offline data and have a small average (importance-weighted) Bellman error. Compared to existing methods, our algorithm simultaneously offers a number of advantages: (1) It achieves the optimal statistical rate of $1/\sqrt{N}$---where $N$ is the size of offline dataset---in converging to the best policy covered in the offline dataset, even when combined with general function approximators. (2) It relies on a weaker \textit{average} notion of policy coverage (compared to the $\ell_\infty$ single-policy concentrability) that exploits the structure of policy visitations. (3) It outperforms the data-collection behavior policy over a wide range of specific hyperparameters. We provide both theoretical analysis and experimental results to validate the effectiveness of our proposed algorithm. The code is available at https://github.com/zhuhl98/ACrab.

Certifiable, adaptive uncertainty estimates for unknown quantities are an essential ingredient of sequential decision-making algorithms. Standard approaches rely on problem-dependent concentration results and are limited to a specific combination of parameterization, noise family, and estimator. In this paper, we revisit the likelihood-based inference principle and propose to use \emph{likelihood ratios} to construct \emph{any-time valid} confidence sequences without requiring specialized treatment in each application scenario. Our method is especially suitable for problems with well-specified likelihoods, and the resulting sets always maintain the prescribed coverage in a model-agnostic manner. The size of the sets depends on a choice of estimator sequence in the likelihood ratio. We discuss how to provably choose the best sequence of estimators and shed light on connections to online convex optimization with algorithms such as Follow-the-Regularized-Leader. To counteract the initially large bias of the estimators, we propose a reweighting scheme that also opens up deployment in non-parametric settings such as RKHS function classes. We provide a \emph{non-asymptotic} analysis of the likelihood ratio confidence sets size for generalized linear models, using insights from convex duality and online learning. We showcase the practical strength of our method on generalized linear bandit problems, survival analysis, and bandits with various additive noise distributions.

Structured state space sequence (S4) models have recently achieved state-of-the-art performance on long-range sequence modeling tasks. These models also have fast inference speeds and parallelisable training, making them potentially useful in many reinforcement learning settings. We propose a modification to a variant of S4 that enables us to initialise and reset the hidden state in parallel, allowing us to tackle reinforcement learning tasks. We show that our modified architecture runs asymptotically faster than Transformers in sequence length and performs better than RNN's on a simple memory-based task. We evaluate our modified architecture on a set of partially-observable environments and find that, in practice, our model outperforms RNN's while also running over five times faster. Then, by leveraging the model’s ability to handle long-range sequences, we achieve strong performance on a challenging meta-learning task in which the agent is given a randomly-sampled continuous control environment, combined with a randomly-sampled linear projection of the environment's observations and actions. Furthermore, we show the resulting model can adapt to out-of-distribution held-out tasks. Overall, the results presented in this paper show that structured state space models are fast and performant for in-context reinforcement learning tasks. We provide code at https://github.com/luchris429/s5rl.

Multi-objective optimization (MOO) has become an influential framework in many machine learning problems with multiple objectives such as learning with multiple criteria and multi-task learning (MTL). In this paper, we propose a new direction-oriented multi-objective formulation by regularizing the common descent direction within a neighborhood of a direction that optimizes a linear combination of objectives such as the average loss in MTL or a weighted loss that places higher emphasis on some tasks than the others. This formulation includes GD and MGDA as special cases, enjoys the direction-oriented benefit as in CAGrad, and facilitates the design of stochastic algorithms. To solve this problem, we propose Stochastic Direction-oriented Multi-objective Gradient descent (SDMGrad) with simple SGD type of updates, and its variant SDMGrad-OS with an efficient objective sampling. We develop a comprehensive convergence analysis for the proposed methods with different loop sizes and regularization coefficients. We show that both SDMGrad and SDMGrad-OS achieve improved sample complexities to find an $\epsilon$-accurate Pareto stationary point while achieving a small $\epsilon$-level distance toward a conflict-avoidant (CA) direction. For a constant-level CA distance, their sample complexities match the best known $\mathcal{O}(\epsilon^{-2})$ without bounded function value assumption. Extensive experiments show that our methods achieve competitive or improved performance compared to existing gradient manipulation approaches in a series of tasks on multi-task supervised learning and reinforcement learning. Code is available at https://github.com/ml-opt-lab/sdmgrad.

An agent learning an option in hierarchical reinforcement learning must solve three problems: identify the option's subgoal (termination condition), learn a policy, and learn where that policy will succeed (initiation set). The termination condition is typically identified first, but the option policy and initiation set must be learned simultaneously, which is challenging because the initiation set depends on the option policy, which changes as the agent learns. Consequently, data obtained from option execution becomes invalid over time, leading to an inaccurate initiation set that subsequently harms downstream task performance. We highlight three issues---data non-stationarity, temporal credit assignment, and pessimism---specific to learning initiation sets, and propose to address them using tools from off-policy value estimation and classification. We show that our method learns higher-quality initiation sets faster than existing methods (in MiniGrid and Montezuma's Revenge), can automatically discover promising grasps for robot manipulation (in Robosuite), and improves the performance of a state-of-the-art option discovery method in a challenging maze navigation task in MuJoCo.

Much of the previous work towards digital agents for graphical user interfaces (GUIs) has relied on text-based representations (derived from HTML or other structured data sources), which are not always readily available. These input representations have been often coupled with custom, task-specific action spaces. This paper focuses on creating agents that interact with the digital world using the same conceptual interface that humans commonly use — via pixel-based screenshots and a generic action space corresponding to keyboard and mouse actions. Building upon recent progress in pixel-based pretraining, we show, for the first time, that it is possible for such agents to outperform human crowdworkers on the MiniWob++ benchmark of GUI-based instruction following tasks.

We study zero-shot generalization in reinforcement learning - optimizing a policy on a set of training tasks to perform well on a similar but unseen test task. To mitigate overfitting, previous work explored different notions of invariance to the task. However, on problems such as the ProcGen Maze, an adequate solution that is invariant to the task visualization does not exist, and therefore invariance-based approaches fail. Our insight is that learning a policy that effectively $\textit{explores}$ the domain is harder to memorize than a policy that maximizes reward for a specific task, and therefore we expect such learned behavior to generalize well; we indeed demonstrate this empirically on several domains that are difficult for invariance-based approaches. Our $\textit{Explore to Generalize}$ algorithm (ExpGen) builds on this insight: we train an additional ensemble of agents that optimize reward. At test time, either the ensemble agrees on an action, and we generalize well, or we take exploratory actions, which generalize well and drive us to a novel part of the state space, where the ensemble may potentially agree again. We show that our approach is the state-of-the-art on tasks of the ProcGen challenge that have thus far eluded effective generalization, yielding a success rate of 83% on the Maze task and 74% on Heist with $200$ training levels. ExpGen can also be combined with an invariance based approach to gain the best of both worlds, setting new state-of-the-art results on ProcGen. Code available at [https://github.com/EvZissel/expgen](https://github.com/EvZissel/expgen).

The major challenge of offline RL is the distribution shift that appears when out-of-distribution actions are queried, which makes the policy improvement direction biased by extrapolation errors. Most existing methods address this problem by penalizing the policy or value for deviating from the behavior policy during policy improvement or evaluation. In this work, we propose a novel MISA framework to approach offline RL from the perspective of Mutual Information between States and Actions in the dataset by directly constraining the policy improvement direction. MISA constructs lower bounds of mutual information parameterized by the policy and Q-values. We show that optimizing this lower bound is equivalent to maximizing the likelihood of a one-step improved policy on the offline dataset. Hence, we constrain the policy improvement direction to lie in the data manifold. The resulting algorithm simultaneously augments the policy evaluation and improvement by adding mutual information regularizations. MISA is a general framework that unifies conservative Q-learning (CQL) and behavior regularization methods (e.g., TD3+BC) as special cases. We introduce 3 different variants of MISA, and empirically demonstrate that tighter mutual information lower bound gives better offline RL performance. In addition, our extensive experiments show MISA significantly outperforms a wide range of baselines on various tasks of the D4RL benchmark, e.g., achieving 742.9 total points on gym-locomotion tasks. Our code is attached and will be released upon publication.