ICML 2024 论文列表

While machine learning shows promise in automated knowledge generation, current techniques such as large language models and micro-targeted influence operations can be exploited for harmful purposes like the proliferation of disinformation. The European Union's Digital Services Act (DSA) is an exemplary policy response addressing these harms generated by online platforms. In this regard, it necessitates a comprehensive evaluation of its impact on curbing the harmful downstream effects of these opaque practices. Despite their harmful applications, we argue that machine learning techniques offer immense, yet under-exploited, potential for unraveling the impacts of regulations like the DSA. Following an analysis that reveals possible limitations in the DSA's provisions, we call for resolute efforts to address methodological barriers around appropriate data access, isolating marginal regulatory effects, and facilitating generalization across different contexts. Given the identified advantages of data-driven approaches to regulatory delivery, we advocate for machine learning research to help quantify the policy impacts on online harms.

In reinforcement learning, it is crucial to have an accurate environment dynamics model to evaluate different policies' value in downstream tasks like offline policy optimization and policy evaluation. However, the learned model is known to be inaccurate in predictions when evaluating target policies different from data-collection policies. In this work, we found that utilizing policy representation for model learning, called policy-conditioned model (PCM) learning, is useful to mitigate the problem, especially when the offline dataset is collected from diversified behavior policies. The reason beyond that is in this case, PCM becomes a meta-dynamics model that is trained to be aware of and focus on the evaluation policies that on-the-fly adjust the model to be suitable to the evaluation policies’ state-action distribution, thus improving the prediction accuracy. Based on that intuition, we propose an easy-to-implement yet effective algorithm of PCM for accurate model learning. We also give a theoretical analysis and experimental evidence to demonstrate the feasibility of reducing value gaps by adapting the dynamics model under different policies. Experiment results show that PCM outperforms the existing SOTA off-policy evaluation methods in the DOPE benchmark by a large margin, and derives significantly better policies in offline policy selection and model predictive control compared with the standard model learning method.

Learning a high-quality transition model is of great importance for sequential decision-making tasks, especially in offline settings. Nevertheless, the complex behaviors of transition dynamics in real-world environments pose challenges for the standard forward models because of their inductive bias towards smooth regressors, conflicting with the inherent nature of transitions such as discontinuity or large curvature. In this work, we propose to model the transition probability implicitly through a scalar-value energy function, which enables not only flexible distribution prediction but also capturing complex transition behaviors. The Energy-based Transition Models (ETM) are shown to accurately fit the discontinuous transition functions and better generalize to out-of-distribution transition data. Furthermore, we demonstrate that energy-based transition models improve the evaluation accuracy and significantly outperform other off-policy evaluation methods in DOPE benchmark. Finally, we show that energy-based transition models also benefit reinforcement learning and outperform prior offline RL algorithms in D4RL Gym-Mujoco tasks.

Learning from preference labels plays a crucial role in fine-tuning large language models --- this is done via supervised learning, on-policy reinforcement learning (RL), or contrastive learning. Different methods come with different implementation tradeoffs, and existing empirical findings present different conclusions, for instance, some results show that online RL is quite important to attain good fine-tuning results, while others find offline methods sufficient. This raises a question: **what kind of approaches are important for fine-tuning with preference data and why?** In this paper, we answer this question by performing a rigorous analysis of a number of fine-tuning techniques on didactic and full-scale LLM problems. Our main finding is that approaches that use on-policy sampling and attempt to push down the likelihood on certain responses (i.e., employ a ''negative gradient'') outperform offline and maximum likelihood objectives. We conceptualize our insights and unify methods that use on-policy sampling or negative gradient under a notion of mode-seeking objectives for categorical distributions. Mode-seeking objectives are able to alter probability mass on specific bins of a categorical distribution at a fast rate compared to maximum likelihood, allowing them to relocate masses across bins more effectively. Our analysis prescribes actionable insights for preference fine-tuning of LLMs and informs how data should be collected for maximal improvement.

We consider an average reward reinforcement learning (RL) problem and work with asymptotic variance as a risk measure to model safety-critical applications. We design a temporal-difference (TD) type algorithm tailored for policy evaluation in this context. Our algorithm is based on linear stochastic approximation of an equivalent formulation of the asymptotic variance in terms of the solution of the Poisson equation. We consider both the tabular and linear function approximation settings, and establish $\tilde {O}(1/k)$ finite time convergence rate, where $k$ is the number of steps of the algorithm. Our work paves the way for developing actor-critic style algorithms for variance-constrained RL. To the best of our knowledge, our result provides the first sequential estimator for asymptotic variance of a Markov chain with provable finite sample guarantees, which is of independent interest.

Various acceleration approaches for Policy Gradient (PG) have been analyzed within the realm of Reinforcement Learning (RL). However, the theoretical understanding of the widely used momentum-based acceleration method on PG remains largely open. In response to this gap, we adapt the celebrated Nesterov's accelerated gradient (NAG) method to policy optimization in RL, termed *Accelerated Policy Gradient* (APG). To demonstrate the potential of APG in achieving fast convergence, we formally prove that with the true gradient and under the softmax policy parametrization, APG converges to an optimal policy at rates: (i) $\tilde{O}(1/t^2)$ with nearly constant step sizes; (ii) $O(e^{-ct})$ with time-varying step sizes. To the best of our knowledge, this is the first characterization of the convergence rates of NAG in the context of RL. Notably, our analysis relies on one interesting finding: Regardless of the parameter initialization, APG ends up entering a locally nearly-concave regime, where APG can significantly benefit from the momentum, within finite iterations. Through numerical validation and experiments on the Atari 2600 benchmarks, we confirm that APG exhibits a $\tilde{O}(1/t^2)$ rate with nearly constant step sizes and a linear convergence rate with time-varying step sizes, significantly improving convergence over the standard PG.

Optimization problems are pervasive in sectors from manufacturing and distribution to healthcare. However, most such problems are still solved heuristically by hand rather than optimally by state-of-the-art solvers because the expertise required to formulate and solve these problems limits the widespread adoption of optimization tools and techniques. This paper introduces OptiMUS, a Large Language Model (LLM)-based agent designed to formulate and solve (mixed integer) linear programming problems from their natural language descriptions. OptiMUS can develop mathematical models, write and debug solver code, evaluate the generated solutions, and improve its model and code based on these evaluations. OptiMUS utilizes a modular structure to process problems, allowing it to handle problems with long descriptions and complex data without long prompts. Experiments demonstrate that OptiMUS outperforms existing state-of-the-art methods on easy datasets by more than $20$% and on hard datasets (including a new dataset, NLP4LP, released with this paper that features long and complex problems) by more than $30$%. The implementation and the datasets are available at https://github.com/teshnizi/OptiMUS.

Interpretable policy learning seeks to estimate intelligible decision policies from observed actions; however, existing models force a tradeoff between accuracy and interpretability, limiting data-driven interpretations of human decision-making processes. Fundamentally, existing approaches are burdened by this tradeoff because they represent the underlying decision process as a universal policy, when in fact human decisions are dynamic and can change drastically under different contexts. Thus, we develop Contextualized Policy Recovery (CPR), which re-frames the problem of modeling complex decision processes as a multi-task learning problem, where each context poses a unique task and complex decision policies can be constructed piece-wise from many simple context-specific policies. CPR models each context-specific policy as a linear map, and generates new policy models _on-demand_ as contexts are updated with new observations. We provide two flavors of the CPR framework: one focusing on exact local interpretability, and one retaining full global interpretability. We assess CPR through studies on simulated and real data, achieving state-of-the-art performance on predicting antibiotic prescription in intensive care units ($+22$% AUROC vs. previous SOTA) and predicting MRI prescription for Alzheimer's patients ($+7.7$% AUROC vs. previous SOTA). With this improvement, CPR closes the accuracy gap between interpretable and black-box methods, allowing high-resolution exploration and analysis of context-specific decision models.

In modern machine learning, models can often fit training data in numerous ways, some of which perform well on unseen (test) data, while others do not. Remarkably, in such cases gradient descent frequently exhibits an implicit bias that leads to excellent performance on unseen data. This implicit bias was extensively studied in supervised learning, but is far less understood in optimal control (reinforcement learning). There, learning a controller applied to a system via gradient descent is known as policy gradient, and a question of prime importance is the extent to which a learned controller extrapolates to unseen initial states. This paper theoretically studies the implicit bias of policy gradient in terms of extrapolation to unseen initial states. Focusing on the fundamental Linear Quadratic Regulator (LQR) problem, we establish that the extent of extrapolation depends on the degree of exploration induced by the system when commencing from initial states included in training. Experiments corroborate our theory, and demonstrate its conclusions on problems beyond LQR, where systems are non-linear and controllers are neural networks. We hypothesize that real-world optimal control may be greatly improved by developing methods for informed selection of initial states to train on.

A natural approach for reinforcement learning is to predict future rewards by unrolling a neural network world model, and to backpropagate through the resulting computational graph to learn a control policy. However, this method often becomes impractical for long horizons, since typical world models induce hard-to-optimize loss landscapes. Transformers are known to efficiently propagate gradients over long horizons: could they be the solution to this problem? Surprisingly, we show that commonly-used transformer world models produce circuitous gradient paths, which can be detrimental to long-range policy gradients. To tackle this challenge, we propose a class of world models called Action-conditioned World Models (AWMs), designed to provide more direct routes for gradient propagation. We integrate such AWMs into a policy gradient framework that underscores the relationship between network architectures and the policy gradient updates they inherently represent. We demonstrate that AWMs can generate optimization landscapes that are easier to navigate even when compared to those from the simulator itself. This property allows transformer AWMs to produce better policies than competitive baselines in realistic long-horizon tasks.

Reasoning from sequences of raw sensory data is a ubiquitous problem across fields ranging from medical devices to robotics. These problems often involve using long sequences of raw sensor data (e.g. magnetometers, piezoresistors) to predict sequences of desirable physical quantities (e.g. force, inertial measurements). While classical approaches are powerful for locally-linear prediction problems, they often fall short when using real-world sensors. These sensors are typically non-linear, are affected by extraneous variables (e.g. vibration), and exhibit data-dependent drift. For many problems, the prediction task is exacerbated by small labeled datasets since obtaining ground-truth labels requires expensive equipment. In this work, we present Hierarchical State-Space models (HiSS), a conceptually simple, new technique for continuous sequential prediction. HiSS stacks structured state-space models on top of each other to create a temporal hierarchy. Across six real-world sensor datasets, from tactile-based state prediction to accelerometer-based inertial measurement, HiSS outperforms state-of-the-art sequence models such as causal Transformers, LSTMs, S4, and Mamba by at least 23% on MSE. Our experiments further indicate that HiSS demonstrates efficient scaling to smaller datasets and is compatible with existing data-filtering techniques. Code, datasets and videos can be found on https://hiss-csp.github.io.

In recent years, trust region on-policy reinforcement learning has achieved impressive results in addressing complex control tasks and gaming scenarios. However, contemporary state-of-the-art algorithms within this category primarily emphasize improvement in expected performance, lacking the ability to control over the worst-case performance outcomes. To address this limitation, we introduce a novel objective function, optimizing which leads to guaranteed monotonic improvement in the lower probability bound of performance with high confidence. Building upon this groundbreaking theoretical advancement, we further introduce a practical solution called Absolute Policy Optimization (APO). Our experiments demonstrate the effectiveness of our approach across challenging continuous control benchmark tasks and extend its applicability to mastering Atari games. Our findings reveal that APO as well as its efficient variation Proximal Absolute Policy Optimization (PAPO) significantly outperforms state-of-the-art policy gradient algorithms, resulting in substantial improvements in worst-case performance, as well as expected performance.

Since the objective functions of reinforcement learning problems are typically highly nonconvex, it is desirable that policy gradient, the most popular algorithm, escapes saddle points and arrives at second-order stationary points. Existing results only consider vanilla policy gradient algorithms with unbiased gradient estimators, but practical implementations under the infinite-horizon discounted reward setting are biased due to finite-horizon sampling. Moreover, actor-critic methods, whose second-order convergence has not yet been established, are also biased due to the critic approximation of the value function. We provide a novel second-order analysis of biased policy gradient methods, including the vanilla gradient estimator computed from Monte-Carlo sampling of trajectories as well as the double-loop actor-critic algorithm, where in the inner loop the critic improves the approximation of the value function via TD(0) learning. Separately, we also establish the convergence of TD(0) on Markov chains irrespective of initial state distribution.

Training reinforcement learning policies using environment interaction data collected from varying policies or dynamics presents a fundamental challenge. Existing works often overlook the distribution discrepancies induced by policy or dynamics shifts, or rely on specialized algorithms with task priors, thus often resulting in suboptimal policy performances and high learning variances. In this paper, we identify a unified strategy for online RL policy learning under diverse settings of policy and dynamics shifts: transition occupancy matching. In light of this, we introduce a surrogate policy learning objective by considering the transition occupancy discrepancies and then cast it into a tractable min-max optimization problem through dual reformulation. Our method, dubbed Occupancy-Matching Policy Optimization (OMPO), features a specialized actor-critic structure equipped with a distribution discriminator and a small-size local buffer. We conduct extensive experiments based on the OpenAI Gym, Meta-World, and Panda Robots environments, encompassing policy shifts under stationary and non-stationary dynamics, as well as domain adaption. The results demonstrate that OMPO outperforms the specialized baselines from different categories in all settings. We also find that OMPO exhibits particularly strong performance when combined with domain randomization, highlighting its potential in RL-based robotics applications.

We prove that the combination of a target network and over-parameterized linear function approximation establishes a weaker convergence condition for bootstrapped value estimation in certain cases, even with off-policy data. Our condition is naturally satisfied for expected updates over the entire state-action space or learning with a batch of complete trajectories from episodic Markov decision processes. Notably, using only a target network or an over-parameterized model does not provide such a convergence guarantee. Additionally, we extend our results to learning with truncated trajectories, showing that convergence is achievable for all tasks with minor modifications, akin to value truncation for the final states in trajectories. Our primary result focuses on temporal difference estimation for prediction, providing high-probability value estimation error bounds and empirical analysis on Baird's counterexample and a Four-room task. Furthermore, we explore the control setting, demonstrating that similar convergence conditions apply to Q-learning.

In this work, we focus on safe policy improvement in multi-agent domains where current state-of-the-art methods cannot be effectively applied because of large state and action spaces. We consider recent results using Monte Carlo Tree Search for Safe Policy Improvement with Baseline Bootstrapping and propose a novel algorithm that scales this approach to multi-agent domains, exploiting the factorization of the transition model and value function. Given a centralized behavior policy and a dataset of trajectories, our algorithm generates an improved policy by selecting joint actions using a novel extension of Max-Plus (or Variable Elimination) that constrains local actions to guarantee safety criteria. An empirical evaluation on multi-agent SysAdmin and multi-UAV Delivery shows that the approach scales to very large domains where state-of-the-art methods cannot work.

Smoothed particle hydrodynamics (SPH) is omnipresent in modern engineering and scientific disciplines. SPH is a class of Lagrangian schemes that discretize fluid dynamics via finite material points that are tracked through the evolving velocity field. Due to the particle-like nature of the simulation, graph neural networks (GNNs) have emerged as appealing and successful surrogates. However, the practical utility of such GNN-based simulators relies on their ability to faithfully model physics, providing accurate and stable predictions over long time horizons - which is a notoriously hard problem. In this work, we identify particle clustering originating from tensile instabilities as one of the primary pitfalls. Based on these insights, we enhance both training and rollout inference of state-of-the-art GNN-based simulators with varying components from standard SPH solvers, including pressure, viscous, and external force components. All Neural SPH-enhanced simulators achieve better performance than the baseline GNNs, often by orders of magnitude in terms of rollout error, allowing for significantly longer rollouts and significantly better physics modeling. Code available under https://github.com/tumaer/neuralsph.

Imitation learning addresses the challenge of learning by observing an expert’s demonstrations without access to reward signals from environments. Most existing imitation learning methods that do not require interacting with environments either model the expert distribution as the conditional probability p(a|s) (e.g., behavioral cloning, BC) or the joint probability p(s, a). Despite the simplicity of modeling the conditional probability with BC, it usually struggles with generalization. While modeling the joint probability can improve generalization performance, the inference procedure is often time-consuming, and the model can suffer from manifold overfitting. This work proposes an imitation learning framework that benefits from modeling both the conditional and joint probability of the expert distribution. Our proposed Diffusion Model-Augmented Behavioral Cloning (DBC) employs a diffusion model trained to model expert behaviors and learns a policy to optimize both the BC loss (conditional) and our proposed diffusion model loss (joint). DBC outperforms baselines in various continuous control tasks in navigation, robot arm manipulation, dexterous manipulation, and locomotion. We design additional experiments to verify the limitations of modeling either the conditional probability or the joint probability of the expert distribution, as well as compare different generative models. Ablation studies justify the effectiveness of our design choices.

*Relative overgeneralization* (RO) occurs in cooperative multi-agent learning tasks when agents converge towards a suboptimal joint policy due to overfitting to suboptimal behaviors of other agents. No methods have been proposed for addressing RO in multi-agent policy gradient (MAPG) methods although these methods produce state-of-the-art results. To address this gap, we propose a general, yet simple, framework to enable optimistic updates in MAPG methods that alleviate the RO problem. Our approach involves clipping the advantage to eliminate negative values, thereby facilitating optimistic updates in MAPG. The optimism prevents individual agents from quickly converging to a local optimum. Additionally, we provide a formal analysis to show that the proposed method retains optimality at a fixed point. In extensive evaluations on a diverse set of tasks including the *Multi-agent MuJoCo* and *Overcooked* benchmarks, our method outperforms strong baselines on 13 out of 19 tested tasks and matches the performance on the rest.