ICLR 2024 Papers

Offline reinforcement learning (RL), where the agent aims to learn the optimal policy based on the data collected by a behavior policy, has attracted increasing attention in recent years. While offline RL with linear function approximation has been extensively studied with optimal results achieved under certain assumptions, many works shift their interest to offline RL with non-linear function approximation. However, limited works on offline RL with non-linear function approximation have instance-dependent regret guarantees. In this paper, we propose an oracle-efficient algorithm, dubbed Pessimistic Nonlinear Least-Square Value Iteration (PNLSVI), for offline RL with non-linear function approximation. Our algorithmic design comprises three innovative components: (1) a variance-based weighted regression scheme that can be applied to a wide range of function classes, (2) a subroutine for variance estimation, and (3) a planning phase that utilizes a pessimistic value iteration approach. Our algorithm enjoys a regret bound that has a tight dependency on the function class complexity and achieves minimax optimal instance-dependent regret when specialized to linear function approximation. Our work extends the previous instance-dependent results within simpler function classes, such as linear and differentiable function to a more general framework. To the best of our knowledge, this is the first statistically optimal algorithm for nonlinear offline RL.

Offline reinforcement learning (RL) offers a promising direction for learning policies from pre-collected datasets without requiring further interactions with the environment. However, existing methods struggle to handle out-of-distribution (OOD) extrapolation errors, especially in sparse reward or scarce data settings. In this paper, we propose a novel training algorithm called Conservative Density Estimation (CDE), which addresses this challenge by explicitly imposing constraints on the state-action occupancy stationary distribution. CDE overcomes the limitations of existing approaches, such as the stationary distribution correction method, by addressing the support mismatch issue in marginal importance sampling. Our method achieves state-of-the-art performance on the D4RL benchmark. Notably, CDE consistently outperforms baselines in challenging tasks with sparse rewards or insufficient data, demonstrating the advantages of our approach in addressing the extrapolation error problem in offline RL.

Structure prediction of large protein complexes (a.k.a., protein multimer mod- elling, PMM) can be achieved through the one-by-one assembly using provided dimer structures and predicted docking paths. However, existing PMM methods struggle with vast search spaces and generalization challenges: (1) The assembly of a N -chain multimer can be depicted using graph structured data, with each chain represented as a node and assembly actions as edges. Thus the assembly graph can be arbitrary acyclic undirected connected graph, leading to the com- binatorial optimization space of N^(N −2) for the PMM problem. (2) Knowledge transfer in the PMM task is non-trivial. The gradually limited data availability as the chain number increases necessitates PMM models that can generalize across multimers of various chains. To address these challenges, we propose GAPN, a Generative Adversarial Policy Network powered by domain-specific rewards and adversarial loss through policy gradient for automatic PMM prediction. Specifi- cally, GAPN learns to efficiently search through the immense assembly space and optimize the direct docking reward through policy gradient. Importantly, we de- sign a adversarial reward function to enhance the receptive field of our model. In this way, GAPN will simultaneously focus on a specific batch of multimers and the global assembly rules learned from multimers with varying chain numbers. Empirically, we have achieved both significant accuracy (measured by RMSD and TM-Score) and efficiency improvements compared to leading complex mod- eling software. GAPN outperforms the state-of-the-art method (MoLPC) with up to 27% improvement in TM-Score, with a speed-up of 600×.

Large language models (LLMs) hold great potential for many natural language applications but risk generating incorrect or toxic content. To probe when an LLM generates unwanted content, the current paradigm is to recruit a $\textit{red team}$ of human testers to design input prompts (i.e., test cases) that elicit undesirable responses from LLMs. However, relying solely on human testers is expensive and time-consuming. Recent works automate red teaming by training a separate red team LLM with reinforcement learning (RL) to generate test cases that maximize the chance of eliciting undesirable responses from the target LLM. However, current RL methods are only able to generate a small number of effective test cases resulting in a low coverage of the span of prompts that elicit undesirable responses from the target LLM. To overcome this limitation, we draw a connection between the problem of increasing the coverage of generated test cases and the well-studied approach of curiosity-driven exploration that optimizes for novelty. Our method of curiosity-driven red teaming (CRT) achieves greater coverage of test cases while mantaining or increasing their effectiveness compared to existing methods. Our method, CRT successfully provokes toxic responses from LLaMA2 model that has been heavily fine-tuned using human preferences to avoid toxic outputs. Code is available at https://github.com/Improbable-AI/curiosity_redteam.

Despite recent progress in offline learning, these methods are still trained and tested on the same environment. In this paper, we compare the generalization abilities of widely used online and offline learning methods such as online reinforcement learning (RL), offline RL, sequence modeling, and behavioral cloning. Our experiments show that offline learning algorithms perform worse on new environments than online learning ones. We also introduce the first benchmark for evaluating generalization in offline learning, collecting datasets of varying sizes and skill-levels from Procgen (2D video games) and WebShop (e-commerce websites). The datasets contain trajectories for a limited number of game levels or natural language instructions and at test time, the agent has to generalize to new levels or instructions. Our experiments reveal that existing offline learning algorithms struggle to match the performance of online RL on both train and test environments. Behavioral cloning is a strong baseline, outperforming state-of-the-art offline RL and sequence modeling approaches when trained on data from multiple environments and tested on new ones. Finally, we find that increasing the diversity of the data, rather than its size, improves performance on new environments for all offline learning algorithms. Our study demonstrates the limited generalization of current offline learning algorithms highlighting the need for more research in this area.

We study pre-training representations for decision-making using video data, which is abundantly available for tasks such as game agents and software testing. Even though significant empirical advances have been made on this problem, a theoretical understanding remains absent. We initiate the theoretical investigation into principled approaches for representation learning and focus on learning the latent state representations of the underlying MDP using video data. We study two types of settings: one where there is iid noise in the observation, and a more challenging setting where there is also the presence of exogenous noise, which is non-iid noise that is temporally correlated, such as the motion of people or cars in the background. We study three commonly used approaches: autoencoding, temporal contrastive learning, and forward modeling. We prove upper bounds for temporal contrastive learning and forward modeling in the presence of only iid noise. We show that these approaches can learn the latent state and use it to do efficient downstream RL with polynomial sample complexity. When exogenous noise is also present, we establish a lower bound result showing that the sample complexity of learning from video data can be exponentially worse than learning from action-labeled trajectory data. This partially explains why reinforcement learning with video pre-training is hard. We evaluate these representational learning methods in two visual domains, yielding results that are consistent with our theoretical findings.

Language model self-improvement (LMSI) techniques have recently gained significant attention as they improve language models without requiring external supervision. A common approach is reinforcement learning from AI feedback (RLAIF), which trains a reward model based on AI preference data and employs a reinforcement learning algorithm to train the language model. However, RLAIF relies on the heuristic assumption that an AI model can provide effective feedback and correct wrong answers, requiring a solid capability of the language model. This paper presents a novel LMSI method, Reinforcement Learning Contemplation (RLC). We disclose that it is simpler for language models to evaluate a sentence than to generate it, even for small language models. Leveraging the gap between the evaluation and generation, RLC evaluates generated answers and updates language model parameters using reinforcement learning to maximize evaluation scores. Through testing on various challenging reasoning tasks and text summarization task, our experiments show that RLC effectively improves language model performance without external supervision, resulting in an answering accuracy increase (from 31.23% to 37.09%) for BigBench-hard reasoning tasks, and a rise in BERTScore for CNN/Daily Mail summarization tasks. Furthermore, RLC can be applied to models of different sizes, showcasing its broad applicability.

Opponent modeling aims at learning the opponent's behaviors, goals, or beliefs to reduce the uncertainty of the competitive environment and assist decision-making. Existing work has mostly focused on learning opponent models online, which is impractical and inefficient in practical scenarios. To this end, we formalize an Offline Opponent Modeling (OOM) problem with the objective of utilizing pre-collected offline datasets to learn opponent models that characterize the opponent from the viewpoint of the controlled agent, which aids in adapting to the unknown fixed policies of the opponent. Drawing on the promises of the Transformers for decision-making, we introduce a general approach, Transformer Against Opponent (TAO), for OOM. Essentially, TAO tackles the problem by harnessing the full potential of the supervised pre-trained Transformers' in-context learning capabilities. The foundation of TAO lies in three stages: an innovative offline policy embedding learning stage, an offline opponent-aware response policy training stage, and a deployment stage for opponent adaptation with in-context learning. Theoretical analysis establishes TAO's equivalence to Bayesian posterior sampling in opponent modeling and guarantees TAO's convergence in opponent policy recognition. Extensive experiments and ablation studies on competitive environments with sparse and dense rewards demonstrate the impressive performance of TAO. Our approach manifests remarkable prowess for fast adaptation, especially in the face of unseen opponent policies, confirming its in-context learning potency.

Current model-based reinforcement learning (MBRL) agents struggle with long-term dependencies. This limits their ability to effectively solve tasks involving extended time gaps between actions and outcomes, or tasks demanding the recalling of distant observations to inform current actions. To improve temporal coherence, we integrate a new family of state space models (SSMs) in world models of MBRL agents to present a new method, Recall to Imagine (R2I). This integration aims to enhance both long-term memory and long-horizon credit assignment. Through a diverse set of illustrative tasks, we systematically demonstrate that R2I not only establishes a new state-of-the-art for challenging memory and credit assignment RL tasks, such as BSuite and POPGym, but also showcases superhuman performance in the complex memory domain of Memory Maze. At the same time, it upholds comparable performance in classic RL tasks, such as Atari and DMC, suggesting the generality of our method. We also show that R2I is faster than the state-of-the-art MBRL method, DreamerV3, resulting in faster wall-time convergence.

This paper studies the sample-efficiency of learning in Partially Observable Markov Decision Processes (POMDPs), a challenging problem in reinforcement learning that is known to be exponentially hard in the worst-case. Motivated by real-world settings such as loading in game playing, we propose an enhanced feedback model called ``multiple observations in hindsight'', where after each episode of interaction with the POMDP, the learner may collect multiple additional observations emitted from the encountered latent states, but may not observe the latent states themselves. We show that sample-efficient learning under this feedback model is possible for two new subclasses of POMDPs: \emph{multi-observation revealing POMDPs} and \emph{distinguishable POMDPs}. Both subclasses generalize and substantially relax \emph{revealing POMDPs}---a widely studied subclass for which sample-efficient learning is possible under standard trajectory feedback. Notably, distinguishable POMDPs only require the emission distributions from different latent states to be \emph{different} instead of \emph{linearly independent} as required in revealing POMDPs.

Markov Decision Processes (MDPs) are a formal framework for modeling and solving sequential decision-making problems. In finite time horizons such problems are relevant for instance for optimal stopping or specific supply chain problems, but also in the training of large language models. In contrast to infinite horizon MDPs optimal policies are not stationary, policies must be learned for every single epoch. In practice all parameters are often trained simultaneously, ignoring the inherent structure suggested by dynamic programming. This paper introduces a combination of dynamic programming and policy gradient called dynamical policy gradient, where the parameters are trained backwards in time. For the tabular softmax parametrisation we carry out the convergence analysis for simultaneous and dynamic policy gradient towards global optima, both in the exact and sampled gradient settings without regularisation. It turns out that the use of dynamic policy gradient training much better exploits the structure of finite-time problems which is reflected in improved convergence bounds.

Meta-reinforcement learning (meta-RL) is a promising framework for tackling challenging domains requiring efficient exploration. Existing meta-RL algorithms are characterized by low sample efficiency, and mostly focus on low-dimensional task distributions. In parallel, model-based RL methods have been successful in solving partially observable MDPs, of which meta-RL is a special case. In this work, we leverage this success and propose a new model-based approach to meta-RL, based on elements from existing state-of-the-art model-based and meta-RL methods. We demonstrate the effectiveness of our approach on common meta-RL benchmark domains, attaining greater return with better sample efficiency (up to $15\times$) while requiring very little hyperparameter tuning. In addition, we validate our approach on a slate of more challenging, higher-dimensional domains, taking a step towards real-world generalizing agents.

Predicting and reasoning about the future lie at the heart of many time-series questions. For example, goal-conditioned reinforcement learning can be viewed as learning representations to predict which states are likely to be visited in the future. While prior methods have used contrastive predictive coding to model time series data, learning representations that encode long-term dependencies usually requires large amounts of data. In this paper, we introduce a temporal difference version of contrastive predictive coding that stitches together pieces of different time series data to decrease the amount of data required to learn predictions of future events. We apply this representation learning method to derive an off-policy algorithm for goal-conditioned RL. Experiments demonstrate that, compared with prior RL methods, ours achieves $2 \times$ median improvement in success rates and can better cope with stochastic environments. In tabular settings, we show that our method is about $20\times$ more sample efficient than the successor representation and $1500 \times$ more sample efficient than the standard (Monte Carlo) version of contrastive predictive coding.

Plasticity, the ability of a neural network to evolve with new data, is crucial for high-performance and sample-efficient visual reinforcement learning (VRL). Although methods like resetting and regularization can potentially mitigate plasticity loss, the influences of various components within the VRL framework on the agent's plasticity are still poorly understood. In this work, we conduct a systematic empirical exploration focusing on three primary underexplored facets and derive the following insightful conclusions: (1) data augmentation is essential in maintaining plasticity; (2) the critic's plasticity loss serves as the principal bottleneck impeding efficient training; and (3) without timely intervention to recover critic's plasticity in the early stages, its loss becomes catastrophic. These insights suggest a novel strategy to address the high replay ratio (RR) dilemma, where exacerbated plasticity loss hinders the potential improvements of sample efficiency brought by increased reuse frequency. Rather than setting a static RR for the entire training process, we propose Adaptive RR, which dynamically adjusts the RR based on the critic’s plasticity level. Extensive evaluations indicate that Adaptive RR not only avoids catastrophic plasticity loss in the early stages but also benefits from more frequent reuse in later phases, resulting in superior sample efficiency.

Training generalist agents is difficult across several axes, requiring us to deal with high-dimensional inputs (space), long horizons (time), and generalization to novel tasks. Recent advances with architectures have allowed for improved scaling along one or two of these axes, but are still computationally prohibitive to use. In this paper, we propose to address all three axes by leveraging Language to Control Diffusion models as a hierarchical planner conditioned on language (LCD). We effectively and efficiently scale diffusion models for planning in extended temporal, state, and task dimensions to tackle long horizon control problems conditioned on natural language instructions, as a step towards generalist agents. Comparing LCD with other state-of-the-art models on the CALVIN language benchmark finds that LCD outperforms other SOTA methods in multi-task success rates, whilst improving inference speed over other comparable diffusion models by 3.3x~15x. We show that LCD can successfully leverage the unique strength of diffusion models to produce coherent long range plans while addressing their weakness in generating low-level details and control.