ICML 2024 Papers

Bilevel optimization is widely applied in many machine learning tasks such as hyper-parameter learning, meta learning and reinforcement learning. Although many algorithms recently have been developed to solve the bilevel optimization problems, they generally rely on the (strongly) convex lower-level problems. More recently, some methods have been proposed to solve the nonconvex-PL bilevel optimization problems, where their upper-level problems are possibly nonconvex, and their lower-level problems are also possibly nonconvex while satisfying Polyak-Łojasiewicz (PL) condition. However, these methods still have a high convergence complexity or a high computation complexity such as requiring compute expensive Hessian/Jacobian matrices and its inverses. In the paper, thus, we propose an efficient Hessian/Jacobian-free method (i.e., HJFBiO) with the optimal convergence complexity to solve the nonconvex-PL bilevel problems. Theoretically, under some mild conditions, we prove that our HJFBiO method obtains an optimal convergence rate of $O(\frac{1}{T})$, where $T$ denotes the number of iterations, and has an optimal gradient complexity of $O(\epsilon^{-1})$ in finding an $\epsilon$-stationary solution. We conduct some numerical experiments on the bilevel PL game and hyper-representation learning task to demonstrate efficiency of our proposed method.

Efficient and effective Out-of-Distribution (OOD) detection is essential for the safe deployment of AI systems. Existing feature space methods, while effective, often incur significant computational overhead due to their reliance on auxiliary models built from training features. In this paper, we propose a computationally-efficient OOD detector without using auxiliary models while still leveraging the rich information embedded in the feature space. Specifically, we detect OOD samples based on their feature distances to decision boundaries. To minimize computational cost, we introduce an efficient closed-form estimation, analytically proven to tightly lower bound the distance. Based on our estimation, we discover that In-Distribution (ID) features tend to be further from decision boundaries than OOD features. Additionally, ID and OOD samples are better separated when compared at equal deviation levels from the mean of training features. By regularizing the distances to decision boundaries based on feature deviation from the mean, we develop a hyperparameter-free, auxiliary model-free OOD detector. Our method matches or surpasses the effectiveness of state-of-the-art methods in extensive experiments while incurring negligible overhead in inference latency. Overall, our approach significantly improves the efficiency-effectiveness trade-off in OOD detection. Code is available at: https://github.com/litianliu/fDBD-OOD.

Deep Reinforcement Learning (DRL) agents have demonstrated impressive success in a wide range of game genres. However, existing research primarily focuses on optimizing DRL competence rather than addressing the challenge of prolonged player interaction. In this paper, we propose a practical DRL agent system for fighting games named _Shūkai_, which has been successfully deployed to Naruto Mobile, a popular fighting game with over 100 million registered users. _Shūkai_ quantifies the state to enhance generalizability, introducing Heterogeneous League Training (HELT) to achieve balanced competence, generalizability, and training efficiency. Furthermore, _Shūkai_ implements specific rewards to align the agent's behavior with human expectations. _Shūkai_'s ability to generalize is demonstrated by its consistent competence across all characters, even though it was trained on only 13% of them. Additionally, HELT exhibits a remarkable 22% improvement in sample efficiency. _Shūkai_ serves as a valuable training partner for players in Naruto Mobile, enabling them to enhance their abilities and skills.

In this paper, we study offline-to-online Imitation Learning (IL) that pretrains an imitation policy from static demonstration data, followed by fast finetuning with minimal environmental interaction. We find the naive combination of existing offline IL and online IL methods tends to behave poorly in this context, because the initial discriminator (often used in online IL) operates randomly and discordantly against the policy initialization, leading to misguided policy optimization and *unlearning* of pretraining knowledge. To overcome this challenge, we propose a principled offline-to-online IL method, named OLLIE, that simultaneously learns a near-expert policy initialization along with an *aligned discriminator initialization*, which can be seamlessly integrated into online IL, achieving smooth and fast finetuning. Empirically, OLLIE consistently and significantly outperforms the baseline methods in **20** challenging tasks, from continuous control to vision-based domains, in terms of performance, demonstration efficiency, and convergence speed. This work may serve as a foundation for further exploration of pretraining and finetuning in the context of IL.

While quantum reinforcement learning (RL) has attracted a surge of attention recently, its theoretical understanding is limited. In particular, it remains elusive how to design provably efficient quantum RL algorithms that can address the exploration-exploitation trade-off. To this end, we propose a novel UCRL-style algorithm that takes advantage of quantum computing for tabular Markov decision processes (MDPs) with $S$ states, $A$ actions, and horizon $H$, and establish an $\mathcal{O}(\mathrm{poly}(S, A, H, \log T))$ worst-case regret for it, where $T$ is the number of episodes. Furthermore, we extend our results to quantum RL with linear function approximation, which is capable of handling problems with large state spaces. Specifically, we develop a quantum algorithm based on value target regression (VTR) for linear mixture MDPs with $d$-dimensional linear representation and prove that it enjoys $\mathcal{O}(\mathrm{poly}(d, H, \log T))$ regret. Our algorithms are variants of UCRL/UCRL-VTR algorithms in classical RL, which also leverage a novel combination of lazy updating mechanisms and quantum estimation subroutines. This is the key to breaking the $\Omega(\sqrt{T})$-regret barrier in classical RL. To the best of our knowledge, this is the first work studying the online exploration in quantum RL with provable logarithmic worst-case regret.

Diffusion models excel at modeling complex data distributions, including those of images, proteins, and small molecules. However, in many cases, our goal is to model parts of the distribution that maximize certain properties: for example, we may want to generate images with high aesthetic quality, or molecules with high bioactivity. It is natural to frame this as a reinforcement learning (RL) problem, in which the objective is to finetune a diffusion model to maximize a reward function that corresponds to some property. Even with access to online queries of the ground-truth reward function, efficiently discovering high-reward samples can be challenging: they might have a low probability in the initial distribution, and there might be many infeasible samples that do not even have a well-defined reward (e.g., unnatural images or physically impossible molecules). In this work, we propose a novel reinforcement learning procedure that efficiently explores on the manifold of feasible samples. We present a theoretical analysis providing a regret guarantee, as well as empirical validation across three domains: images, biological sequences, and molecules.

Many roboticists dream of presenting a robot with a task in the evening and returning the next morning to find the robot capable of solving the task. What is preventing us from achieving this? Sim-to-real reinforcement learning (RL) has achieved impressive performance on challenging robotics tasks, but requires substantial human effort to set up the task in a way that is amenable to RL. It's our position that algorithmic improvements in policy optimization and other ideas should be guided towards resolving the primary bottleneck of shaping the training environment, i.e., designing observations, actions, rewards and simulation dynamics. Most practitioners don't tune the RL algorithm, but other environment parameters to obtain a desirable controller. We posit that scaling RL to diverse robotic tasks will only be achieved if the community focuses on automating environment shaping procedures.

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.

Temporal difference (TD) learning algorithms with neural network function parameterization have well-established empirical success in many practical large-scale reinforcement learning tasks. However, theoretical understanding of these algorithms remains challenging due to the nonlinearity of the action-value approximation. In this paper, we develop an improved non-asymptotic analysis of the neural TD method with a general $L$-layer neural network. New proof techniques are developed and an improved new $\tilde{\mathcal{O}}(\epsilon^{-1})$ sample complexity is derived. To our best knowledge, this is the first finite-time analysis of neural TD that achieves an $\tilde{\mathcal{O}}(\epsilon^{-1})$ complexity under the Markovian sampling, as opposed to the best known $\tilde{\mathcal{O}}(\epsilon^{-2})$ complexity in the existing literature.

Reinforcement Learning (RL) for constrained MDPs (CMDPs) is an increasingly important problem for various applications. Often, the average criterion is more suitable than the discounted criterion. Yet, RL for average-CMDPs (ACMDPs) remains a challenging problem. Algorithms designed for discounted constrained RL problems often do not perform well for the average CMDP setting. In this paper, we introduce a new policy optimization with function approximation algorithm for constrained MDPs with the average criterion. The Average-Constrained Policy Optimization (ACPO) algorithm is inspired by trust region-based policy optimization algorithms. We develop basic sensitivity theory for average CMDPs, and then use the corresponding bounds in the design of the algorithm. We provide theoretical guarantees on its performance, and through extensive experimental work in various challenging OpenAI Gym environments, show its superior empirical performance when compared to other state-of-the-art algorithms adapted for the ACMDPs.

Despite the recent successes of multi-agent reinforcement learning (MARL) algorithms, efficiently adapting to co-players in mixed-motive environments remains a significant challenge. One feasible approach is to hierarchically model co-players' behavior based on inferring their characteristics. However, these methods often encounter difficulties in efficient reasoning and utilization of inferred information. To address these issues, we propose Hierarchical Opponent modeling and Planning (HOP), a novel multi-agent decision-making algorithm that enables few-shot adaptation to unseen policies in mixed-motive environments. HOP is hierarchically composed of two modules: an opponent modeling module that infers others' goals and learns corresponding goal-conditioned policies, and a planning module that employs Monte Carlo Tree Search (MCTS) to identify the best response. Our approach improves efficiency by updating beliefs about others' goals both across and within episodes and by using information from the opponent modeling module to guide planning. Experimental results demonstrate that in mixed-motive environments, HOP exhibits superior few-shot adaptation capabilities when interacting with various unseen agents, and excels in self-play scenarios. Furthermore, the emergence of social intelligence during our experiments underscores the potential of our approach in complex multi-agent environments.

The label smoothness assumption is at the core of Graph Convolutional Networks (GCNs): nodes in a local region have similar labels. Thus, GCN performs local feature smoothing operation to adhere to this assumption. However, there exist some nodes whose labels obtained by feature smoothing conflict with the label smoothness assumption. We find that the label smoothness assumption and the process of feature smoothing are both problematic on these nodes, and call these nodes out of GCN's control (OOC nodes). In this paper, first, we design the corresponding algorithm to locate the OOC nodes, then we summarize the characteristics of OOC nodes that affect their representation learning, and based on their characteristics, we present DaGCN, an efficient framework that can facilitate the OOC nodes. Extensive experiments verify the superiority of the proposed method and demonstrate that current advanced GCNs are improvements specifically on OOC nodes; the remaining nodes under GCN's control (UC nodes) are already optimally represented by vanilla GCN on most datasets.

We study the statistical complexity of offline decision-making with function approximation, establishing (near) minimax-optimal rates for stochastic contextual bandits and Markov decision processes. The performance limits are captured by the pseudo-dimension of the (value) function class and a new characterization of the behavior policy that *strictly* subsumes all the previous notions of data coverage in the offline decision-making literature. In addition, we seek to understand the benefits of using offline data in online decision-making and show nearly minimax-optimal rates in a wide range of regimes.

Value functions are an essential component in deep reinforcement learning (RL), that are typically trained via mean squared error regression to match bootstrapped target values. However, scaling value-based RL methods to large networks has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We show that training value functions with categorical cross-entropy significantly enhances performance and scalability across various domains, including single-task RL on Atari 2600 games, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving *state-of-the-art results* on these domains. Through careful analysis, we show that categorical cross-entropy mitigates issues inherent to value-based RL, such as noisy targets and non-stationarity. We argue that shifting to categorical cross-entropy for training value functions can substantially improve the scalability of deep RL at little-to-no cost.

We study the problem of maintaining a decision tree in the fully-dynamic setting, where the dataset is updated by an adversarial sequence of insertions and deletions. We present the first algorithm with strong guarantees on both the quality of the tree and the worst-case update time (the maximum time spent between two consecutive dataset updates). For instance, we can maintain a tree where each node has Gini gain within $\beta$ of the optimum, while guaranteeing an update time $O(d \beta^{-3} \log^4 n )$, where $d$ is the number of features and $n$ the maximum size of the dataset. This is optimal up to polylogarithmic factors, as any dynamic algorithm must have update time in $\Omega(d)$. Similar guarantees hold for the variance and information gain, for classification and regression, and even for *boosted* trees. This shows that many popular decision trees such as ID3 or C4.5 can be efficiently be made dynamic, answering an open question of Bressan, Damay and Sozio (AAAI 2023). We also show that, under the 3SUM conjecture or the Orthogonal Vectors Hypothesis, the update time must be polynomial in $1/\beta$.

Large Language Models (LLMs) tend to prioritize adherence to user prompts over providing veracious responses, leading to the sycophancy issue. When challenged by users, LLMs tend to admit mistakes and provide inaccurate responses even if they initially provided the correct answer. Recent works propose to employ supervised fine-tuning (SFT) to mitigate the sycophancy issue, while it typically leads to the degeneration of LLMs' general capability. To address the challenge, we propose a novel supervised pinpoint tuning (SPT), where the region-of-interest modules are tuned for a given objective. Specifically, SPT first reveals and verifies a small percentage (<5%) of the basic modules, which significantly affect a particular behavior of LLMs. i.e., sycophancy. Subsequently, SPT merely fine-tunes these identified modules while freezing the rest. To verify the effectiveness of the proposed SPT, we conduct comprehensive experiments, demonstrating that SPT significantly mitigates the sycophancy issue of LLMs (even better than SFT). Moreover, SPT introduces limited or even no side effects on the general capability of LLMs. Our results shed light on how to precisely, effectively, and efficiently explain and improve the targeted ability of LLMs.

Deep reinforcement learning (DRL) has significantly advanced the field of combinatorial optimization (CO). However, its practicality is hindered by the necessity for a large number of reward evaluations, especially in scenarios involving computationally intensive function assessments. To enhance the sample efficiency, we propose a simple but effective method, called *symmetric replay training (SRT)*, which can be easily integrated into various DRL methods. Our method leverages high-reward samples to encourage exploration of the under-explored symmetric regions without additional online interactions - *free*. Through replay training, the policy is trained to maximize the likelihood of the symmetric trajectories of discovered high-rewarded samples. Experimental results demonstrate the consistent improvement of our method in sample efficiency across diverse DRL methods applied to real-world tasks, such as molecular optimization and hardware design.

In this paper, we consider multi-objective reinforcement learning, which arises in many real-world problems with multiple optimization goals. We approach the problem with a max-min framework focusing on fairness among the multiple goals and develop a relevant theory and a practical model-free algorithm under the max-min framework. The developed theory provides a theoretical advance in multi-objective reinforcement learning, and the proposed algorithm demonstrates a notable performance improvement over existing baseline methods.

In this paper we adopt a representation-centric perspective on exploration in reinforcement learning, viewing exploration fundamentally as a density estimation problem. We investigate the effectiveness of clustering representations for exploration in 3-D environments, based on the observation that the importance of pixel changes between transitions is less pronounced in 3-D environments compared to 2-D environments, where pixel changes between transitions are typically distinct and significant. We propose a method that performs episodic and global clustering on random representations and on pre-trained DINO representations to count states, i.e, estimate pseudo-counts. Surprisingly, even random features can be clustered effectively to count states in 3-D environments, however when these become visually more complex, pre-trained DINO representations are more effective thanks to the pre-trained inductive biases in the representations. Overall, this presents a pathway for integrating pre-trained biases into exploration. We evaluate our approach on the VizDoom and Habitat environments, demonstrating that our method surpasses other well-known exploration methods in these settings.