NeurIPS 2023 Papers

Imitation learning from human feedback studies how to train well-performed imitation agents with an annotator's relative comparison of two demonstrations (one demonstration is better/worse than the other), which is usually easier to collect than the perfect expert data required by traditional imitation learning. However, in many real-world applications, it is still expensive or even impossible to provide a clear pairwise comparison between two demonstrations with similar quality. This motivates us to study the problem of imitation learning with vague feedback, where the data annotator can only distinguish the paired demonstrations correctly when their quality differs significantly, i.e., one from the expert and another from the non-expert. By modeling the underlying demonstration pool as a mixture of expert and non-expert data, we show that the expert policy distribution can be recovered when the proportion $\alpha$ of expert data is known. We also propose a mixture proportion estimation method for the unknown $\alpha$ case. Then, we integrate the recovered expert policy distribution with generative adversarial imitation learning to form an end-to-end algorithm. Experiments show that our methods outperform standard and preference-based imitation learning methods on various tasks.

We investigate replicable learning algorithms. Informally a learning algorithm is replicable if the algorithm outputs the same canonical hypothesis over multiple runs with high probability, even when different runs observe a different set of samples from the unknown data distribution. In general, such a strong notion of replicability is not achievable. Thus we consider two feasible notions of replicability called {\em list replicability} and {\em certificate replicability}. Intuitively, these notions capture the degree of (non) replicability. The goal is to design learning algorithms with optimal list and certificate complexities while minimizing the sample complexity. Our contributions are the following. 1. We first study the learning task of estimating the biases of $d$ coins, up to an additive error of $\varepsilon$, by observing samples. For this task, we design a $(d+1)$-list replicable algorithm. To complement this result, we establish that the list complexity is optimal, i.e there are no learning algorithms with a list size smaller than $d+1$ for this task. We also design learning algorithms with certificate complexity $\tilde{O}(\log d)$. The sample complexity of both these algorithms is $\tilde{O}(\frac{d^2}{\varepsilon^2})$ where $\varepsilon$ is the approximation error parameter (for a constant error probability). 2. In the PAC model, we show that any hypothesis class that is learnable with $d$-nonadaptive statistical queries can be learned via a $(d+1)$-list replicable algorithm and also via a $\tilde{O}(\log d)$-certificate replicable algorithm. The sample complexity of both these algorithms is $\tilde{O}(\frac{d^2}{\nu^2})$ where $\nu$ is the approximation error of the statistical query. We also show that for the concept class \dtep, the list complexity is exactly $d+1$ with respect to the uniform distribution. To establish our upper bound results we use rounding schemes induced by geometric partitions with certain properties. We use Sperner/KKM Lemma to establish the lower bound results.

Language models (LMs) often exhibit undesirable text generation behaviors, including generating false, toxic, or irrelevant outputs. Reinforcement learning from human feedback (RLHF)---where human preference judgments on LM outputs are transformed into a learning signal---has recently shown promise in addressing these issues. However, such holistic feedback conveys limited information on long text outputs; it does not indicate which aspects of the outputs influenced user preference; e.g., which parts contain what type(s) of errors. In this paper, we use fine-grained human feedback (e.g., which sentence is false, which sub-sentence is irrelevant) as an explicit training signal. We introduce Fine-Grained RLHF, a framework that enables training and learning from reward functions that are fine-grained in two respects: (1) density, providing a reward after every segment (e.g., a sentence) is generated; and (2) incorporating multiple reward models associated with different feedback types (e.g., factual incorrectness, irrelevance, and information incompleteness). We conduct experiments on detoxification and long-form question answering to illustrate how learning with this reward function leads to improved performance, supported by both automatic and human evaluation. Additionally, we show that LM behaviors can be customized using different combinations of fine-grained reward models. We release all data, collected human feedback, and codes at https://FineGrainedRLHF.github.io.

Grouping is ubiquitous in natural systems and is essential for promoting efficiency in team coordination. This paper proposes a novel formulation of Group-oriented Multi-Agent Reinforcement Learning (GoMARL), which learns automatic grouping without domain knowledge for efficient cooperation. In contrast to existing approaches that attempt to directly learn the complex relationship between the joint action-values and individual utilities, we empower subgroups as a bridge to model the connection between small sets of agents and encourage cooperation among them, thereby improving the learning efficiency of the whole team. In particular, we factorize the joint action-values as a combination of group-wise values, which guide agents to improve their policies in a fine-grained fashion. We present an automatic grouping mechanism to generate dynamic groups and group action-values. We further introduce a hierarchical control for policy learning that drives the agents in the same group to specialize in similar policies and possess diverse strategies for various groups. Experiments on the StarCraft II micromanagement tasks and Google Research Football scenarios verify our method's effectiveness. Extensive component studies show how grouping works and enhances performance.

Prior research on policy adaptation in multi-agent games has often relied on online interaction with the target agent in training, which can be expensive and impractical in real-world scenarios. Inspired by recent progress in offline reinforcement learn- ing, this paper studies offline policy adaptation, which aims to utilize the target agent’s behavior data to exploit its weakness or enable effective cooperation. We investigate its distinct challenges of distributional shift and risk-free deviation, and propose a novel learning objective, conservative offline adaptation, that optimizes the worst-case performance against any dataset consistent proxy models. We pro- pose an efficient algorithm called Constrained Self-Play (CSP) that incorporates dataset information into regularized policy learning. We prove that CSP learns a near-optimal risk-free offline adaptation policy upon convergence. Empirical results demonstrate that CSP outperforms non-conservative baselines in various environments, including Maze, predator-prey, MuJoCo, and Google Football.

Many machine learning approaches for decision making, such as reinforcement learning, rely on simulators or predictive models to forecast the time-evolution of quantities of interest, e.g., the state of an agent or the reward of a policy. Forecasts of such complex phenomena are commonly described by highly nonlinear dynamical systems, making their use in optimization-based decision-making challenging. Koopman operator theory offers a beneficial paradigm for addressing this problem by characterizing forecasts via linear time-invariant (LTI) ODEs, turning multi-step forecasts into sparse matrix multiplication. Though there exists a variety of learning approaches, they usually lack crucial learning-theoretic guarantees, making the behavior of the obtained models with increasing data and dimensionality unclear. We address the aforementioned by deriving a universal Koopman-invariant reproducing kernel Hilbert space (RKHS) that solely spans transformations into LTI dynamical systems. The resulting Koopman Kernel Regression (KKR) framework enables the use of statistical learning tools from function approximation for novel convergence results and generalization error bounds under weaker assumptions than existing work. Our experiments demonstrate superior forecasting performance compared to Koopman operator and sequential data predictors in RKHS.

In the field of unsupervised skill discovery (USD), a major challenge is limited exploration, primarily due to substantial penalties when skills deviate from their initial trajectories. To enhance exploration, recent methodologies employ auxiliary rewards to maximize the epistemic uncertainty or entropy of states. However, we have identified that the effectiveness of these rewards declines as the environmental complexity rises. Therefore, we present a novel USD algorithm, skill **disco**very with gui**dance** (**DISCO-DANCE**), which (1) selects the guide skill that possesses the highest potential to reach unexplored states, (2) guides other skills to follow guide skill, then (3) the guided skills are dispersed to maximize their discriminability in unexplored states. Empirical evaluation demonstrates that DISCO-DANCE outperforms other USD baselines in challenging environments, including two navigation benchmarks and a continuous control benchmark. Qualitative visualizations and code of DISCO-DANCE are available at https://mynsng.github.io/discodance/.

Antibodies have become an important class of therapeutic agents to treat human diseases. To accelerate therapeutic antibody discovery, computational methods, especially machine learning, have attracted considerable interest for predicting specific interactions between antibody candidates and target antigens such as viruses and bacteria. However, the publicly available datasets in existing works have notable limitations, such as small sizes and the lack of non-binding samples and exact amino acid sequences. To overcome these limitations, we have developed AVIDa-hIL6, a large-scale dataset for predicting antigen-antibody interactions in the variable domain of heavy chain of heavy chain antibodies (VHHs), produced from an alpaca immunized with the human interleukin-6 (IL-6) protein, as antigens. By leveraging the simple structure of VHHs, which facilitates identification of full-length amino acid sequences by DNA sequencing technology, AVIDa-hIL6 contains 573,891 antigen-VHH pairs with amino acid sequences. All the antigen-VHH pairs have reliable labels for binding or non-binding, as generated by a novel labeling method. Furthermore, via introduction of artificial mutations, AVIDa-hIL6 contains 30 different mutants in addition to wild-type IL-6 protein. This characteristic provides opportunities to develop machine learning models for predicting changes in antibody binding by antigen mutations. We report experimental benchmark results on AVIDa-hIL6 by using machine learning models. The results indicate that the existing models have potential, but further research is needed to generalize them to predict effective antibodies against unknown mutants. The dataset is available at https://avida-hil6.cognanous.com.

Despite the recent advancements in offline reinforcement learning via supervised learning (RvS) and the success of the decision transformer (DT) architecture in various domains, DTs have fallen short in several challenging benchmarks. The root cause of this underperformance lies in their inability to seamlessly connect segments of suboptimal trajectories. To overcome this limitation, we present a novel approach to enhance RvS methods by integrating intermediate targets. We introduce the Waypoint Transformer (WT), using an architecture that builds upon the DT framework and conditioned on automatically-generated waypoints. The results show a significant increase in the final return compared to existing RvS methods, with performance on par or greater than existing state-of-the-art temporal difference learning-based methods. Additionally, the performance and stability improvements are largest in the most challenging environments and data configurations, including AntMaze Large Play/Diverse and Kitchen Mixed/Partial.

Offline reinforcement learning (RL) has received considerable attention in recent years due to its attractive capability of learning policies from offline datasets without environmental interactions. Despite some success in the single-agent setting, offline multi-agent RL (MARL) remains to be a challenge. The large joint state-action space and the coupled multi-agent behaviors pose extra complexities for offline policy optimization. Most existing offline MARL studies simply apply offline data-related regularizations on individual agents, without fully considering the multi-agent system at the global level. In this work, we present OMIGA, a new offline multi-agent RL algorithm with implicit global-to-local value regularization. OMIGA provides a principled framework to convert global-level value regularization into equivalent implicit local value regularizations and simultaneously enables in-sample learning, thus elegantly bridging multi-agent value decomposition and policy learning with offline regularizations. Based on comprehensive experiments on the offline multi-agent MuJoCo and StarCraft II micro-management tasks, we show that OMIGA achieves superior performance over the state-of-the-art offline MARL methods in almost all tasks.

Despite the success of neural-based combinatorial optimization methods for end-to-end heuristic learning, out-of-distribution generalization remains a challenge. In this paper, we present a novel formulation of Combinatorial Optimization Problems (COPs) as Markov Decision Processes (MDPs) that effectively leverages common symmetries of COPs to improve out-of-distribution robustness. Starting from a direct MDP formulation of a constructive method, we introduce a generic way to reduce the state space, based on Bisimulation Quotienting (BQ) in MDPs. Then, for COPs with a recursive nature, we specialize the bisimulation and show how the reduced state exploits the symmetries of these problems and facilitates MDP solving. Our approach is principled and we prove that an optimal policy for the proposed BQ-MDP actually solves the associated COPs. We illustrate our approach on five classical problems: the Euclidean and Asymmetric Traveling Salesman, Capacitated Vehicle Routing, Orienteering and Knapsack Problems. Furthermore, for each problem, we introduce a simple attention-based policy network for the BQ-MDPs, which we train by imitation of (near) optimal solutions of small instances from a single distribution. We obtain new state-of-the-art results for the five COPs on both synthetic and realistic benchmarks. Notably, in contrast to most existing neural approaches, our learned policies show excellent generalization performance to much larger instances than seen during training, without any additional search procedure. Our code is available at: [link](https://github.com/naver/bq-nco).

We present an oracle-efficient relaxation for the adversarial contextual bandits problem, where the contexts are sequentially drawn i.i.d from a known distribution and the cost sequence is chosen by an online adversary. Our algorithm has a regret bound of $O(T^{\frac{2}{3}}(K\log(|\Pi|))^{\frac{1}{3}})$ and makes at most $O(K)$ calls per round to an offline optimization oracle, where $K$ denotes the number of actions, $T$ denotes the number of rounds and $\Pi$ denotes the set of policies. This is the first result to improve the prior best bound of $O((TK)^{\frac{2}{3}}(\log(|\Pi|))^{\frac{1}{3}})$ as obtained by Syrgkanis et al. at NeurIPS 2016, and the first to match the original bound of Langford and Zhang at NeurIPS 2007 which was obtained for the stochastic case.

We propose regularity as a novel reward signal for intrinsically-motivated reinforcement learning. Taking inspiration from child development, we postulate that striving for structure and order helps guide exploration towards a subspace of tasks that are not favored by naive uncertainty-based intrinsic rewards. Our generalized formulation of Regularity as Intrinsic Reward (RaIR) allows us to operationalize it within model-based reinforcement learning. In a synthetic environment, we showcase the plethora of structured patterns that can emerge from pursuing this regularity objective. We also demonstrate the strength of our method in a multi-object robotic manipulation environment. We incorporate RaIR into free play and use it to complement the model’s epistemic uncertainty as an intrinsic reward. Doing so, we witness the autonomous construction of towers and other regular structures during free play, which leads to a substantial improvement in zero-shot downstream task performance on assembly tasks.

Restricting the variance of a policy’s return is a popular choice in risk-averse Reinforcement Learning (RL) due to its clear mathematical definition and easy interpretability. Traditional methods directly restrict the total return variance. Recent methods restrict the per-step reward variance as a proxy. We thoroughly examine the limitations of these variance-based methods, such as sensitivity to numerical scale and hindering of policy learning, and propose to use an alternative risk measure, Gini deviation, as a substitute. We study various properties of this new risk measure and derive a policy gradient algorithm to minimize it. Empirical evaluation in domains where risk-aversion can be clearly defined, shows that our algorithm can mitigate the limitations of variance-based risk measures and achieves high return with low risk in terms of variance and Gini deviation when others fail to learn a reasonable policy.

Bilevel Optimization has witnessed notable progress recently with new emerging efficient algorithms. However, its application in the Federated Learning setting remains relatively underexplored, and the impact of Federated Learning's inherent challenges on the convergence of bilevel algorithms remain obscure. In this work, we investigate Federated Bilevel Optimization problems and propose a communication-efficient algorithm, named FedBiOAcc. The algorithm leverages an efficient estimation of the hyper-gradient in the distributed setting and utilizes the momentum-based variance-reduction acceleration. Remarkably, FedBiOAcc achieves a communication complexity $O(\epsilon^{-1})$, a sample complexity $O(\epsilon^{-1.5})$ and the linear speed up with respect to the number of clients. We also analyze a special case of the Federated Bilevel Optimization problems, where lower level problems are locally managed by clients. We prove that FedBiOAcc-Local, a modified version of FedBiOAcc, converges at the same rate for this type of problems. Finally, we validate the proposed algorithms through two real-world tasks: Federated Data-cleaning and Federated Hyper-representation Learning. Empirical results show superior performance of our algorithms.

We investigate robust model-free reinforcement learning algorithms designed for environments that may be dynamic or even adversarial. Traditional state-based policies often struggle to accommodate the challenges imposed by the presence of unmodeled disturbances in such settings. Moreover, optimizing linear state-based policies pose an obstacle for efficient optimization, leading to nonconvex objectives, even in benign environments like linear dynamical systems. Drawing inspiration from recent advancements in model-based control, we intro- duce a novel class of policies centered on disturbance signals. We define several categories of these signals, which we term pseudo-disturbances, and develop corresponding policy classes based on them. We provide efficient and practical algorithms for optimizing these policies. Next, we examine the task of online adaptation of reinforcement learning agents in the face of adversarial disturbances. Our methods seamlessly integrate with any black-box model-free approach, yielding provable regret guarantees when dealing with linear dynamics. These regret guarantees unconditionally improve the best-known results for bandit linear control in having no dependence on the state-space dimension. We evaluate our method over various standard RL benchmarks and demonstrate improved robustness.

We consider the problem of interactive decision making, encompassing structured bandits and reinforcement learning with general function approximation. Recently, Foster et al. (2021) introduced the Decision-Estimation Coefficient, a measure of statistical complexity that lower bounds the optimal regret for interactive decision making, as well as a meta-algorithm, Estimation-to-Decisions, which achieves upper bounds in terms of the same quantity. Estimation-to-Decisions is a reduction, which lifts algorithms for (supervised) online estimation into algorithms for decision making. In this paper, we show that by combining Estimation-to-Decisions with a specialized form of "optimistic" estimation introduced by Zhang (2022), it is possible to obtain guarantees that improve upon those of Foster et al. (2021) by accommodating more lenient notions of estimation error. We use this approach to derive regret bounds for model-free reinforcement learning with value function approximation, and give structural results showing when it can and cannot help more generally.

We focus on the task of approximating the optimal value function in deep reinforcement learning. This iterative process is comprised of solving a sequence of optimization problems where the loss function changes per iteration. The common approach to solving this sequence of problems is to employ modern variants of the stochastic gradient descent algorithm such as Adam. These optimizers maintain their own internal parameters such as estimates of the first-order and the second-order moments of the gradient, and update them over time. Therefore, information obtained in previous iterations is used to solve the optimization problem in the current iteration. We demonstrate that this can contaminate the moment estimates because the optimization landscape can change arbitrarily from one iteration to the next one. To hedge against this negative effect, a simple idea is to reset the internal parameters of the optimizer when starting a new iteration. We empirically investigate this resetting idea by employing various optimizers in conjunction with the Rainbow algorithm. We demonstrate that this simple modification significantly improves the performance of deep RL on the Atari benchmark.

Exploration strategy plays an important role in reinforcement learning, especially in sparse-reward tasks. In cooperative multi-agent reinforcement learning~(MARL), designing a suitable exploration strategy is much more challenging due to the large state space and the complex interaction among agents. Currently, mainstream exploration methods in MARL either contribute to exploring the unfamiliar states which are large and sparse, or measuring the interaction among agents with high computational costs. We found an interesting phenomenon that different kinds of exploration plays a different role in different MARL scenarios, and choosing a suitable one is often more effective than designing an exquisite algorithm. In this paper, we propose a exploration method that incorporate the \underline{C}uri\underline{O}sity-based and \underline{IN}fluence-based exploration~(COIN) which is simple but effective in various situations. First, COIN measures the influence of each agent on the other agents based on mutual information theory and designs it as intrinsic rewards which are applied to each individual value function. Moreover, COIN computes the curiosity-based intrinsic rewards via prediction errors which are added to the extrinsic reward. For integrating the two kinds of intrinsic rewards, COIN utilizes a novel framework in which they complement each other and lead to a sufficient and effective exploration on cooperative MARL tasks. We perform extensive experiments on different challenging benchmarks, and results across different scenarios show the superiority of our method.