This manuscript gives a big-picture, up-to-date overview of the field of (deep) reinforcement learning and sequential decision making, covering value-based methods, policy-based methods, model-based methods, multi-agent RL, LLMs and RL, and various other topics (e.g., offline RL, hierarchical RL, intrinsic reward).
Diversity in demonstration selection is crucial for enhancing model generalization, as it enables a broader coverage of structures and concepts. However, constructing an appropriate set of demonstrations has remained a focal point of research. This paper presents the Relevance-Diversity Enhanced Selection (RDES), an innovative approach that leverages reinforcement learning to optimize the selection of diverse reference demonstrations for text classification tasks using Large Language Models (LLMs), especially in few-shot prompting scenarios. RDES employs a Q-learning framework to dynamically identify demonstrations that maximize both diversity and relevance to the classification objective by calculating a diversity score based on label distribution among selected demonstrations. This method ensures a balanced representation of reference data, leading to improved classification accuracy. Through extensive experiments on four benchmark datasets and involving 12 closed-source and open-source LLMs, we demonstrate that RDES significantly enhances classification accuracy compared to ten established baselines. Furthermore, we investigate the incorporation of Chain-of-Thought (CoT) reasoning in the reasoning process, which further enhances the model's predictive performance. The results underscore the potential of reinforcement learning to facilitate adaptive demonstration selection and deepen the understanding of classification challenges.
The high costs and risks involved in extensive environment interactions hinder the practical application of current online safe reinforcement learning (RL) methods. While offline safe RL addresses this by learning policies from static datasets, the performance therein is usually limited due to reliance on data quality and challenges with out-of-distribution (OOD) actions. Inspired by recent successes in offline-to-online (O2O) RL, it is crucial to explore whether offline safe RL can be leveraged to facilitate faster and safer online policy learning, a direction that has yet to be fully investigated. To fill this gap, we first demonstrate that naively applying existing O2O algorithms from standard RL would not work well in the safe RL setting due to two unique challenges: \emph{erroneous Q-estimations}, resulted from offline-online objective mismatch and offline cost sparsity, and \emph{Lagrangian mismatch}, resulted from difficulties in aligning Lagrange multipliers between offline and online policies. To address these challenges, we introduce \textbf{Marvel}, a novel framework for O2O safe RL, comprising two key components that work in concert: \emph{Value Pre-Alignment} to align the Q-functions with the underlying truth before online learning, and \emph{Adaptive PID Control} to effectively adjust the Lagrange multipliers during online finetuning. Extensive experiments demonstrate that Marvel significantly outperforms existing baselines in both reward maximization and safety constraint satisfaction. By introducing the first policy-finetuning based framework for O2O safe RL, which is compatible with many offline and online safe RL methods, our work has the great potential to advance the field towards more efficient and practical safe RL solutions.
This paper proposes \emph{Episodic and Lifelong Exploration via Maximum ENTropy} (ELEMENT), a novel, multiscale, intrinsically motivated reinforcement learning (RL) framework that is able to explore environments without using any extrinsic reward and transfer effectively the learned skills to downstream tasks. We advance the state of the art in three ways. First, we propose a multiscale entropy optimization to take care of the fact that previous maximum state entropy, for lifelong exploration with millions of state observations, suffers from vanishing rewards and becomes very expensive computationally across iterations. Therefore, we add an episodic maximum entropy over each episode to speedup the search further. Second, we propose a novel intrinsic reward for episodic entropy maximization named \emph{average episodic state entropy} which provides the optimal solution for a theoretical upper bound of the episodic state entropy objective. Third, to speed the lifelong entropy maximization, we propose a $k$ nearest neighbors ($k$NN) graph to organize the estimation of the entropy and updating processes that reduces the computation substantially. Our ELEMENT significantly outperforms state-of-the-art intrinsic rewards in both episodic and lifelong setups. Moreover, it can be exploited in task-agnostic pre-training, collecting data for offline reinforcement learning, etc.
The exploration \& exploitation dilemma poses significant challenges in reinforcement learning (RL). Recently, curiosity-based exploration methods achieved great success in tackling hard-exploration problems. However, they necessitate extensive hyperparameter tuning on different environments, which heavily limits the applicability and accessibility of this line of methods. In this paper, we characterize this problem via analysis of the agent behavior, concluding the fundamental difficulty of choosing a proper hyperparameter. We then identify the difficulty and the instability of the optimization when the agent learns with curiosity. We propose our method, hyperparameter robust exploration (\textbf{Hyper}), which extensively mitigates the problem by effectively regularizing the visitation of the exploration and decoupling the exploitation to ensure stable training. We theoretically justify that \textbf{Hyper} is provably efficient under function approximation setting and empirically demonstrate its appealing performance and robustness in various environments.
Lane change decision-making is a complex task due to intricate vehicle-vehicle and vehicle-infrastructure interactions. Existing algorithms for lane-change control often depend on vehicles with a certain level of autonomy (e.g., autonomous or connected autonomous vehicles). To address the challenges posed by low penetration rates of autonomous vehicles and the high costs of precise data collection, this study proposes a dynamic lane change regulation design based on multi-agent reinforcement learning (MARL) to enhance freeway traffic efficiency. The proposed framework leverages multi-lane macroscopic traffic models that describe spatial-temporal dynamics of the density and speed for each lane. Lateral traffic flow between adjacent lanes, resulting from aggregated lane-changing behaviors, is modeled as source terms exchanged between the partial differential equations (PDEs). We propose a lane change regulation strategy using MARL, where one agent is placed at each discretized lane grid. The state of each agent is represented by aggregated vehicle attributes within its grid, generated from the SUMO microscopic simulation environment. The agent's actions are lane-change regulations for vehicles in its grid. Specifically, lane-change regulation signals are computed at a centralized traffic management center and then broadcast to connected vehicles in the corresponding lane grids. Compared to vehicle-level maneuver control, this approach achieves a higher regulation rate by leveraging vehicle connectivity while introducing no critical safety concerns, and accommodating varying levels of connectivity and autonomy within the traffic system. The proposed model is simulated and evaluated in varied traffic scenarios and demand conditions. Experimental results demonstrate that the method improves overall traffic efficiency with minimal additional energy consumption while maintaining driving safety.
The reliable deployment of deep reinforcement learning in real-world settings requires the ability to generalize across a variety of conditions, including both in-distribution scenarios seen during training as well as novel out-of-distribution scenarios. In this work, we present a framework for dynamics generalization in deep reinforcement learning that unifies these two distinct types of generalization within a single architecture. We introduce a robust adaptation module that provides a mechanism for identifying and reacting to both in-distribution and out-of-distribution environment dynamics, along with a joint training pipeline that combines the goals of in-distribution adaptation and out-of-distribution robustness. Our algorithm GRAM achieves strong generalization performance across in-distribution and out-of-distribution scenarios upon deployment, which we demonstrate through extensive simulation and hardware locomotion experiments on a quadruped robot.
We propose to learn legged robot locomotion skills by watching thousands of wild animal videos from the internet, such as those featured in nature documentaries. Indeed, such videos offer a rich and diverse collection of plausible motion examples, which could inform how robots should move. To achieve this, we introduce Reinforcement Learning from Wild Animal Videos (RLWAV), a method to ground these motions into physical robots. We first train a video classifier on a large-scale animal video dataset to recognize actions from RGB clips of animals in their natural habitats. We then train a multi-skill policy to control a robot in a physics simulator, using the classification score of a third-person camera capturing videos of the robot's movements as a reward for reinforcement learning. Finally, we directly transfer the learned policy to a real quadruped Solo. Remarkably, despite the extreme gap in both domain and embodiment between animals in the wild and robots, our approach enables the policy to learn diverse skills such as walking, jumping, and keeping still, without relying on reference trajectories nor skill-specific rewards.
Traffic simulations are commonly used to optimize traffic flow, with reinforcement learning (RL) showing promising potential for automated traffic signal control. Multi-agent reinforcement learning (MARL) is particularly effective for learning control strategies for traffic lights in a network using iterative simulations. However, existing methods often assume perfect vehicle detection, which overlooks real-world limitations related to infrastructure availability and sensor reliability. This study proposes a co-simulation framework integrating CARLA and SUMO, which combines high-fidelity 3D modeling with large-scale traffic flow simulation. Cameras mounted on traffic light poles within the CARLA environment use a YOLO-based computer vision system to detect and count vehicles, providing real-time traffic data as input for adaptive signal control in SUMO. MARL agents, trained with four different reward structures, leverage this visual feedback to optimize signal timings and improve network-wide traffic flow. Experiments in the test-bed demonstrate the effectiveness of the proposed MARL approach in enhancing traffic conditions using real-time camera-based detection. The framework also evaluates the robustness of MARL under faulty or sparse sensing and compares the performance of YOLOv5 and YOLOv8 for vehicle detection. Results show that while better accuracy improves performance, MARL agents can still achieve significant improvements with imperfect detection, demonstrating adaptability for real-world scenarios.
Machine learning systems are increasingly being used in critical decision making such as healthcare, finance, and criminal justice. Concerns around their fairness have resulted in several bias mitigation techniques that emphasize the need for high-quality data to ensure fairer decisions. However, the role of earlier stages of machine learning pipelines in mitigating model bias has not been explored well. In this paper, we focus on the task of acquiring additional labeled data points for training the downstream machine learning model to rapidly improve its fairness. Since not all data points in a data pool are equally beneficial to the task of fairness, we generate an ordering in which data points should be acquired. We present DataSift, a data acquisition framework based on the idea of data valuation that relies on partitioning and multi-armed bandits to determine the most valuable data points to acquire. Over several iterations, DataSift selects a partition and randomly samples a batch of data points from the selected partition, evaluates the benefit of acquiring the batch on model fairness, and updates the utility of partitions depending on the benefit. To further improve the effectiveness and efficiency of evaluating batches, we leverage influence functions that estimate the effect of acquiring a batch without retraining the model. We empirically evaluate DataSift on several real-world and synthetic datasets and show that the fairness of a machine learning model can be significantly improved even while acquiring a few data points.
One explanation for how people can plan efficiently despite limited cognitive resources is that we possess a set of adaptive planning strategies and know when and how to use them. But how are these strategies acquired? While previous research has studied how individuals learn to choose among existing strategies, little is known about the process of forming new planning strategies. In this work, we propose that new planning strategies are discovered through metacognitive reinforcement learning. To test this, we designed a novel experiment to investigate the discovery of new planning strategies. We then present metacognitive reinforcement learning models and demonstrate their capability for strategy discovery as well as show that they provide a better explanation of human strategy discovery than alternative learning mechanisms. However, when fitted to human data, these models exhibit a slower discovery rate than humans, leaving room for improvement.
Inverse Reinforcement Learning (IRL) has demonstrated effectiveness in a variety of imitation tasks. In this paper, we introduce an IRL framework designed to extract rewarding features from expert trajectories affected by delayed disturbances. Instead of relying on direct observations, our approach employs an efficient off-policy adversarial training framework to derive expert features and recover optimal policies from augmented delayed observations. Empirical evaluations in the MuJoCo environment under diverse delay settings validate the effectiveness of our method. Furthermore, we provide a theoretical analysis showing that recovering expert policies from augmented delayed observations outperforms using direct delayed observations.
We investigate the problem of teaching a robot manipulator to perform dynamic non-prehensile object transport, also known as the `robot waiter' task, from a limited set of real-world demonstrations. We propose an approach that combines batch reinforcement learning (RL) with model-predictive control (MPC) by pretraining an ensemble of value functions from demonstration data, and utilizing them online within an uncertainty-aware MPC scheme to ensure robustness to limited data coverage. Our approach is straightforward to integrate with off-the-shelf MPC frameworks and enables learning solely from task space demonstrations with sparsely labeled transitions, while leveraging MPC to ensure smooth joint space motions and constraint satisfaction. We validate the proposed approach through extensive simulated and real-world experiments on a Franka Panda robot performing the robot waiter task and demonstrate robust deployment of value functions learned from 50-100 demonstrations. Furthermore, our approach enables generalization to novel objects not seen during training and can improve upon suboptimal demonstrations. We believe that such a framework can reduce the burden of providing extensive demonstrations and facilitate rapid training of robot manipulators to perform non-prehensile manipulation tasks. Project videos and supplementary material can be found at: https://sites.google.com/view/cvmpc.
Recently, a novel paradigm has been proposed for reinforcement learning-based NAS agents, that revolves around the incremental improvement of a given architecture. We assess the abilities of such reinforcement learning agents to transfer between different tasks. We perform our evaluation using the Trans-NASBench-101 benchmark, and consider the efficacy of the transferred agents, as well as how quickly they can be trained. We find that pretraining an agent on one task benefits the performance of the agent in another task in all but 1 task when considering final performance. We also show that the training procedure for an agent can be shortened significantly by pretraining it on another task. Our results indicate that these effects occur regardless of the source or target task, although they are more pronounced for some tasks than for others. Our results show that transfer learning can be an effective tool in mitigating the computational cost of the initial training procedure for reinforcement learning-based NAS agents.
Partial observability of the underlying states generally presents significant challenges for reinforcement learning (RL). In practice, certain \emph{privileged information}, e.g., the access to states from simulators, has been exploited in training and has achieved prominent empirical successes. To better understand the benefits of privileged information, we revisit and examine several simple and practically used paradigms in this setting. Specifically, we first formalize the empirical paradigm of \emph{expert distillation} (also known as \emph{teacher-student} learning), demonstrating its pitfall in finding near-optimal policies. We then identify a condition of the partially observable environment, the \emph{deterministic filter condition}, under which expert distillation achieves sample and computational complexities that are \emph{both} polynomial. Furthermore, we investigate another useful empirical paradigm of \emph{asymmetric actor-critic}, and focus on the more challenging setting of observable partially observable Markov decision processes. We develop a belief-weighted asymmetric actor-critic algorithm with polynomial sample and quasi-polynomial computational complexities, in which one key component is a new provable oracle for learning belief states that preserve \emph{filter stability} under a misspecified model, which may be of independent interest. Finally, we also investigate the provable efficiency of partially observable multi-agent RL (MARL) with privileged information. We develop algorithms featuring \emph{centralized-training-with-decentralized-execution}, a popular framework in empirical MARL, with polynomial sample and (quasi-)polynomial computational complexities in both paradigms above. Compared with a few recent related theoretical studies, our focus is on understanding practically inspired algorithmic paradigms, without computationally intractable oracles.
This paper proposes an online environment poisoning algorithm tailored for reinforcement learning agents operating in a black-box setting, where an adversary deliberately manipulates training data to lead the agent toward a mischievous policy. In contrast to prior studies that primarily investigate white-box settings, we focus on a scenario characterized by \textit{unknown} environment dynamics to the attacker and a \textit{flexible} reinforcement learning algorithm employed by the targeted agent. We first propose an attack scheme that is capable of poisoning the reward functions and state transitions. The poisoning task is formalized as a constrained optimization problem, following the framework of \cite{ma2019policy}. Given the transition probabilities are unknown to the attacker in a black-box environment, we apply a stochastic gradient descent algorithm, where the exact gradients are approximated using sample-based estimates. A penalty-based method along with a bilevel reformulation is then employed to transform the problem into an unconstrained counterpart and to circumvent the double-sampling issue. The algorithm's effectiveness is validated through a maze environment.
This paper presents a model-free deep reinforcement learning framework for informative path planning with heterogeneous fleets of autonomous surface vehicles to locate and collect plastic waste. The system employs two teams of vehicles: scouts and cleaners. Coordination between these teams is achieved through a deep reinforcement approach, allowing agents to learn strategies to maximize cleaning efficiency. The primary objective is for the scout team to provide an up-to-date contamination model, while the cleaner team collects as much waste as possible following this model. This strategy leads to heterogeneous teams that optimize fleet efficiency through inter-team cooperation supported by a tailored reward function. Different trainings of the proposed algorithm are compared with other state-of-the-art heuristics in two distinct scenarios, one with high convexity and another with narrow corridors and challenging access. According to the obtained results, it is demonstrated that deep reinforcement learning based algorithms outperform other benchmark heuristics, exhibiting superior adaptability. In addition, training with greedy actions further enhances performance, particularly in scenarios with intricate layouts.
Generative models, particularly diffusion models, have achieved remarkable success in density estimation for multimodal data, drawing significant interest from the reinforcement learning (RL) community, especially in policy modeling in continuous action spaces. However, existing works exhibit significant variations in training schemes and RL optimization objectives, and some methods are only applicable to diffusion models. In this study, we compare and analyze various generative policy training and deployment techniques, identifying and validating effective designs for generative policy algorithms. Specifically, we revisit existing training objectives and classify them into two categories, each linked to a simpler approach. The first approach, Generative Model Policy Optimization (GMPO), employs a native advantage-weighted regression formulation as the training objective, which is significantly simpler than previous methods. The second approach, Generative Model Policy Gradient (GMPG), offers a numerically stable implementation of the native policy gradient method. We introduce a standardized experimental framework named GenerativeRL. Our experiments demonstrate that the proposed methods achieve state-of-the-art performance on various offline-RL datasets, offering a unified and practical guideline for training and deploying generative policies.
Efforts to interpret reinforcement learning (RL) models often rely on high-level techniques such as attribution or probing, which provide only correlational insights and coarse causal control. This work proposes replacing nonlinearities in convolutional neural networks (ConvNets) with bilinear variants, to produce a class of models for which these limitations can be addressed. We show bilinear model variants perform comparably in model-free reinforcement learning settings, and give a side by side comparison on ProcGen environments. Bilinear layers' analytic structure enables weight-based decomposition. Previous work has shown bilinearity enables quantifying functional importance through eigendecomposition, to identify interpretable low rank structure. We show how to adapt the decomposition to convolution layers by applying singular value decomposition to vectors of interest, to separate the channel and spatial dimensions. Finally, we propose a methodology for causally validating concept-based probes, and illustrate its utility by studying a maze-solving agent's ability to track a cheese object.
Designing efficient algorithms for multi-agent reinforcement learning (MARL) is fundamentally challenging because the size of the joint state and action spaces grows exponentially in the number of agents. These difficulties are exacerbated when balancing sequential global decision-making with local agent interactions. In this work, we propose a new algorithm $\texttt{SUBSAMPLE-MFQ}$ ($\textbf{Subsample}$-$\textbf{M}$ean-$\textbf{F}$ield-$\textbf{Q}$-learning) and a decentralized randomized policy for a system with $n$ agents. For any $k\leq n$, our algorithm learns a policy for the system in time polynomial in $k$. We prove that this learned policy converges to the optimal policy on the order of $\tilde{O}(1/\sqrt{k})$ as the number of subsampled agents $k$ increases. In particular, this bound is independent of the number of agents $n$.