Deep reinforcement learning (DRL) has demonstrated remarkable performance in many continuous control tasks. However, a significant obstacle to the real-world application of DRL is the lack of safety guarantees. Although DRL agents can satisfy system safety in expectation through reward shaping, designing agents to consistently meet hard constraints (e.g., safety specifications) at every time step remains a formidable challenge. In contrast, existing work in the field of safe control provides guarantees on persistent satisfaction of hard safety constraints. However, these methods require explicit analytical system dynamics models to synthesize safe control, which are typically inaccessible in DRL settings. In this paper, we present a model-free safe control algorithm, the implicit safe set algorithm, for synthesizing safeguards for DRL agents that ensure provable safety throughout training. The proposed algorithm synthesizes a safety index (barrier certificate) and a subsequent safe control law solely by querying a black-box dynamic function (e.g., a digital twin simulator). Moreover, we theoretically prove that the implicit safe set algorithm guarantees finite time convergence to the safe set and forward invariance for both continuous-time and discrete-time systems. We validate the proposed algorithm on the state-of-the-art Safety Gym benchmark, where it achieves zero safety violations while gaining $95\% \pm 9\%$ cumulative reward compared to state-of-the-art safe DRL methods. Furthermore, the resulting algorithm scales well to high-dimensional systems with parallel computing.
While originally developed for continuous control problems, Proximal Policy Optimization (PPO) has emerged as the work-horse of a variety of reinforcement learning (RL) applications, including the fine-tuning of generative models. Unfortunately, PPO requires multiple heuristics to enable stable convergence (e.g. value networks, clipping), and is notorious for its sensitivity to the precise implementation of these components. In response, we take a step back and ask what a minimalist RL algorithm for the era of generative models would look like. We propose REBEL, an algorithm that cleanly reduces the problem of policy optimization to regressing the relative reward between two completions to a prompt in terms of the policy, enabling strikingly lightweight implementation. In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL, which allows us to match the strongest known theoretical guarantees in terms of convergence and sample complexity in the RL literature. REBEL can also cleanly incorporate offline data and be extended to handle the intransitive preferences we frequently see in practice. Empirically, we find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO, all while being simpler to implement and more computationally efficient than PPO. When fine-tuning Llama-3-8B-Instruct, REBEL achieves strong performance in AlpacaEval 2.0, MT-Bench, and Open LLM Leaderboard.
This paper presents AFU, an off-policy deep RL algorithm addressing in a new way the challenging "max-Q problem" in Q-learning for continuous action spaces, with a solution based on regression and conditional gradient scaling. AFU has an actor but its critic updates are entirely independent from it. As a consequence, the actor can be chosen freely. In the initial version, AFU-alpha, we employ the same stochastic actor as in Soft Actor-Critic (SAC), but we then study a simple failure mode of SAC and show how AFU can be modified to make actor updates less likely to become trapped in local optima, resulting in a second version of the algorithm, AFU-beta. Experimental results demonstrate the sample efficiency of both versions of AFU, marking it as the first model-free off-policy algorithm competitive with state-of-the-art actor-critic methods while departing from the actor-critic perspective.
We address the crucial yet underexplored stability properties of the Hamilton--Jacobi--Bellman (HJB) equation in model-free reinforcement learning contexts, specifically for Lipschitz continuous optimal control problems. We bridge the gap between Lipschitz continuous optimal control problems and classical optimal control problems in the viscosity solutions framework, offering new insights into the stability of the value function of Lipschitz continuous optimal control problems. By introducing structural assumptions on the dynamics and reward functions, we further study the rate of convergence of value functions. Moreover, we introduce a generalized framework for Lipschitz continuous control problems that incorporates the original problem and leverage it to propose a new HJB-based reinforcement learning algorithm. The stability properties and performance of the proposed method are tested with well-known benchmark examples in comparison with existing approaches.
In robotic control tasks, policies trained by reinforcement learning (RL) in simulation often experience a performance drop when deployed on physical hardware, due to modeling error, measurement error, and unpredictable perturbations in the real world. Robust RL methods account for this issue by approximating a worst-case value function during training, but they can be sensitive to approximation errors in the value function and its gradient before training is complete. In this paper, we hypothesize that Lipschitz regularization can help condition the approximated value function gradients, leading to improved robustness after training. We test this hypothesis by combining Lipschitz regularization with an application of Fast Gradient Sign Method to reduce approximation errors when evaluating the value function under adversarial perturbations. Our empirical results demonstrate the benefits of this approach over prior work on a number of continuous control benchmarks.
Representation rank is an important concept for understanding the role of Neural Networks (NNs) in Deep Reinforcement learning (DRL), which measures the expressive capacity of value networks. Existing studies focus on unboundedly maximizing this rank; nevertheless, that approach would introduce overly complex models in the learning, thus undermining performance. Hence, fine-tuning representation rank presents a challenging and crucial optimization problem. To address this issue, we find a guiding principle for adaptive control of the representation rank. We employ the Bellman equation as a theoretical foundation and derive an upper bound on the cosine similarity of consecutive state-action pairs representations of value networks. We then leverage this upper bound to propose a novel regularizer, namely BEllman Equation-based automatic rank Regularizer (BEER). This regularizer adaptively regularizes the representation rank, thus improving the DRL agent's performance. We first validate the effectiveness of automatic control of rank on illustrative experiments. Then, we scale up BEER to complex continuous control tasks by combining it with the deterministic policy gradient method. Among 12 challenging DeepMind control tasks, BEER outperforms the baselines by a large margin. Besides, BEER demonstrates significant advantages in Q-value approximation. Our code is available at https://github.com/sweetice/BEER-ICLR2024.
Linear Temporal Logic (LTL) offers a precise means for constraining the behavior of reinforcement learning agents. However, in many settings where both satisfaction and optimality conditions are present, LTL is insufficient to capture both. Instead, LTL-constrained policy optimization, where the goal is to optimize a scalar reward under LTL constraints, is needed. This constrained optimization problem proves difficult in deep Reinforcement Learning (DRL) settings, where learned policies often ignore the LTL constraint due to the sparse nature of LTL satisfaction. To alleviate the sparsity issue, we introduce Cycle Experience Replay (CyclER), a novel reward shaping technique that exploits the underlying structure of the LTL constraint to guide a policy towards satisfaction by encouraging partial behaviors compliant with the constraint. We provide a theoretical guarantee that optimizing CyclER will achieve policies that satisfy the LTL constraint with near-optimal probability. We evaluate CyclER in three continuous control domains. Our experimental results show that optimizing CyclER in tandem with the existing scalar reward outperforms existing reward-shaping methods at finding performant LTL-satisfying policies.
Distributed Distributional DrQ is a model-free and off-policy RL algorithm for continuous control tasks based on the state and observation of the agent, which is an actor-critic method with the data-augmentation and the distributional perspective of critic value function. Aim to learn to control the agent and master some tasks in a high-dimensional continuous space. DrQ-v2 uses DDPG as the backbone and achieves out-performance in various continuous control tasks. Here Distributed Distributional DrQ uses Distributed Distributional DDPG as the backbone, and this modification aims to achieve better performance in some hard continuous control tasks through the better expression ability of distributional value function and distributed actor policies.
Structured state space sequence (S4) models have recently achieved state-of-the-art performance on long-range sequence modeling tasks. These models also have fast inference speeds and parallelisable training, making them potentially useful in many reinforcement learning settings. We propose a modification to a variant of S4 that enables us to initialise and reset the hidden state in parallel, allowing us to tackle reinforcement learning tasks. We show that our modified architecture runs asymptotically faster than Transformers in sequence length and performs better than RNN's on a simple memory-based task. We evaluate our modified architecture on a set of partially-observable environments and find that, in practice, our model outperforms RNN's while also running over five times faster. Then, by leveraging the model's ability to handle long-range sequences, we achieve strong performance on a challenging meta-learning task in which the agent is given a randomly-sampled continuous control environment, combined with a randomly-sampled linear projection of the environment's observations and actions. Furthermore, we show the resulting model can adapt to out-of-distribution held-out tasks. Overall, the results presented in this paper show that structured state space models are fast and performant for in-context reinforcement learning tasks. We provide code at https://github.com/luchris429/popjaxrl.
Deep reinforcement learning agents for continuous control are known to exhibit significant instability in their performance over time. In this work, we provide a fresh perspective on these behaviors by studying the return landscape: the mapping between a policy and a return. We find that popular algorithms traverse noisy neighborhoods of this landscape, in which a single update to the policy parameters leads to a wide range of returns. By taking a distributional view of these returns, we map the landscape, characterizing failure-prone regions of policy space and revealing a hidden dimension of policy quality. We show that the landscape exhibits surprising structure by finding simple paths in parameter space which improve the stability of a policy. To conclude, we develop a distribution-aware procedure which finds such paths, navigating away from noisy neighborhoods in order to improve the robustness of a policy. Taken together, our results provide new insight into the optimization, evaluation, and design of agents.
Neural Cellular Automata (NCA) is a class of Cellular Automata where the update rule is parameterized by a neural network that can be trained using gradient descent. In this paper, we focus on NCA models used for texture synthesis, where the update rule is inspired by partial differential equations (PDEs) describing reaction-diffusion systems. To train the NCA model, the spatio-temporal domain is discretized, and Euler integration is used to numerically simulate the PDE. However, whether a trained NCA truly learns the continuous dynamic described by the corresponding PDE or merely overfits the discretization used in training remains an open question. We study NCA models at the limit where space-time discretization approaches continuity. We find that existing NCA models tend to overfit the training discretization, especially in the proximity of the initial condition, also called "seed". To address this, we propose a solution that utilizes uniform noise as the initial condition. We demonstrate the effectiveness of our approach in preserving the consistency of NCA dynamics across a wide range of spatio-temporal granularities. Our improved NCA model enables two new test-time interactions by allowing continuous control over the speed of pattern formation and the scale of the synthesized patterns. We demonstrate this new NCA feature in our interactive online demo. Our work reveals that NCA models can learn continuous dynamics and opens new venues for NCA research from a dynamical system's perspective.
Recent reinforcement learning approaches have shown surprisingly strong capabilities of bang-bang policies for solving continuous control benchmarks. The underlying coarse action space discretizations often yield favourable exploration characteristics while final performance does not visibly suffer in the absence of action penalization in line with optimal control theory. In robotics applications, smooth control signals are commonly preferred to reduce system wear and energy efficiency, but action costs can be detrimental to exploration during early training. In this work, we aim to bridge this performance gap by growing discrete action spaces from coarse to fine control resolution, taking advantage of recent results in decoupled Q-learning to scale our approach to high-dimensional action spaces up to dim(A) = 38. Our work indicates that an adaptive control resolution in combination with value decomposition yields simple critic-only algorithms that yield surprisingly strong performance on continuous control tasks.
Closed-loop control of nonlinear dynamical systems with partial-state observability demands expert knowledge of a diverse, less standardized set of theoretical tools. Moreover, it requires a delicate integration of controller and estimator designs to achieve the desired system behavior. To establish a general controller synthesis framework, we explore the Decision Transformer (DT) architecture. Specifically, we first frame the control task as predicting the current optimal action based on past observations, actions, and rewards, eliminating the need for a separate estimator design. Then, we leverage the pre-trained language models, i.e., the Generative Pre-trained Transformer (GPT) series, to initialize DT and subsequently train it for control tasks using low-rank adaptation (LoRA). Our comprehensive experiments across five distinct control tasks, ranging from maneuvering aerospace systems to controlling partial differential equations (PDEs), demonstrate DT's capability to capture the parameter-agnostic structures intrinsic to control tasks. DT exhibits remarkable zero-shot generalization abilities for completely new tasks and rapidly surpasses expert performance levels with a minimal amount of demonstration data. These findings highlight the potential of DT as a foundational controller for general control applications.
Recent advancements in deep reinforcement learning (RL) have demonstrated notable progress in sample efficiency, spanning both model-based and model-free paradigms. Despite the identification and mitigation of specific bottlenecks in prior works, the agent's exploration ability remains under-emphasized in the realm of sample-efficient RL. This paper investigates how to achieve sample-efficient exploration in continuous control tasks. We introduce an RL algorithm that incorporates a predictive model and off-policy learning elements, where an online planner enhanced by a novelty-aware terminal value function is employed for sample collection. Leveraging the forward predictive error within a latent state space, we derive an intrinsic reward without incurring parameters overhead. This reward establishes a solid connection to model uncertainty, allowing the agent to effectively overcome the asymptotic performance gap. Through extensive experiments, our method shows competitive or even superior performance compared to prior works, especially the sparse reward cases.
In standard reinforcement learning settings, agents typically assume immediate feedback about the effects of their actions after taking them. However, in practice, this assumption may not hold true due to physical constraints and can significantly impact the performance of learning algorithms. In this paper, we address observation delays in partially observable environments. We propose leveraging world models, which have shown success in integrating past observations and learning dynamics, to handle observation delays. By reducing delayed POMDPs to delayed MDPs with world models, our methods can effectively handle partial observability, where existing approaches achieve sub-optimal performance or degrade quickly as observability decreases. Experiments suggest that one of our methods can outperform a naive model-based approach by up to 250%. Moreover, we evaluate our methods on visual delayed environments, for the first time showcasing delay-aware reinforcement learning continuous control with visual observations.
With the advent of universal function approximators in the domain of reinforcement learning, the number of practical applications leveraging deep reinforcement learning (DRL) has exploded. Decision-making in autonomous vehicles (AVs) has emerged as a chief application among them, taking the sensor data or the higher-order kinematic variables as the input and providing a discrete choice or continuous control output. There has been a continuous effort to understand the black-box nature of the DRL models, but so far, there hasn't been any discussion (to the best of authors' knowledge) about how the models learn the physical process. This presents an overwhelming limitation that restricts the real-world deployment of DRL in AVs. Therefore, in this research work, we try to decode the knowledge learnt by the attention-based DRL framework about the physical process. We use a continuous proximal policy optimization-based DRL algorithm as the baseline model and add a multi-head attention framework in an open-source AV simulation environment. We provide some analytical techniques for discussing the interpretability of the trained models in terms of explainability and causality for spatial and temporal correlations. We show that the weights in the first head encode the positions of the neighboring vehicles while the second head focuses on the leader vehicle exclusively. Also, the ego vehicle's action is causally dependent on the vehicles in the target lane spatially and temporally. Through these findings, we reliably show that these techniques can help practitioners decipher the results of the DRL algorithms.
In preference-based reinforcement learning (PbRL), a reward function is learned from a type of human feedback called preference. To expedite preference collection, recent works have leveraged \emph{offline preferences}, which are preferences collected for some offline data. In this scenario, the learned reward function is fitted on the offline data. If a learning agent exhibits behaviors that do not overlap with the offline data, the learned reward function may encounter generalizability issues. To address this problem, the present study introduces a framework that consolidates offline preferences and \emph{virtual preferences} for PbRL, which are comparisons between the agent's behaviors and the offline data. Critically, the reward function can track the agent's behaviors using the virtual preferences, thereby offering well-aligned guidance to the agent. Through experiments on continuous control tasks, this study demonstrates the effectiveness of incorporating the virtual preferences in PbRL.
A key aspect of intelligence is the ability to demonstrate a broad spectrum of behaviors for adapting to unexpected situations. Over the past decade, advancements in deep reinforcement learning have led to groundbreaking achievements to solve complex continuous control tasks. However, most approaches return only one solution specialized for a specific problem. We introduce Quality-Diversity Actor-Critic (QDAC), an off-policy actor-critic deep reinforcement learning algorithm that leverages a value function critic and a successor features critic to learn high-performing and diverse behaviors. In this framework, the actor optimizes an objective that seamlessly unifies both critics using constrained optimization to (1) maximize return, while (2) executing diverse skills. Compared with other Quality-Diversity methods, QDAC achieves significantly higher performance and more diverse behaviors on six challenging continuous control locomotion tasks. We also demonstrate that we can harness the learned skills to adapt better than other baselines to five perturbed environments. Finally, qualitative analyses showcase a range of remarkable behaviors: adaptive-intelligent-robotics.github.io/QDAC.
In this paper, we study the problem of ensuring safety with a few shots of samples for partially unknown systems. We first characterize a fundamental limit when producing safe actions is not possible due to insufficient information or samples. Then, we develop a technique that can generate provably safe actions and recovery behaviors using a minimum number of samples. In the performance analysis, we also establish Nagumos theorem - like results with relaxed assumptions, which is potentially useful in other contexts. Finally, we discuss how the proposed method can be integrated into a policy gradient algorithm to assure safety and stability with a handful of samples without stabilizing initial policies or generative models to probe safe actions.
In deep reinforcement learning, estimating the value function to evaluate the quality of states and actions is essential. The value function is often trained using the least squares method, which implicitly assumes a Gaussian error distribution. However, a recent study suggested that the error distribution for training the value function is often skewed because of the properties of the Bellman operator, and violates the implicit assumption of normal error distribution in the least squares method. To address this, we proposed a method called Symmetric Q-learning, in which the synthetic noise generated from a zero-mean distribution is added to the target values to generate a Gaussian error distribution. We evaluated the proposed method on continuous control benchmark tasks in MuJoCo. It improved the sample efficiency of a state-of-the-art reinforcement learning method by reducing the skewness of the error distribution.