Value approximation using deep neural networks is at the heart of off-policy deep reinforcement learning, and is often the primary module that provides learning signals to the rest of the algorithm. While multi-layer perceptron networks are universal function approximators, recent works in neural kernel regression suggest the presence of a spectral bias, where fitting high-frequency components of the value function requires exponentially more gradient update steps than the low-frequency ones. In this work, we re-examine off-policy reinforcement learning through the lens of kernel regression and propose to overcome such bias via a composite neural tangent kernel. With just a single line-change, our approach, the Fourier feature networks (FFN) produce state-of-the-art performance on challenging continuous control domains with only a fraction of the compute. Faster convergence and better off-policy stability also make it possible to remove the target network without suffering catastrophic divergences, which further reduces TD}(0)'s estimation bias on a few tasks.
Though deep reinforcement learning (DRL) has obtained substantial success, it may encounter catastrophic failures due to the intrinsic uncertainty of both transition and observation. Most of the existing methods for safe reinforcement learning can only handle transition disturbance or observation disturbance since these two kinds of disturbance affect different parts of the agent; besides, the popular worst-case return may lead to overly pessimistic policies. To address these issues, we first theoretically prove that the performance degradation under transition disturbance and observation disturbance depends on a novel metric of Value Function Range (VFR), which corresponds to the gap in the value function between the best state and the worst state. Based on the analysis, we adopt conditional value-at-risk (CVaR) as an assessment of risk and propose a novel reinforcement learning algorithm of CVaR-Proximal-Policy-Optimization (CPPO) which formalizes the risk-sensitive constrained optimization problem by keeping its CVaR under a given threshold. Experimental results show that CPPO achieves a higher cumulative reward and is more robust against both observation and transition disturbances on a series of continuous control tasks in MuJoCo.
Many Deep Reinforcement Learning (D-RL) algorithms rely on simple forms of exploration such as the additive action noise often used in continuous control domains. Typically, the scaling factor of this action noise is chosen as a hyper-parameter and is kept constant during training. In this paper, we focus on action noise in off-policy deep reinforcement learning for continuous control. We analyze how the learned policy is impacted by the noise type, noise scale, and impact scaling factor reduction schedule. We consider the two most prominent types of action noise, Gaussian and Ornstein-Uhlenbeck noise, and perform a vast experimental campaign by systematically varying the noise type and scale parameter, and by measuring variables of interest like the expected return of the policy and the state-space coverage during exploration. For the latter, we propose a novel state-space coverage measure $\operatorname{X}_{\mathcal{U}\text{rel}}$ that is more robust to estimation artifacts caused by points close to the state-space boundary than previously-proposed measures. Larger noise scales generally increase state-space coverage. However, we found that increasing the space coverage using a larger noise scale is often not beneficial. On the contrary, reducing the noise scale over the training process reduces the variance and generally improves the learning performance. We conclude that the best noise type and scale are environment dependent, and based on our observations derive heuristic rules for guiding the choice of the action noise as a starting point for further optimization.
Adversarial Imitation Learning (AIL) is a class of popular state-of-the-art Imitation Learning algorithms commonly used in robotics. In AIL, an artificial adversary's misclassification is used as a reward signal that is optimized by any standard Reinforcement Learning (RL) algorithm. Unlike most RL settings, the reward in AIL is $differentiable$ but current model-free RL algorithms do not make use of this property to train a policy. The reward is AIL is also shaped since it comes from an adversary. We leverage the differentiability property of the shaped AIL reward function and formulate a class of Actor Residual Critic (ARC) RL algorithms. ARC algorithms draw a parallel to the standard Actor-Critic (AC) algorithms in RL literature and uses a residual critic, $C$ function (instead of the standard $Q$ function) to approximate only the discounted future return (excluding the immediate reward). ARC algorithms have similar convergence properties as the standard AC algorithms with the additional advantage that the gradient through the immediate reward is exact. For the discrete (tabular) case with finite states, actions, and known dynamics, we prove that policy iteration with $C$ function converges to an optimal policy. In the continuous case with function approximation and unknown dynamics, we experimentally show that ARC aided AIL outperforms standard AIL in simulated continuous-control and real robotic manipulation tasks. ARC algorithms are simple to implement and can be incorporated into any existing AIL implementation with an AC algorithm. Video and link to code are available at: https://sites.google.com/view/actor-residual-critic.
Model-based approaches to reinforcement learning (MBRL) exhibit favorable performance in practice, but their theoretical guarantees in large spaces are mostly restricted to the setting when transition model is Gaussian or Lipschitz, and demands a posterior estimate whose representational complexity grows unbounded with time. In this work, we develop a novel MBRL method (i) which relaxes the assumptions on the target transition model to belong to a generic family of mixture models; (ii) is applicable to large-scale training by incorporating a compression step such that the posterior estimate consists of a Bayesian coreset of only statistically significant past state-action pairs; and (iii) exhibits a sublinear Bayesian regret. To achieve these results, we adopt an approach based upon Stein's method, which, under a smoothness condition on the constructed posterior and target, allows distributional distance to be evaluated in closed form as the kernelized Stein discrepancy (KSD). The aforementioned compression step is then computed in terms of greedily retaining only those samples which are more than a certain KSD away from the previous model estimate. Experimentally, we observe that this approach is competitive with several state-of-the-art RL methodologies, and can achieve up-to 50 percent reduction in wall clock time in some continuous control environments.
Online imitation learning is the problem of how best to mimic expert demonstrations, given access to the environment or an accurate simulator. Prior work has shown that in the infinite sample regime, exact moment matching achieves value equivalence to the expert policy. However, in the finite sample regime, even if one has no optimization error, empirical variance can lead to a performance gap that scales with $H^2 / N$ for behavioral cloning and $H / \sqrt{N}$ for online moment matching, where $H$ is the horizon and $N$ is the size of the expert dataset. We introduce the technique of replay estimation to reduce this empirical variance: by repeatedly executing cached expert actions in a stochastic simulator, we compute a smoother expert visitation distribution estimate to match. In the presence of general function approximation, we prove a meta theorem reducing the performance gap of our approach to the parameter estimation error for offline classification (i.e. learning the expert policy). In the tabular setting or with linear function approximation, our meta theorem shows that the performance gap incurred by our approach achieves the optimal $\widetilde{O} \left( \min({H^{3/2}} / {N}, {H} / {\sqrt{N}} \right)$ dependency, under significantly weaker assumptions compared to prior work. We implement multiple instantiations of our approach on several continuous control tasks and find that we are able to significantly improve policy performance across a variety of dataset sizes.
Continual learning (CL) enables the development of models and agents that learn from a sequence of tasks while addressing the limitations of standard deep learning approaches, such as catastrophic forgetting. In this work, we investigate the factors that contribute to the performance differences between task-agnostic CL and multi-task (MTL) agents. We pose two hypotheses: (1) task-agnostic methods might provide advantages in settings with limited data, computation, or high dimensionality, and (2) faster adaptation may be particularly beneficial in continual learning settings, helping to mitigate the effects of catastrophic forgetting. To investigate these hypotheses, we introduce a replay-based recurrent reinforcement learning (3RL) methodology for task-agnostic CL agents. We assess 3RL on a synthetic task and the Meta-World benchmark, which includes 50 unique manipulation tasks. Our results demonstrate that 3RL outperforms baseline methods and can even surpass its multi-task equivalent in challenging settings with high dimensionality. We also show that the recurrent task-agnostic agent consistently outperforms or matches the performance of its transformer-based counterpart. These findings provide insights into the advantages of task-agnostic CL over task-aware MTL approaches and highlight the potential of task-agnostic methods in resource-constrained, high-dimensional, and multi-task environments.
Training deep reinforcement learning (DRL) models usually requires high computation costs. Therefore, compressing DRL models possesses immense potential for training acceleration and model deployment. However, existing methods that generate small models mainly adopt the knowledge distillation-based approach by iteratively training a dense network. As a result, the training process still demands massive computing resources. Indeed, sparse training from scratch in DRL has not been well explored and is particularly challenging due to non-stationarity in bootstrap training. In this work, we propose a novel sparse DRL training framework, "the Rigged Reinforcement Learning Lottery" (RLx2), which builds upon gradient-based topology evolution and is capable of training a sparse DRL model based entirely on a sparse network. Specifically, RLx2 introduces a novel multi-step TD target mechanism with a dynamic-capacity replay buffer to achieve robust value learning and efficient topology exploration in sparse models. It also reaches state-of-the-art sparse training performance in several tasks, showing 7.5\times-20\times model compression with less than 3% performance degradation and up to 20\times and 50\times FLOPs reduction for training and inference, respectively.
We propose Taylor Series Imitation Learning (TaSIL), a simple augmentation to standard behavior cloning losses in the context of continuous control. TaSIL penalizes deviations in the higher-order Taylor series terms between the learned and expert policies. We show that experts satisfying a notion of $\textit{incremental input-to-state stability}$ are easy to learn, in the sense that a small TaSIL-augmented imitation loss over expert trajectories guarantees a small imitation loss over trajectories generated by the learned policy. We provide sample-complexity bounds for TaSIL that scale as $\tilde{\mathcal{O}}(1/n)$ in the realizable setting, for $n$ the number of expert demonstrations. Finally, we demonstrate experimentally the relationship between the robustness of the expert policy and the order of Taylor expansion required in TaSIL, and compare standard Behavior Cloning, DART, and DAgger with TaSIL-loss-augmented variants. In all cases, we show significant improvement over baselines across a variety of MuJoCo tasks.
Deep reinforcement learning (DRL) gives the promise that an agent learns good policy from high-dimensional information, whereas representation learning removes irrelevant and redundant information and retains pertinent information. In this work, we demonstrate that the learned representation of the $Q$-network and its target $Q$-network should, in theory, satisfy a favorable distinguishable representation property. Specifically, there exists an upper bound on the representation similarity of the value functions of two adjacent time steps in a typical DRL setting. However, through illustrative experiments, we show that the learned DRL agent may violate this property and lead to a sub-optimal policy. Therefore, we propose a simple yet effective regularizer called Policy Evaluation with Easy Regularization on Representation (PEER), which aims to maintain the distinguishable representation property via explicit regularization on internal representations. And we provide the convergence rate guarantee of PEER. Implementing PEER requires only one line of code. Our experiments demonstrate that incorporating PEER into DRL can significantly improve performance and sample efficiency. Comprehensive experiments show that PEER achieves state-of-the-art performance on all 4 environments on PyBullet, 9 out of 12 tasks on DMControl, and 19 out of 26 games on Atari. To the best of our knowledge, PEER is the first work to study the inherent representation property of Q-network and its target. Our code is available at https://sites.google.com/view/peer-cvpr2023/.
A crucial challenge in reinforcement learning is to reduce the number of interactions with the environment that an agent requires to master a given task. Transfer learning proposes to address this issue by re-using knowledge from previously learned tasks. However, determining which source task qualifies as the most appropriate for knowledge extraction, as well as the choice regarding which algorithm components to transfer, represent severe obstacles to its application in reinforcement learning. The goal of this paper is to address these issues with modular multi-source transfer learning techniques. The proposed techniques automatically learn how to extract useful information from source tasks, regardless of the difference in state-action space and reward function. We support our claims with extensive and challenging cross-domain experiments for visual control.
Embodied agents in continuous control domains have had limited exposure to tasks allowing to explore musculoskeletal properties that enable agile and nimble behaviors in biological beings. The sophistication behind neuro-musculoskeletal control can pose new challenges for the motor learning community. At the same time, agents solving complex neural control problems allow impact in fields such as neuro-rehabilitation, as well as collaborative-robotics. Human biomechanics underlies complex multi-joint-multi-actuator musculoskeletal systems. The sensory-motor system relies on a range of sensory-contact rich and proprioceptive inputs that define and condition muscle actuation required to exhibit intelligent behaviors in the physical world. Current frameworks for musculoskeletal control do not support physiological sophistication of the musculoskeletal systems along with physical world interaction capabilities. In addition, they are neither embedded in complex and skillful motor tasks nor are computationally effective and scalable to study large-scale learning paradigms. Here, we present MyoSuite -- a suite of physiologically accurate biomechanical models of elbow, wrist, and hand, with physical contact capabilities, which allow learning of complex and skillful contact-rich real-world tasks. We provide diverse motor-control challenges: from simple postural control to skilled hand-object interactions such as turning a key, twirling a pen, rotating two balls in one hand, etc. By supporting physiological alterations in musculoskeletal geometry (tendon transfer), assistive devices (exoskeleton assistance), and muscle contraction dynamics (muscle fatigue, sarcopenia), we present real-life tasks with temporal changes, thereby exposing realistic non-stationary conditions in our tasks which most continuous control benchmarks lack.
Efficient exploration is a crucial challenge in deep reinforcement learning. Several methods, such as behavioral priors, are able to leverage offline data in order to efficiently accelerate reinforcement learning on complex tasks. However, if the task at hand deviates excessively from the demonstrated task, the effectiveness of such methods is limited. In our work, we propose to learn features from offline data that are shared by a more diverse range of tasks, such as correlation between actions and directedness. Therefore, we introduce state-free priors, which directly model temporal consistency in demonstrated trajectories, and are capable of driving exploration in complex tasks, even when trained on data collected on simpler tasks. Furthermore, we introduce a novel integration scheme for action priors in off-policy reinforcement learning by dynamically sampling actions from a probabilistic mixture of policy and action prior. We compare our approach against strong baselines and provide empirical evidence that it can accelerate reinforcement learning in long-horizon continuous control tasks under sparse reward settings.
It is desirable for an agent to be able to solve a rich variety of problems that can be specified through language in the same environment. A popular approach towards obtaining such agents is to reuse skills learned in prior tasks to generalise compositionally to new ones. However, this is a challenging problem due to the curse of dimensionality induced by the combinatorially large number of ways high-level goals can be combined both logically and temporally in language. To address this problem, we propose a framework where an agent first learns a sufficient set of skill primitives to achieve all high-level goals in its environment. The agent can then flexibly compose them both logically and temporally to provably achieve temporal logic specifications in any regular language, such as regular fragments of linear temporal logic. This provides the agent with the ability to map from complex temporal logic task specifications to near-optimal behaviours zero-shot. We demonstrate this experimentally in a tabular setting, as well as in a high-dimensional video game and continuous control environment. Finally, we also demonstrate that the performance of skill machines can be improved with regular off-policy reinforcement learning algorithms when optimal behaviours are desired.
The current work on reinforcement learning (RL) from demonstrations often assumes the demonstrations are samples from an optimal policy, an unrealistic assumption in practice. When demonstrations are generated by sub-optimal policies or have sparse state-action pairs, policy learned from sub-optimal demonstrations may mislead an agent with incorrect or non-local action decisions. We propose a new method called Local Ensemble and Reparameterization with Split and Merge of expert policies (LEARN-SAM) to improve efficiency and make better use of the sub-optimal demonstrations. First, LEARN-SAM employs a new concept, the lambda-function, based on a discrepancy measure between the current state to demonstrated states to "localize" the weights of the expert policies during learning. Second, LEARN-SAM employs a split-and-merge (SAM) mechanism by separating the helpful parts in each expert demonstration and regrouping them into new expert policies to use the demonstrations selectively. Both the lambda-function and SAM mechanism help boost the learning speed. Theoretically, we prove the invariant property of reparameterized policy before and after the SAM mechanism, providing theoretical guarantees for the convergence of the employed policy gradient method. We demonstrate the superiority of the LEARN-SAM method and its robustness with varying demonstration quality and sparsity in six experiments on complex continuous control problems of low to high dimensions, compared to existing methods on RL from demonstration.
Experience replay plays a crucial role in improving the sample efficiency of deep reinforcement learning agents. Recent advances in experience replay propose using Mixup (Zhang et al., 2018) to further improve sample efficiency via synthetic sample generation. We build upon this technique with Neighborhood Mixup Experience Replay (NMER), a geometrically-grounded replay buffer that interpolates transitions with their closest neighbors in state-action space. NMER preserves a locally linear approximation of the transition manifold by only applying Mixup between transitions with vicinal state-action features. Under NMER, a given transition's set of state action neighbors is dynamic and episode agnostic, in turn encouraging greater policy generalizability via inter-episode interpolation. We combine our approach with recent off-policy deep reinforcement learning algorithms and evaluate on continuous control environments. We observe that NMER improves sample efficiency by an average 94% (TD3) and 29% (SAC) over baseline replay buffers, enabling agents to effectively recombine previous experiences and learn from limited data.
We present an algorithm for Inverse Reinforcement Learning (IRL) from expert state observations only. Our approach decouples reward modelling from policy learning, unlike state-of-the-art adversarial methods which require updating the reward model during policy search and are known to be unstable and difficult to optimize. Our method, IL-flOw, recovers the expert policy by modelling state-state transitions, by generating rewards using deep density estimators trained on the demonstration trajectories, avoiding the instability issues of adversarial methods. We demonstrate that using the state transition log-probability density as a reward signal for forward reinforcement learning translates to matching the trajectory distribution of the expert demonstrations, and experimentally show good recovery of the true reward signal as well as state of the art results for imitation from observation on locomotion and robotic continuous control tasks.
This study focuses on two important problems related to applying offline model-based optimization to real-world industrial control problems. The first problem is how to create a reliable probabilistic model that accurately captures the dynamics present in noisy industrial data. The second problem is how to reliably optimize control parameters without actively collecting feedback from industrial systems. Specifically, we introduce a novel cGAN ensemble-based uncertainty-aware surrogate model for reliable offline model-based optimization in industrial control problems. The effectiveness of the proposed method is demonstrated through extensive experiments conducted on two representative cases, namely a discrete control case and a continuous control case. The results of these experiments show that our method outperforms several competitive baselines in the field of offline model-based optimization for industrial control.
While reinforcement learning (RL) provides a framework for learning through trial and error, translating RL algorithms into the real world has remained challenging. A major hurdle to real-world application arises from the development of algorithms in an episodic setting where the environment is reset after every trial, in contrast with the continual and non-episodic nature of the real-world encountered by embodied agents such as humans and robots. Prior works have considered an alternating approach where a forward policy learns to solve the task and the backward policy learns to reset the environment, but what initial state distribution should the backward policy reset the agent to? Assuming access to a few demonstrations, we propose a new method, MEDAL, that trains the backward policy to match the state distribution in the provided demonstrations. This keeps the agent close to the task-relevant states, allowing for a mix of easy and difficult starting states for the forward policy. Our experiments show that MEDAL matches or outperforms prior methods on three sparse-reward continuous control tasks from the EARL benchmark, with 40% gains on the hardest task, while making fewer assumptions than prior works.
Actor-critic Reinforcement Learning (RL) algorithms have achieved impressive performance in continuous control tasks. However, they still suffer two nontrivial obstacles, i.e., low sample efficiency and overestimation bias. To this end, we propose Simultaneous Double Q-learning with Conservative Advantage Learning (SDQ-CAL). Our SDQ-CAL boosts the Double Q-learning for off-policy actor-critic RL based on a modification of the Bellman optimality operator with Advantage Learning. Specifically, SDQ-CAL improves sample efficiency by modifying the reward to facilitate the distinction from experience between the optimal actions and the others. Besides, it mitigates the overestimation issue by updating a pair of critics simultaneously upon double estimators. Extensive experiments reveal that our algorithm realizes less biased value estimation and achieves state-of-the-art performance in a range of continuous control benchmark tasks. We release the source code of our method at: \url{https://github.com/LQNew/SDQ-CAL}.