Action and observation delays commonly occur in many Reinforcement Learning applications, such as remote control scenarios. We study the anatomy of randomly delayed environments, and show that partially resampling trajectory fragments in hindsight allows for off-policy multi-step value estimation. We apply this principle to derive Delay-Correcting Actor-Critic (DCAC), an algorithm based on Soft Actor-Critic with significantly better performance in environments with delays. This is shown theoretically and also demonstrated practically on a delay-augmented version of the MuJoCo continuous control benchmark.
Multitask Reinforcement Learning is a promising way to obtain models with better performance, generalisation, data efficiency, and robustness. Most existing work is limited to compatible settings, where the state and action space dimensions are the same across tasks. Graph Neural Networks (GNN) are one way to address incompatible environments, because they can process graphs of arbitrary size. They also allow practitioners to inject biases encoded in the structure of the input graph. Existing work in graph-based continuous control uses the physical morphology of the agent to construct the input graph, i.e., encoding limb features as node labels and using edges to connect the nodes if their corresponded limbs are physically connected. In this work, we present a series of ablations on existing methods that show that morphological information encoded in the graph does not improve their performance. Motivated by the hypothesis that any benefits GNNs extract from the graph structure are outweighed by difficulties they create for message passing, we also propose Amorpheus, a transformer-based approach. Further results show that, while Amorpheus ignores the morphological information that GNNs encode, it nonetheless substantially outperforms GNN-based methods that use the morphological information to define the message-passing scheme.
Model predictive control (MPC) has been successful in applications involving the control of complex physical systems. This class of controllers leverages the information provided by an approximate model of the system's dynamics to simulate the effect of control actions. MPC methods also present a few hyper-parameters which may require a relatively expensive tuning process by demanding interactions with the physical system. Therefore, we investigate fine-tuning MPC methods in the context of stochastic MPC, which presents extra challenges due to the randomness of the controller's actions. In these scenarios, performance outcomes present noise, which is not homogeneous across the domain of possible hyper-parameter settings, but which varies in an input-dependent way. To address these issues, we propose a Bayesian optimisation framework that accounts for heteroscedastic noise to tune hyper-parameters in control problems. Empirical results on benchmark continuous control tasks and a physical robot support the proposed framework's suitability relative to baselines, which do not take heteroscedasticity into account.
As discussed in previous studies, the efficacy of evolutionary or reinforcement learning algorithms for continuous control optimization can be enhanced by including a neural module dedicated to feature extraction trained through self-supervised methods. In this paper we report additional experiments supporting this hypothesis and we demonstrate how the advantage provided by feature extraction is not limited to problems that benefit from dimensionality reduction or that involve agents operating on the basis of allocentric perception. We introduce a method that permits to continue the training of the feature-extraction module during the training of the policy network and that increases the efficacy of feature extraction. Finally, we compare alternative feature-extracting methods and we show that sequence-to-sequence learning yields better results than the methods considered in previous studies.
In this work, we study the interaction of strategic agents in continuous action Cournot games with limited information feedback. Cournot game is the essential market model for many socio-economic systems where agents learn and compete without the full knowledge of the system or each other. We consider the dynamics of the policy gradient algorithm, which is a widely adopted continuous control reinforcement learning algorithm, in concave Cournot games. We prove the convergence of policy gradient dynamics to the Nash equilibrium when the price function is linear or the number of agents is two. This is the first result (to the best of our knowledge) on the convergence property of learning algorithms with continuous action spaces that do not fall in the no-regret class.
We present a novel approach (DyNODE) that captures the underlying dynamics of a system by incorporating control in a neural ordinary differential equation framework. We conduct a systematic evaluation and comparison of our method and standard neural network architectures for dynamics modeling. Our results indicate that a simple DyNODE architecture when combined with an actor-critic reinforcement learning (RL) algorithm that uses model predictions to improve the critic's target values, outperforms canonical neural networks, both in sample efficiency and predictive performance across a diverse range of continuous tasks that are frequently used to benchmark RL algorithms. This approach provides a new avenue for the development of models that are more suited to learn the evolution of dynamical systems, particularly useful in the context of model-based reinforcement learning. To assist related work, we have made code available at https://github.com/vmartinezalvarez/DyNODE .
Continuous control is a widely applicable area of reinforcement learning. The main players of this area are actor-critic methods that utilize policy gradients of neural approximators as a common practice. The focus of our study is to show the characteristics of the actor loss function which is the essential part of the optimization. We exploit low dimensional visualizations of the loss function and provide comparisons for loss landscapes of various algorithms. Furthermore, we apply our approach to multi-store dynamic inventory control, a notoriously difficult problem in supply chain operations, and explore the shape of the loss function associated with the optimal policy. We modelled and solved the problem using reinforcement learning while having a loss landscape in favor of optimality.
For over a decade, model-based reinforcement learning has been seen as a way to leverage control-based domain knowledge to improve the sample-efficiency of reinforcement learning agents. While model-based agents are conceptually appealing, their policies tend to lag behind those of model-free agents in terms of final reward, especially in non-trivial environments. In response, researchers have proposed model-based agents with increasingly complex components, from ensembles of probabilistic dynamics models, to heuristics for mitigating model error. In a reversal of this trend, we show that simple model-based agents can be derived from existing ideas that not only match, but outperform state-of-the-art model-free agents in terms of both sample-efficiency and final reward. We find that a model-free soft value estimate for policy evaluation and a model-based stochastic value gradient for policy improvement is an effective combination, achieving state-of-the-art results on a high-dimensional humanoid control task, which most model-based agents are unable to solve. Our findings suggest that model-based policy evaluation deserves closer attention.
Reinforcement learning (RL) in low-data and risk-sensitive domains requires performant and flexible deployment policies that can readily incorporate constraints during deployment. One such class of policies are the semi-parametric H-step lookahead policies, which select actions using trajectory optimization over a dynamics model for a fixed horizon with a terminal value function. In this work, we investigate a novel instantiation of H-step lookahead with a learned model and a terminal value function learned by a model-free off-policy algorithm, named Learning Off-Policy with Online Planning (LOOP). We provide a theoretical analysis of this method, suggesting a tradeoff between model errors and value function errors and empirically demonstrate this tradeoff to be beneficial in deep reinforcement learning. Furthermore, we identify the "Actor Divergence" issue in this framework and propose Actor Regularized Control (ARC), a modified trajectory optimization procedure. We evaluate our method on a set of robotic tasks for Offline and Online RL and demonstrate improved performance. We also show the flexibility of LOOP to incorporate safety constraints during deployment with a set of navigation environments. We demonstrate that LOOP is a desirable framework for robotics applications based on its strong performance in various important RL settings. Project video and details can be found at https://hari-sikchi.github.io/loop .
Many Reinforcement Learning (RL) approaches use joint control signals (positions, velocities, torques) as action space for continuous control tasks. We propose to lift the action space to a higher level in the form of subgoals for a motion generator (a combination of motion planner and trajectory executor). We argue that, by lifting the action space and by leveraging sampling-based motion planners, we can efficiently use RL to solve complex, long-horizon tasks that could not be solved with existing RL methods in the original action space. We propose ReLMoGen -- a framework that combines a learned policy to predict subgoals and a motion generator to plan and execute the motion needed to reach these subgoals. To validate our method, we apply ReLMoGen to two types of tasks: 1) Interactive Navigation tasks, navigation problems where interactions with the environment are required to reach the destination, and 2) Mobile Manipulation tasks, manipulation tasks that require moving the robot base. These problems are challenging because they are usually long-horizon, hard to explore during training, and comprise alternating phases of navigation and interaction. Our method is benchmarked on a diverse set of seven robotics tasks in photo-realistic simulation environments. In all settings, ReLMoGen outperforms state-of-the-art Reinforcement Learning and Hierarchical Reinforcement Learning baselines. ReLMoGen also shows outstanding transferability between different motion generators at test time, indicating a great potential to transfer to real robots.
State-of-the-art reinforcement learning algorithms mostly rely on being allowed to directly interact with their environment to collect millions of observations. This makes it hard to transfer their success to industrial control problems, where simulations are often very costly or do not exist, and exploring in the real environment can potentially lead to catastrophic events. Recently developed, model-free, offline RL algorithms, can learn from a single dataset (containing limited exploration) by mitigating extrapolation error in value functions. However, the robustness of the training process is still comparatively low, a problem known from methods using value functions. To improve robustness and stability of the learning process, we use dynamics models to assess policy performance instead of value functions, resulting in MOOSE (MOdel-based Offline policy Search with Ensembles), an algorithm which ensures low model bias by keeping the policy within the support of the data. We compare MOOSE with state-of-the-art model-free, offline RL algorithms { BRAC,} BEAR and BCQ on the Industrial Benchmark and MuJoCo continuous control tasks in terms of robust performance, and find that MOOSE outperforms its model-free counterparts in almost all considered cases, often even by far.
Deep reinforcement learning (DRL) has achieved significant success in various robot tasks: manipulation, navigation, etc. However, complex visual observations in natural environments remains a major challenge. This paper presents Contrastive Variational Reinforcement Learning (CVRL), a model-based method that tackles complex visual observations in DRL. CVRL learns a contrastive variational model by maximizing the mutual information between latent states and observations discriminatively, through contrastive learning. It avoids modeling the complex observation space unnecessarily, as the commonly used generative observation model often does, and is significantly more robust. CVRL achieves comparable performance with state-of-the-art model-based DRL methods on standard Mujoco tasks. It significantly outperforms them on Natural Mujoco tasks and a robot box-pushing task with complex observations, e.g., dynamic shadows. The CVRL code is available publicly at https://github.com/Yusufma03/CVRL.
Distribution-based search algorithms are an effective approach for evolutionary reinforcement learning of neural network controllers. In these algorithms, gradients of the total reward with respect to the policy parameters are estimated using a population of solutions drawn from a search distribution, and then used for policy optimization with stochastic gradient ascent. A common choice in the community is to use the Adam optimization algorithm for obtaining an adaptive behavior during gradient ascent, due to its success in a variety of supervised learning settings. As an alternative to Adam, we propose to enhance classical momentum-based gradient ascent with two simple techniques: gradient normalization and update clipping. We argue that the resulting optimizer called ClipUp (short for "clipped updates") is a better choice for distribution-based policy evolution because its working principles are simple and easy to understand and its hyperparameters can be tuned more intuitively in practice. Moreover, it removes the need to re-tune hyperparameters if the reward scale changes. Experiments show that ClipUp is competitive with Adam despite its simplicity and is effective on challenging continuous control benchmarks, including the Humanoid control task based on the Bullet physics simulator.
This paper introduces two simple techniques to improve off-policy Reinforcement Learning (RL) algorithms. First, we formulate off-policy RL as a stochastic proximal point iteration. The target network plays the role of the variable of optimization and the value network computes the proximal operator. Second, we exploits the two value functions commonly employed in state-of-the-art off-policy algorithms to provide an improved action value estimate through bootstrapping with limited increase of computational resources. Further, we demonstrate significant performance improvement over state-of-the-art algorithms on standard continuous-control RL benchmarks.
In reinforcement learning, it is typical to use the empirically observed transitions and rewards to estimate the value of a policy via either model-based or Q-fitting approaches. Although straightforward, these techniques in general yield biased estimates of the true value of the policy. In this work, we investigate the potential for statistical bootstrapping to be used as a way to take these biased estimates and produce calibrated confidence intervals for the true value of the policy. We identify conditions - specifically, sufficient data size and sufficient coverage - under which statistical bootstrapping in this setting is guaranteed to yield correct confidence intervals. In practical situations, these conditions often do not hold, and so we discuss and propose mechanisms that can be employed to mitigate their effects. We evaluate our proposed method and show that it can yield accurate confidence intervals in a variety of conditions, including challenging continuous control environments and small data regimes.
We propose a novel solution to challenging sparse-reward, continuous control problems that require hierarchical planning at multiple levels of abstraction. Our solution, dubbed AlphaNPI-X, involves three separate stages of learning. First, we use off-policy reinforcement learning algorithms with experience replay to learn a set of atomic goal-conditioned policies, which can be easily repurposed for many tasks. Second, we learn self-models describing the effect of the atomic policies on the environment. Third, the self-models are harnessed to learn recursive compositional programs with multiple levels of abstraction. The key insight is that the self-models enable planning by imagination, obviating the need for interaction with the world when learning higher-level compositional programs. To accomplish the third stage of learning, we extend the AlphaNPI algorithm, which applies AlphaZero to learn recursive neural programmer-interpreters. We empirically show that AlphaNPI-X can effectively learn to tackle challenging sparse manipulation tasks, such as stacking multiple blocks, where powerful model-free baselines fail.
The Predictive Information is the mutual information between the past and the future, I(X_past; X_future). We hypothesize that capturing the predictive information is useful in RL, since the ability to model what will happen next is necessary for success on many tasks. To test our hypothesis, we train Soft Actor-Critic (SAC) agents from pixels with an auxiliary task that learns a compressed representation of the predictive information of the RL environment dynamics using a contrastive version of the Conditional Entropy Bottleneck (CEB) objective. We refer to these as Predictive Information SAC (PI-SAC) agents. We show that PI-SAC agents can substantially improve sample efficiency over challenging baselines on tasks from the DM Control suite of continuous control environments. We evaluate PI-SAC agents by comparing against uncompressed PI-SAC agents, other compressed and uncompressed agents, and SAC agents directly trained from pixels. Our implementation is given on GitHub.
Prioritized Experience Replay (ER) has been empirically shown to improve sample efficiency across many domains and attracted great attention; however, there is little theoretical understanding of why such prioritized sampling helps and its limitations. In this work, we take a deep look at the prioritized ER. In a supervised learning setting, we show the equivalence between the error-based prioritized sampling method for mean squared error and uniform sampling for cubic power loss. We then provide theoretical insight into why it improves convergence rate upon uniform sampling during early learning. Based on the insight, we further point out two limitations of the prioritized ER method: 1) outdated priorities and 2) insufficient coverage of the sample space. To mitigate the limitations, we propose our model-based stochastic gradient Langevin dynamics sampling method. We show that our method does provide states distributed close to an ideal prioritized sampling distribution estimated by the brute-force method, which does not suffer from the two limitations. We conduct experiments on both discrete and continuous control problems to show our approach's efficacy and examine the practical implication of our method in an autonomous driving application.
The field of reinforcement learning can be split into model-based and model-free methods. Here, we unify these approaches by casting model-free policy optimisation as amortised variational inference, and model-based planning as iterative variational inference, within a `control as hybrid inference' (CHI) framework. We present an implementation of CHI which naturally mediates the balance between iterative and amortised inference. Using a didactic experiment, we demonstrate that the proposed algorithm operates in a model-based manner at the onset of learning, before converging to a model-free algorithm once sufficient data have been collected. We verify the scalability of our algorithm on a continuous control benchmark, demonstrating that it outperforms strong model-free and model-based baselines. CHI thus provides a principled framework for harnessing the sample efficiency of model-based planning while retaining the asymptotic performance of model-free policy optimisation.
In a reward-free environment, what is a suitable intrinsic objective for an agent to pursue so that it can learn an optimal task-agnostic exploration policy? In this paper, we argue that the entropy of the state distribution induced by finite-horizon trajectories is a sensible target. Especially, we present a novel and practical policy-search algorithm, Maximum Entropy POLicy optimization (MEPOL), to learn a policy that maximizes a non-parametric, $k$-nearest neighbors estimate of the state distribution entropy. In contrast to known methods, MEPOL is completely model-free as it requires neither to estimate the state distribution of any policy nor to model transition dynamics. Then, we empirically show that MEPOL allows learning a maximum-entropy exploration policy in high-dimensional, continuous-control domains, and how this policy facilitates learning a variety of meaningful reward-based tasks downstream.