Deep reinforcement learning (RL) agents often fail to generalize to unseen environments (yet semantically similar to trained agents), particularly when they are trained on high-dimensional state spaces, such as images. In this paper, we propose a simple technique to improve a generalization ability of deep RL agents by introducing a randomized (convolutional) neural network that randomly perturbs input observations. It enables trained agents to adapt to new domains by learning robust features invariant across varied and randomized environments. Furthermore, we consider an inference method based on the Monte Carlo approximation to reduce the variance induced by this randomization. We demonstrate the superiority of our method across 2D CoinRun, 3D DeepMind Lab exploration and 3D robotics control tasks: it significantly outperforms various regularization and data augmentation methods for the same purpose.

People readily generalise prior knowledge to novel situations and stimuli. Advances in machine learning and artificial intelligence have begun to approximate and even surpass human performance in specific domains, but machine learning systems struggle to generalise information to untrained situations. We present and model that demonstrates human-like extrapolatory generalisation by learning and explicitly representing an open-ended set of relations characterising regularities within the domains it is exposed to. First, when trained to play one video game (e.g., Breakout). the model generalises to a new game (e.g., Pong) with different rules, dimensions, and characteristics in a single shot. Second, the model can learn representations from a different domain (e.g., 3D shape images) that support learning a video game and generalising to a new game in one shot. By exploiting well-established principles from cognitive psychology and neuroscience, the model learns structured representations without feedback, and without requiring knowledge of the relevant relations to be given a priori. We present additional simulations showing that the representations that the model learns support cross-domain generalisation. The model's ability to generalise between different games demonstrates the flexible generalisation afforded by a capacity to learn not only statistical relations, but also other relations that are useful for characterising the domain to be learned. In turn, this kind of flexible, relational generalisation is only possible because the model is capable of representing relations explicitly, a capacity that is notably absent in extant statistical machine learning algorithms.

-- In this paper, we present a synergic learning algorithm to address the task of indirect manipulation of an unknown deformable tissue. Tissue manipulation is a common yet challenging task in various surgical interventions, which makes it a good candidate for robotic automation. We propose using a linear approximate Q-learning method in which human knowledge contributes to selecting useful yet simple features of tissue manipulation while the algorithm learns to take optimal actions and accomplish the task. The algorithm is implemented and evaluated on a simulation using the OpenCV and CHAI3D libraries. Successful simulation results for four different configurations which are based on realistic tissue manipulation scenarios are presented. Results indicate that with a careful selection of relatively simple and intuitive features, the developed Q-learning algorithm can successfully learn an optimal policy without any prior knowledge of tissue dynamics or camera intrinsic/extrinsic calibration parameters. Robot-Assisted Surgery (RAS) is becoming the norm of many operating room procedures, as it enables enhanced precision, dexterity, and feedback.

Hierarchical Reinforcement Learning (HRL) is a promising approach to solving long-horizon problems with sparse and delayed rewards. Many existing HRL algorithms either use pre-trained low-level skills that are unadaptable, or require domain-specific information to define low-level rewards. In this paper, we aim to adapt low-level skills to downstream tasks while maintaining the generality of reward design. We propose an HRL framework which sets auxiliary rewards for low-level skill training based on the advantage function of the high-level policy. This auxiliary reward enables efficient, simultaneous learning of the high-level policy and low-level skills without using task-specific knowledge. In addition, we also theoretically prove that optimizing low-level skills with this auxiliary reward will increase the task return for the joint policy. Experimental results show that our algorithm dramatically outperforms other state-of-the-art HRL methods in Mujoco domains. We also find both low-level and high-level policies trained by our algorithm transferable.

We consider off-policy policy evaluation when the trajectory data are generated by multiple behavior policies. Recent work has shown the key role played by the state or state-action stationary distribution corrections in the infinite horizon context for off-policy policy evaluation. We propose estimated mixture policy (EMP), a novel class of partially policy-agnostic methods to accurately estimate those quantities. With careful analysis, we show that EMP gives rise to estimates with reduced variance for estimating the state stationary distribution correction while it also offers a useful induction bias for estimating the state-action stationary distribution correction. In extensive experiments with both continuous and discrete environments, we demonstrate that our algorithm offers significantly improved accuracy compared to the state-of-the-art methods.

-- It is expected that many human drivers will still prefer to drive themselves even if the self-driving technologies are ready. T o enable A Vs to safely and efficiently maneuver in this mixed traffic, it is critical that the A Vs can understand how humans cope with risks and make driving-related decisions. On the other hand, the driving environment is highly dynamic and ever-changing, and it is thus difficult to enumerate all the scenarios and hard-code the controllers. T o face up these challenges, in this work, we incorporate a human decision-making model in reinforcement learning to control A Vs for safe and efficient operations. Specifically, we adapt regret theory to describe a human driver's lane-changing behavior, and fit the personalized models to individual drivers for predicting their lane-changing decisions. The predicted decisions are incorporated in the safety constraints for reinforcement learning in training and in implementation. We then use an extended version of double deep Q-network (DDQN) to train our A V controller within the safety set. By doing so, the amount of collisions in training is reduced to zero, while the training accuracy is not impinged. I. INTRODUCTION Autonomous driving has attracted significant research interest in the past two decades as it offers the potential to release drivers from exhausting driving. While great progresses have been made in the field of perception, path planning, and controls, high-level decision-making remains a big challenge due to the involvement of complex, cluttered environment and the dynamic, uncertain behaviors of other traffic users. Some recent works have been applying reinforcement learning (RL) methods to autonomous driving and promising performance [1] has been reported. RL-based methods can learn the decision-making and driving behaviors which are hard, if not infeasible, for traditional rule-based designs, and often with much less human effort. However, it is reported in [2] that when using RL-based methods lots of collisions happen before the agent starts to behave properly.

Combining model-based and model-free deep reinforcement learning has shown great promise for improving sample efficiency on complex control tasks while still retaining high performance. Incorporating imagination is a recent effort in this direction inspired by human mental simulation of motor behavior. We propose a learning-adaptive imagination approach which, unlike previous approaches, takes into account the reliability of the learned dynamics model used for imagining the future. Our approach learns an ensemble of disjoint local dynamics models in latent space and derives an intrinsic reward based on learning progress, motivating the controller to take actions leading to data that improves the models. The learned models are used to generate imagined experiences, augmenting the training set of real experiences. We evaluate our approach on learning vision-based robotic grasping and show that it significantly improves sample efficiency and achieves near-optimal performance in a sparse reward environment.

RLCard is an open-source toolkit for reinforcement learning research in card games. It supports various card environments with easy-to-use interfaces, including Blackjack, Leduc Hold'em, Texas Hold'em, UNO, Dou Dizhu and Mahjong. The goal of RLCard is to bridge reinforcement learning and imperfect information games, and push forward the research of reinforcement learning in domains with multiple agents, large state and action space, and sparse reward. In this paper, we provide an overview of the key components in RLCard, a discussion of the design principles, a brief introduction of the interfaces, and comprehensive evaluations of the environments.

Robots can learn the right reward function by querying a human expert. Existing approaches attempt to choose questions where the robot is most uncertain about the human's response; however, they do not consider how easy it will be for the human to answer! In this paper we explore an information gain formulation for optimally selecting questions that naturally account for the human's ability to answer. Our approach identifies questions that optimize the trade-off between robot and human uncertainty, and determines when these questions become redundant or costly. Simulations and a user study show our method not only produces easy questions, but also ultimately results in faster reward learning.

A BSTRACT The field of Deep Reinforcement Learning (DRL) has recently seen a surge in the popularity of maximum entropy reinforcement learning algorithms. Their popularity stems from the intuitive interpretation of the maximum entropy objective and their superior sample efficiency on standard benchmarks. In this paper, we seek to understand the primary contribution of the entropy term to the performance of maximum entropy algorithms. For the Mujoco benchmark, we demonstrate that the entropy term in Soft Actor Critic (SAC) principally addresses the bounded nature of the action spaces. With this insight, we propose a simple normalization scheme which allows a streamlined algorithm without entropy maximization match the performance of SAC. Our experimental results demonstrate a need to revisit the benefits of entropy regularization in DRL. We also propose a simple nonuniform sampling method for selecting transitions from the replay buffer during training. We further show that the streamlined algorithm with the simple nonuniform sampling scheme outperforms SAC and achieves state-of-the-art performance on challenging continuous control tasks. 1 I NTRODUCTION Off-policy deep Reinforcement Learning (RL) algorithms aim to improve sample efficiency by reusing past experience. Recently a number of new off-policy Deep Reinforcement Learning algorithms have been proposed for control tasks with continuous state and action spaces, including Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3) (Lillicrap et al., 2015; Fuji-moto et al., 2018). TD3, in particular, has been shown to be significantly more sample efficient than popular on-policy methods for a wide range of Mujoco benchmarks. The field of Deep Reinforcement Learning (DRL) has also recently seen a surge in the popularity of maximum entropy reinforcement learning algorithms. Their popularity stems from the intuitive interpretation of the maximum entropy objective and their superior sample efficiency on standard benchmarks.