The two most common perspectives on Reinforcement learning (RL) are optimization and dynamic programming. Methods that compute the gradients of the non-differentiable expected reward objective, such as the REINFORCE trick are commonly grouped into the optimization perspective, whereas methods that employ TD-learning or Q-learning are dynamic programming methods. While these methods have shown considerable success in recent years, these methods are still quite challenging to apply to new problems. In contrast deep supervised learning has been extremely successful and we may hence ask: Can we use supervised learning to perform RL? In this blog post we discuss a mental model for RL, based on the idea that RL can be viewed as doing supervised learning on the "good data".

Learning from visual observations is a fundamental yet challenging problem in Reinforcement Learning (RL). Although algorithmic advances combined with convolutional neural networks have proved to be a recipe for success, current methods are still lacking on two fronts: (a) data-efficiency of learning and (b) generalization to new environments. To this end, we present Reinforcement Learning with Augmented Data (RAD), a simple plug-and-play module that can enhance most RL algorithms. We perform the first extensive study of general data augmentations for RL on both pixel-based and state-based inputs, and introduce two new data augmentations - random translate and random amplitude scale. We show that augmentations such as random translate, crop, color jitter, patch cutout, random convolutions, and amplitude scale can enable simple RL algorithms to outperform complex state-of-the-art methods across common benchmarks. RAD sets a new state-of-the-art in terms of data-efficiency and final performance on the DeepMind Control Suite benchmark for pixel-based control as well as OpenAI Gym benchmark for state-based control. We further demonstrate that RAD significantly improves test-time generalization over existing methods on several OpenAI ProcGen benchmarks.

This paper presents a finite time convergence analysis for a decentralized stochastic approximation (SA) scheme. The scheme generalizes several algorithms for decentralized machine learning and multi-agent reinforcement learning. Our proof technique involves separating the iterates into their respective consensual parts and consensus error. The consensus error is bounded in terms of the stationarity of the consensual part, while the updates of the consensual part can be analyzed as a perturbed SA scheme. Under the Markovian noise and time varying communication graph assumptions, the decentralized SA scheme has an expected convergence rate of ${\cal O}(\log T/ \sqrt{T} )$, where $T$ is the iteration number, in terms of squared norms of gradient for nonlinear SA with smooth but non-convex cost function. This rate is comparable to the best known performances of SA in a centralized setting with a non-convex potential function.

This paper investigates a hybrid solution which combines deep reinforcement learning (RL) and classical trajectory planning for the following in front application. Here, an autonomous robot aims to stay ahead of a person as the person freely walks around. Following in front is a challenging problem as the user's intended trajectory is unknown and needs to be estimated, explicitly or implicitly, by the robot. In addition, the robot needs to find a feasible way to safely navigate ahead of human trajectory. Our deep RL module implicitly estimates human trajectory and produces short-term navigational goals to guide the robot. These goals are used by a trajectory planner to smoothly navigate the robot to the short-term goals, and eventually in front of the user. We employ curriculum learning in the deep RL module to efficiently achieve a high return. Our system outperforms the state-of-the-art in following ahead and is more reliable compared to end-to-end alternatives in both the simulation and real world experiments. In contrast to a pure deep RL approach, we demonstrate zero-shot transfer of the trained policy from simulation to the real world.

Abstract-- Current robot platforms available for research are either very expensive or unable to handle the abuse of exploratory controls in reinforcement learning. We develop RealAnt, a minimal low-cost physical version of the popular'Ant' benchmark used in reinforcement learning. RealAnt costs only 350 AC ($410) in materials and can be assembled in less than an hour. We validate the platform with reinforcement learning experiments and provide baseline results on a set of benchmark tasks. We demonstrate that the TD3 algorithm can learn to walk the RealAnt from less than 45 minutes of experience. We also provide simulator versions of the robot (with the same dimensions, state-action spaces, and delayed noisy observations) in the MuJoCo and PyBullet simulators.

Reward shaping is an effective technique for incorporating domain knowledge into reinforcement learning (RL). Existing approaches such as potential-based reward shaping normally make full use of a given shaping reward function. However, since the transformation of human knowledge into numeric reward values is often imperfect due to reasons such as human cognitive bias, completely utilizing the shaping reward function may fail to improve the performance of RL algorithms. In this paper, we consider the problem of adaptively utilizing a given shaping reward function. We formulate the utilization of shaping rewards as a bi-level optimization problem, where the lower level is to optimize policy using the shaping rewards and the upper level is to optimize a parameterized shaping weight function for true reward maximization. We formally derive the gradient of the expected true reward with respect to the shaping weight function parameters and accordingly propose three learning algorithms based on different assumptions. Experiments in sparse-reward cartpole and MuJoCo environments show that our algorithms can fully exploit beneficial shaping rewards, and meanwhile ignore unbeneficial shaping rewards or even transform them into beneficial ones.

The main novelty of the proposed approach is that it allows a robot to learn an end-to-end policy which can adapt to changes in the environment during execution. While goal conditioning of policies has been studied in the RL literature, such approaches are not easily extended to cases where the robot's goal can change during execution. This is something that humans are naturally able to do. However, it is difficult for robots to learn such reflexes (i.e., to naturally respond to dynamic environments), especially when the goal location is not explicitly provided to the robot, and instead needs to be perceived through a vision sensor. In the current work, we present a method that can achieve such behavior by combining traditional kinematic planning, deep learning, and deep reinforcement learning in a synergistic fashion to generalize to arbitrary environments. We demonstrate the proposed approach for several reaching and pick-and-place tasks in simulation, as well as on a real system of a 6-DoF industrial manipulator. A video describing our work could be found \url{https://youtu.be/hE-Ew59GRPQ}.

Based on simply watching how an agent acts in the environment it is hard to tell anything about why it behaves this way and how it works internally. That's why it is crucial to establish metrics that tell WHY the agent performs in a certain way. This is challenging especially when the agent doesn't behave the way we would like it to behave, … which is like always. Every AI practitioner knows that whatever we work on, most of the time it won't simply work out of the box (they wouldn't pay us so much for it otherwise). In this blog post, you'll learn what to keep track of to inspect/debug your agent learning trajectory. I'll assume you are already familiar with the Reinforcement Learning (RL) agent-environment setting (see Figure 1) and you've heard about at least some of the most common RL algorithms and environments. Nevertheless, don't worry if you are just beginning your journey with RL.

Researchers in recent years have deployed reinforcement learning (RL) agents to solve increasingly challenging problems. As the trend continues, so has the development of new methods that enable the injection of "priors" (prior knowledge) into agents to help them better understand the structure of the world and come up with more effective solution strategies. In a new paper, researchers from Google, OpenAI, and DeepMind introduce "behaviour priors," a framework designed to capture common movement and interaction patterns that are shared across a set of related tasks or contexts. The researchers discuss how such behaviour patterns can be captured using probabilistic trajectory models and how they can be integrated effectively into RL schemes, such as for facilitating multi-task and transfer learning. Their method for learning behaviour priors can lead to significant speedups on complex tasks, the researchers say.

Reinforcement learning (RL) is used to automate decision-making in a variety of domains, including games, autoscaling, finance, robotics, recommendations, and supply chain. Launched at AWS re:Invent 2018, Amazon SageMaker RL helps you quickly build, train, and deploy policies learned by RL. Ray is an open-source distributed execution framework that makes it easy to scale your Python applications. Amazon SageMaker RL uses the RLlib library that builds on the Ray framework to train RL policies. This post walks you through the tools available in Ray and Amazon SageMaker RL that help you address challenges such as scale, security, iterative development, and operational cost when you use RL in production.