Conditional Value at Risk (CVaR) is a prominent risk measure that is being used extensively in various domains. We develop a new formula for the gradient of the CVaR in the form of a conditional expectation. Based on this formula, we propose a novel sampling-based estimator for the gradient of the CVaR, in the spirit of the likelihood-ratio method. We analyze the bias of the estimator, and prove the convergence of a corresponding stochastic gradient descent algorithm to a local CVaR optimum. Our method allows to consider CVaR optimization in new domains. As an example, we consider a reinforcement learning application, and learn a risk-sensitive controller for the game of Tetris.

Understanding spatio-temporal resource preferences is paramount in the design of policies for sustainable development. Unfortunately, resource preferences are often unknown to policy-makers and have to be inferred from data. In this paper we consider the problem of inferring agents' preferences from observed movement trajectories, and formulate it as an Inverse Reinforcement Learning (IRL) problem . With the goal of informing policy-making, we take a probabilistic approach and consider generative models that can be used to simulate behavior under new circumstances such as changes in resource availability, access policies, or climate. We study the Dynamic Discrete Choice (DDC) models from econometrics and prove that they generalize the Max-Entropy IRL model, a widely used probabilistic approach from the machine learning literature. Furthermore, we develop SPL-GD, a new learning algorithm for DDC models that is considerably faster than the state of the art and scales to very large datasets. We consider an application in the context of pastoralism in the arid and semi-arid regions of Africa, where migratory pastoralists face regular risks due to resource availability, droughts, and resource degradation from climate change and development. We show how our approach based on satellite and survey data can accurately model migratory pastoralism in East Africa and that it considerably outperforms other approaches on a large-scale real-world dataset of pastoralists' movements in Ethiopia collected over 3 years.

We present a fully autonomous robotic system for grasping objects in dense clutter. The objects are unknown and have arbitrary shapes. Therefore, we cannot rely on prior models. Instead, the robot learns online, from scratch, to manipulate the objects by trial and error. Grasping objects in clutter is significantly harder than grasping isolated objects, because the robot needs to push and move objects around in order to create sufficient space for the fingers. These pre-grasping actions do not have an immediate utility, and may result in unnecessary delays. The utility of a pre-grasping action can be measured only by looking at the complete chain of consecutive actions and effects. This is a sequential decision-making problem that can be cast in the reinforcement learning framework. We solve this problem by learning the stochastic transitions between the observed states, using nonparametric density estimation. The learned transition function is used only for re-calculating the values of the executed actions in the observed states, with different policies. Values of new state-actions are obtained by regressing the values of the executed actions. The state of the system at a given time is a depth (3D) image of the scene. We use spectral clustering for detecting the different objects in the image. The performance of our system is assessed on a robot with real-world objects.

Reinforcement learning is a powerful machine learning paradigm that allows agents to autonomously learn to maximize a scalar reward. However, it often suffers from poor initial performance and long learning times. This paper discusses how collecting online human feedback, both in real time and post hoc, can potentially improve the performance of such learning systems. We use the game Pac-Man to simulate a navigation setting and show that workers are able to accurately identify both when a sub-optimal action is executed, and what action should have been performed instead. Our results demonstrate that the crowd is capable of generating helpful input. We conclude with a discussion the types of errors that occur most commonly when engaging human workers for this task, and a discussion of how such data could be used to improve learning. Our work serves as a critical first step in designing systems that use real-time human feedback to improve the learning performance of automated systems on-the-fly. Figure 1: This screenshot shows the web interface of the user study with game layout, and components of the Pac-Man game: 1) Pac-Man, 2) 4 Ghosts, 3) Pills, and 4) Power Pills.

In the context of soccer robotics, ball dribbling is a complex behavior where a robot player attempts to maneuver the ball in a very controlled way, while moving towards a desired target. To learn when and how to modify the robot’s velocity vector is a complex problem, hardly solvable in an effective way with methods based on identification of the system dynamics and/or kinematics and mathematical models. We propose a decentralized reinforcement learning strategy, where each component of the omnidirectional biped walk (𝑣𝑥,𝑣𝑦,𝑣𝜃) is learned in parallel with single-agents working in a multiagent task. Moreover, we propose an approach to accelerate the decentralized learning based on knowledge transfer from simple linear controllers. Obtained results are successful; with less human effort, and less required designer knowledge, the decentralized reinforcement learning scheme shows better performances than the current dribbling engine used by UChile Robotics Team in the SPL robot soccer competitions. The proposed decentralized rein- forcement learning scheme achieves asymptotic performance after 1500 episodes and can be accelerated up to 70% by using our approach to share actions.

We introduce a novel feature set for reinforcement learning in visual domains (e.g. video games) designed to capture pairwise, position-invariant, spatial relationships between objects on the screen. The feature set is simple to implement and computationally practical, but nevertheless allows for substantial improvement over existing baselines in a wide variety of Atari 2600 games. In the most dramatic results the features allow multiple orders of magnitude improvement in performance.

Recently it has been shown that deep neural networks can learn to play Atari games by directly observing raw pixels of the playing area. We show how apprenticeship learning can be applied in this setting so that an agent can learn to perform a task (i.e. play a game) by observing the expert, without any explicitly provided knowledge of the game’s internal state or objectives.

Reinforcement learning enables a robot to learn behavior through trial-and-error. However, knowledge is usually built from scratch and learning may take a long time. Many approaches have been proposed to transfer the knowledge learned in one task and reuse it in another new similar task to speed up learning in the target task.A very effective knowledge to be transferred is an abstract policy, which generalizes the learned policies in source tasks to extend the domain of tasks that can reuse them.There are inductive and deductive methods to generate abstract policies.However, there is a lack of deeper analysis to assess not only the effectiveness of each type of policy, but also the way in which each policy is used to accelerate the learning in a new task.In this paper we propose two simple inductive methods and we use a deductive method to generate stochastic abstract policies from source tasks. We also propose two strategies to use the abstract policy during learning in a new task: the hard and the soft strategy. We make a comparative analysis between the three types of policies and the two strategies of use in a robotic navigation domain.We show that these techniques are effective in improving the agent learning performance, especially during the early stages of the learning process, when the agent is completely unaware of the new task.

This article discusses the notion of context transfer in reinforcement learning tasks. Context transfer, as defined in this article, implies knowledge transfer between tasks that share the same environment's dynamics and reward function, but have different state and action spaces. For example, we have a working mobile robot in an environment. At some point, we decide to upgrade its sensors and/or actuators. Any change in these modules will result in a different description of the agent-environment model, and the trained knowledge is no longer applicable. We consider the tasks of the old and new robots, as the source and target tasks, respectively. The Markov decision process (MDP) of these tasks, under certain conditions, are called Q-transferable tasks, and the problem of knowledge transfer between them is called context transfer. We investigate the relation of the MDPs of these tasks.

Skills are generally designed by a domain expert, but designing a `good' set of skills can be challenging in high-dimensional, complex domains. In some cases, the skills may contain useful prior knowledge but cannot solve the task, resulting in a sub-optimal solution or no solution at all. Given a `poor' set of skills, we would like to dynamically improve them. The main contribution of this paper is showing that Interrupting Options (IO) can improve the initial skill set in a high-dimensional, complex domain by learning when to switch between skills. Furthermore, we discuss some of the pitfalls we ran into while trying to get IO to work.