Online multi-task learning is an important capability for lifelong learning agents, enabling them to acquire models for diverse tasks over time and rapidly learn new tasks by building upon prior experience. However, recent progress toward lifelong reinforcement learning (RL) has been limited to learning from within a single task domain. For truly versatile lifelong learning, the agent must be able to autonomously transfer knowledge between different task domains. A few methods for cross-domain transfer have been developed, but these methods are computationally inefficient for scenarios where the agent must learn tasks consecutively. In this paper, we develop the first cross-domain lifelong RL framework. Our approach efficiently optimizes a shared repository of transferable knowledge and learns projection matrices that specialize that knowledge to different task domains. We provide rigorous theoretical guarantees on the stability of this approach, and empirically evaluate its performance on diverse dynamical systems. Our results show that the proposed method can learn effectively from interleaved task domains and rapidly acquire high performance in new domains.

Among various traditional art forms, brush stroke drawing is one of the widely used styles in modern computer graphic tools such as GIMP, Photoshop and Painter. In this paper, we develop an AI-aided art authoring (A4) system of non-photorealistic rendering that allows users to automatically generate brush stroke paintings in a specific artist's style. Within the reinforcement learning framework of brush stroke generation proposed, our contribution in this paper is to learn artists' drawing styles from video-captured stroke data by inverse reinforcement learning. Through experiments, we demonstrate that our system can successfully learn artists' styles and render pictures with consistent and smooth brush strokes.

When a transition probability matrix is represented as the product of two stochastic matrices, swapping the factors of the multiplication yields another transition matrix that retains some fundamental characteristics of the original. Since the new matrix can be much smaller than its precursor, replacing the former for the latter can lead to significant savings in terms of computational effort. This strategy, dubbed the "stochastic-factorization trick," can be used to compute the stationary distribution of a Markov chain, to determine the fundamental matrix of an absorbing chain, and to compute a decision policy via dynamic programming or reinforcement learning. In this paper we show that the stochastic-factorization trick can also provide benefits in terms of the number of samples needed to estimate a transition matrix. We introduce a probabilistic interpretation of a stochastic factorization and build on the resulting model to develop an algorithm to compute the factorization directly from data. If the transition matrix can be well approximated by a low-order stochastic factorization, estimating its factors instead of the original matrix reduces significantly the number of parameters to be estimated. Thus, when compared to estimating the transition matrix directly via maximum likelihood, the proposed method is able to compute approximations of roughly the same quality using less data. We illustrate the effectiveness of the proposed algorithm by using it to help a reinforcement learning agent learn how to play the game of blackjack.

Recently, several Web-scale knowledge harvesting systems have been built, each of which is competent at extracting information from certain types of data (e.g., unstructured text, structured tables on the web, etc.). In order to determine the response to a new query posed to such systems (e.g., is sugar a healthy food?), it is useful to integrate opinions from multiple systems. If a response is desired within a specific time budget (e.g., in less than 2 seconds), then maybe only a subset of these resources can be queried. In this paper, we address the problem of knowledge integration for on-demand time-budgeted query answering. We propose a new method, AskWorld, which learns a policy that chooses which queries to send to which resources, by accommodating varying budget constraints that are available only at query (test) time. Through extensive experiments on real world datasets, we demonstrate AskWorld’s capability in selecting most informative resources to query within test-time constraints, resulting in improved performance compared to competitive baselines.

Solomonoff induction is held as a gold standard for learning, but it is known to be incomputable. We quantify its incomputability by placing various flavors of Solomonoff's prior M in the arithmetical hierarchy. We also derive computability bounds for knowledge-seeking agents, and give a limit-computable weakly asymptotically optimal reinforcement learning agent.

This technical note presents a new approach to carrying out the kind of exploration achieved by Thompson sampling, but without explicitly maintaining or sampling from posterior distributions. The approach is based on a bootstrap technique that uses a combination of observed and artificially generated data. The latter serves to induce a prior distribution which, as we will demonstrate, is critical to effective exploration. We explain how the approach can be applied to multi-armed bandit and reinforcement learning problems and how it relates to Thompson sampling. The approach is particularly well-suited for contexts in which exploration is coupled with deep learning, since in these settings, maintaining or generating samples from a posterior distribution becomes computationally infeasible.

The true online TD({\lambda}) algorithm has recently been proposed (van Seijen and Sutton, 2014) as a universal replacement for the popular TD({\lambda}) algorithm, in temporal-difference learning and reinforcement learning. True online TD({\lambda}) has better theoretical properties than conventional TD({\lambda}), and the expectation is that it also results in faster learning. In this paper, we put this hypothesis to the test. Specifically, we compare the performance of true online TD({\lambda}) with that of TD({\lambda}) on challenging examples, random Markov reward processes, and a real-world myoelectric prosthetic arm. We use linear function approximation with tabular, binary, and non-binary features. We assess the algorithms along three dimensions: computational cost, learning speed, and ease of use. Our results confirm the strength of true online TD({\lambda}): 1) for sparse feature vectors, the computational overhead with respect to TD({\lambda}) is minimal; for non-sparse features the computation time is at most twice that of TD({\lambda}), 2) across all domains/representations the learning speed of true online TD({\lambda}) is often better, but never worse than that of TD({\lambda}), and 3) true online TD({\lambda}) is easier to use, because it does not require choosing between trace types, and it is generally more stable with respect to the step-size. Overall, our results suggest that true online TD({\lambda}) should be the first choice when looking for an efficient, general-purpose TD method.

Data-efficient learning in continuous state-action spaces using very high-dimensional observations remains a key challenge in developing fully autonomous systems. In this paper, we consider one instance of this challenge, the pixels to torques problem, where an agent must learn a closed-loop control policy from pixel information only. We introduce a data-efficient, model-based reinforcement learning algorithm that learns such a closed-loop policy directly from pixel information. The key ingredient is a deep dynamical model that uses deep auto-encoders to learn a low-dimensional embedding of images jointly with a predictive model in this low-dimensional feature space. Joint learning ensures that not only static but also dynamic properties of the data are accounted for. This is crucial for long-term predictions, which lie at the core of the adaptive model predictive control strategy that we use for closed-loop control. Compared to state-of-the-art reinforcement learning methods for continuous states and actions, our approach learns quickly, scales to high-dimensional state spaces and is an important step toward fully autonomous learning from pixels to torques.

Several authors have recently developed risk-sensitive policy gradient methods that augment the standard expected cost minimization problem with a measure of variability in cost. These studies have focused on specific risk-measures, such as the variance or conditional value at risk (CVaR). In this work, we extend the policy gradient method to the whole class of coherent risk measures, which is widely accepted in finance and operations research, among other fields. We consider both static and time-consistent dynamic risk measures. For static risk measures, our approach is in the spirit of policy gradient algorithms and combines a standard sampling approach with convex programming. For dynamic risk measures, our approach is actor-critic style and involves explicit approximation of value function. Most importantly, our contribution presents a unified approach to risk-sensitive reinforcement learning that generalizes and extends previous results.

This study is aimed at answering the famous question of how the approximation errors at each iteration of Approximate Dynamic Programming (ADP) affect the quality of the final results considering the fact that errors at each iteration affect the next iteration. To this goal, convergence of Value Iteration scheme of ADP for deterministic nonlinear optimal control problems with undiscounted cost functions is investigated while considering the errors existing in approximating respective functions. The boundedness of the results around the optimal solution is obtained based on quantities which are known in a general optimal control problem and assumptions which are verifiable. Moreover, since the presence of the approximation errors leads to the deviation of the results from optimality, sufficient conditions for stability of the system operated by the result obtained after a finite number of value iterations, along with an estimation of its region of attraction, are derived in terms of a calculable upper bound of the control approximation error. Finally, the process of implementation of the method on an orbital maneuver problem is investigated through which the assumptions made in the theoretical developments are verified and the sufficient conditions are applied for guaranteeing stability and near optimality.