Lucian Busoniu

A Comprehensive Survey of Multiagent Reinforcement Learning

Multiagent systems are rapidly finding applications in a variety of domains, including robotics, distributed control, telecommunications, and economics. The complexity of many tasks arising in these domains makes them difficult to solve with preprogrammed agent behaviors. The agents must, instead, discover a solution on their own, using learning. A significant part of the research on multiagent learning concerns reinforcement learning techniques. This paper provides a comprehensive survey of multiagent reinforcement learning (MARL). A central issue in the field is the formal statement of the multiagent learning goal. Different viewpoints on this issue have led to the proposal of many different goals, among which two focal points can be distinguished: stability of the agents learning dynamics, and adaptation to the changing behavior of the other agents. The MARL algorithms described in the literature aim---either explicitly or implicitly---at one of these two goals or at a combination of both, in a fully cooperative, fully competitive, or more general setting. A representative selection of these algorithms is discussed in detail in this paper, together with the specific issues that arise in each category. Additionally, the benefits and challenges of MARL are described along with some of the problem domains where the MARL techniques have been applied. Finally, an outlook for the field is provided.

Reinforcement Learning and Dynamic Programming Using Function Approximators

From household appliances to applications in robotics, engineered systems involving complex dynamics can only be as effective as the algorithms that control them. While Dynamic Programming (DP) has provided researchers with a way to optimally solve decision and control problems involving complex dynamic systems, its practical value was limited by algorithms that lacked the capacity to scale up to realistic problems. However, in recent years, dramatic developments in Reinforcement Learning (RL), the model-free counterpart of DP, changed our understanding of what is possible. Those developments led to the creation of reliable methods that can be applied even when a mathematical model of the system is unavailable, allowing researchers to solve challenging control problems in engineering, as well as in a variety of other disciplines, including economics, medicine, and artificial intelligence. Reinforcement Learning and Dynamic Programming Using Function Approximators provides a comprehensive and unparalleled exploration of the field of RL and DP. With a focus on continuous-variable problems, this seminal text details essential developments that have substantially altered the field over the past decade. In its pages, pioneering experts provide a concise introduction to classical RL and DP, followed by an extensive presentation of the state-of-the-art and novel methods in RL and DP with approximation. Combining algorithm development with theoretical guarantees, they elaborate on their work with illustrative examples and insightful comparisons. Three individual chapters are dedicated to representative algorithms from each of the major classes of techniques: value iteration, policy iteration, and policy search. The features and performance of these algorithms are highlighted in extensive experimental studies on a range of control applications. The recent development of applications involving complex systems has led to a surge of interest in RL and DP methods and the subsequent need for a quality resource on the subject. For graduate students and others new to the field, this book offers a thorough introduction to both the basics and emerging methods. And for those researchers and practitioners working in the fields of optimal and adaptive control, machine learning, artificial intelligence, and operations research, this resource offers a combination of practical algorithms, theoretical analysis, and comprehensive examples that they will be able to adapt and apply to their own work. Access the authors website at www.dcsc.tudelft.nl/rlbook/ for additional material, including computer code used in the studies and information concerning new developments.

A Survey of Actor-Critic Reinforcement Learning: Standard and Natural Policy Gradients

Policy-gradient-based actor-critic algorithms are amongst the most popular algorithms in the reinforcement learning framework. Their advantage of being able to search for optimal policies using low-variance gradient estimates has made them useful in several real-life applications, such as robotics, power control, and finance. Although general surveys on reinforcement learning techniques already exist, no survey is specifically dedicated to actor-critic algorithms in particular. This paper, therefore, describes the state of the art of actor-critic algorithms, with a focus on methods that can work in an online setting and use function approximation in order to deal with continuous state and action spaces. After starting with a discussion on the concepts of reinforcement learning and the origins of actor-critic algorithms, this paper describes the workings of the natural gradient, which has made its way into many actor-critic algorithms over the past few years. A review of several standard and natural actor-critic algorithms is given, and the paper concludes with an overview of application areas and a discussion on open issues.

Experience Replay for Real-Time Reinforcement Learning Control

Reinforcement-learning (RL) algorithms can automatically learn optimal control strategies for nonlinear, possibly stochastic systems. A promising approach for RL control is experience replay (ER), which learns quickly from a limited amount of data, by repeatedly presenting these data to an underlying RL algorithm. Despite its benefits, ER RL has been studied only sporadically in the literature, and its applications have largely been confined to simulated systems. Therefore, in this paper, we evaluate ER RL on real-time control experiments that involve a pendulum swing-up problem and the vision-based control of a goalkeeper robot. These real-time experiments are complemented by simulation studies and comparisons with traditional RL. As a preliminary, we develop a general ER framework that can be combined with essentially any incremental RL technique, and instantiate this framework for the approximate Q-learning and SARSA algorithms. The successful real-time learning results that are presented here are highly encouraging for the applicability of ER RL in practice.

Multi-agent Reinforcement Learning: An Overview

Multi-agent systems can be used to address problems in a variety of domains, including robotics, distributed control, telecommunications, and economics. The complexity of many tasks arising in these domains makes them difficult to solve with preprogrammed agent behaviors. The agents must instead discover a solution on their own, using learning. A significant part of the research on multi-agent learning concerns reinforcement learning techniques. This chapter reviews a representative selection of multi-agent reinforcement learning algorithms for fully cooperative, fully competitive, and more general (neither cooperative nor competitive) tasks. The benefits and challenges of multi-agent reinforcement learning are described. A central challenge in the field is the formal statement of a multi-agent learning goal; this chapter reviews the learning goals proposed in the literature. The problem domains where multi-agent reinforcement learning techniques have been applied are briefly discussed. Several multi-agent reinforcement learning algorithms are applied to an illustrative example involving the coordinated transportation of an object by two cooperative robots. In an outlook for the multi-agent reinforcement learning field, a set of important open issues are identified, and promising research directions to address these issues are outlined.

Cross-Entropy Optimization of Control Policies With Adaptive Basis Functions

This paper introduces an algorithm for direct search of control policies in continuous-state discrete-action Markov decision processes. The algorithm looks for the best closed-loop policy that can be represented using a given number of basis functions (BFs), where a discrete action is assigned to each BF. The type of the BFs and their number are specified in advance and determine the complexity of the representation. Considerable flexibility is achieved by optimizing the locations and shapes of the BFs, together with the action assignments. The optimization is carried out with the cross-entropy method and evaluates the policies by their empirical return from a representative set of initial states. The return for each representative state is estimated using Monte Carlo simulations. The resulting algorithm for cross-entropy policy search with adaptive BFs is extensively evaluated in problems with two to six state variables, for which it reliably obtains good policies with only a small number of BFs. In these experiments, cross-entropy policy search requires vastly fewer BFs than value-function techniques with equidistant BFs, and outperforms policy search with a competing optimization algorithm called DIRECT.

Online least-squares policy iteration for reinforcement learning control

Reinforcement learning is a promising paradigm for learning optimal control. We consider policy iteration (PI) algorithms for reinforcement learning, which iteratively evaluate and improve control policies. State-of-the-art, least-squares techniques for policy evaluation are sample-efficient and have relaxed convergence requirements. However, they are typically used in offline PI, whereas a central goal of reinforcement learning is to develop online algorithms. Therefore, we propose an online PI algorithm that evaluates policies with the so-called least-squares temporal difference for Q-functions (LSTD-Q). The crucial difference between this online least-squares policy iteration (LSPI) algorithm and its offline counterpart is that, in the online case, policy improvements must be performed once every few state transitions, using only an incomplete evaluation of the current policy. In an extensive experimental evaluation, online LSPI is found to work well for a wide range of its parameters, and to learn successfully in a real-time example. Online LSPI also compares favorably with offline LSPI and with a different flavor of online PI, which instead of LSTD-Q employs another least-squares method for policy evaluation.

Multi-Agent Reinforcement Learning: A Survey

Multi-agent systems are rapidly finding applications in a variety of domains, including robotics, distributed control, telecommunications, economics. Many tasks arising in these domains require that the agents learn behaviors online. A significant part of the research on multi-agent learning concerns reinforcement learning techniques. However, due to different viewpoints on central issues, such as the formal statement of the learning goal, a large number of different methods and approaches have been introduced. In this paper we aim to present an integrated survey of the field. First, the issue of the multi-agent learning goal is discussed, after which a representative selection of algorithms is reviewed. Finally, open issues are identified and future research directions are outlined

Approximate dynamic programming with a fuzzy parameterization

Dynamic programming (DP) is a powerful paradigm for general, nonlinear optimal control. Computing exact DP solutions is in general only possible when the process states and the control actions take values in a small discrete set. In practice, it is necessary to approximate the solutions. Therefore, we propose an algorithm for approximate DP that relies on a fuzzy partition of the state space, and on a discretization of the action space. This fuzzy Q-iteration algorithm works for deterministic processes, under the discounted return criterion. We prove that fuzzy Q-iteration asymptotically converges to a solution that lies within a bound of the optimal solution. A bound on the suboptimality of the solution obtained in a finite number of iterations is also derived. Under continuity assumptions on the dynamics and on the reward function, we show that fuzzy Q-iteration is consistent, i.e., that it asymptotically obtains the optimal solution as the approximation accuracy increases. These properties hold both when the parameters of the approximator are updated in a synchronous fashion, and when they are updated asynchronously. The asynchronous algorithm is proven to converge at least as fast as the synchronous one. The performance of fuzzy Q-iteration is illustrated in a two-link manipulator control problem.

Vision and Control for UAVs: A Survey of General Methods and of Inexpensive Platforms for Infrastructure Inspection

Unmanned aerial vehicles (UAVs) have gained significant attention in recent years. Low-cost platforms using inexpensive sensor payloads have been shown to provide satisfactory flight and navigation capabilities. In this report, we survey vision and control methods that can be applied to low-cost UAVs, and we list some popular inexpensive platforms and application fields where they are useful. We also highlight the sensor suites used where this information is available. We overview, among others, feature detection and tracking, optical flow and visual servoing, low-level stabilization and high-level planning methods. We then list popular low-cost UAVs, selecting mainly quadrotors. We discuss applications, restricting our focus to the field of infrastructure inspection. Finally, as an example, we formulate two use-cases for railway inspection, a less explored application field, and illustrate the usage of the vision and control techniques reviewed by selecting appropriate ones to tackle these use-cases. To select vision methods, we run a thorough set of experimental evaluations.