Showing papers on "Reinforcement learning published in 2011"

Approximate dynamic programming : solving the curses of dimensionality

Abstract: Preface. Acknowledgments. 1. The challenges of dynamic programming. 1.1 A dynamic programming example: a shortest path problem. 1.2 The three curses of dimensionality. 1.3 Some real applications. 1.4 Problem classes. 1.5 The many dialects of dynamic programming. 1.6 What is new in this book? 1.7 Bibliographic notes. 2. Some illustrative models. 2.1 Deterministic problems. 2.2 Stochastic problems. 2.3 Information acquisition problems. 2.4 A simple modeling framework for dynamic programs. 2.5 Bibliographic notes. Problems. 3. Introduction to Markov decision processes. 3.1 The optimality equations. 3.2 Finite horizon problems. 3.3 Infinite horizon problems. 3.4 Value iteration. 3.5 Policy iteration. 3.6 Hybrid valuepolicy iteration. 3.7 The linear programming method for dynamic programs. 3.8 Monotone policies. 3.9 Why does it work? 3.10 Bibliographic notes. Problems 4. Introduction to approximate dynamic programming. 4.1 The three curses of dimensionality (revisited). 4.2 The basic idea. 4.3 Sampling random variables . 4.4 ADP using the postdecision state variable. 4.5 Lowdimensional representations of value functions. 4.6 So just what is approximate dynamic programming? 4.7 Experimental issues. 4.8 Dynamic programming with missing or incomplete models. 4.9 Relationship to reinforcement learning. 4.10 But does it work? 4.11 Bibliographic notes. Problems. 5. Modeling dynamic programs. 5.1 Notational style. 5.2 Modeling time. 5.3 Modeling resources. 5.4 The states of our system. 5.5 Modeling decisions. 5.6 The exogenous information process. 5.7 The transition function. 5.8 The contribution function. 5.9 The objective function. 5.10 A measuretheoretic view of information. 5.11 Bibliographic notes. Problems. 6. Stochastic approximation methods. 6.1 A stochastic gradient algorithm. 6.2 Some stepsize recipes. 6.3 Stochastic stepsizes. 6.4 Computing bias and variance. 6.5 Optimal stepsizes. 6.6 Some experimental comparisons of stepsize formulas. 6.7 Convergence. 6.8 Why does it work? 6.9 Bibliographic notes. Problems. 7. Approximating value functions. 7.1 Approximation using aggregation. 7.2 Approximation methods using regression models. 7.3 Recursive methods for regression models. 7.4 Neural networks. 7.5 Batch processes. 7.6 Why does it work? 7.7 Bibliographic notes. Problems. 8. ADP for finite horizon problems. 8.1 Strategies for finite horizon problems. 8.2 Qlearning. 8.3 Temporal difference learning. 8.4 Policy iteration. 8.5 Monte Carlo value and policy iteration. 8.6 The actorcritic paradigm. 8.7 Bias in value function estimation. 8.8 State sampling strategies. 8.9 Starting and stopping. 8.10 A taxonomy of approximate dynamic programming strategies. 8.11 Why does it work? 8.12 Bibliographic notes. Problems. 9. Infinite horizon problems. 9.1 From finite to infinite horizon. 9.2 Algorithmic strategies. 9.3 Stepsizes for infinite horizon problems. 9.4 Error measures. 9.5 Direct ADP for online applications. 9.6 Finite horizon models for steady state applications. 9.7 Why does it work? 9.8 Bibliographic notes. Problems. 10. Exploration vs. exploitation. 10.1 A learning exercise: the nomadic trucker. 10.2 Learning strategies. 10.3 A simple information acquisition problem. 10.4 Gittins indices and the information acquisition problem. 10.5 Variations. 10.6 The knowledge gradient algorithm. 10.7 Information acquisition in dynamic programming. 10.8 Bibliographic notes. Problems. 11. Value function approximations for special functions. 11.1 Value functions versus gradients. 11.2 Linear approximations. 11.3 Piecewise linear approximations. 11.4 The SHAPE algorithm. 11.5 Regression methods. 11.6 Cutting planes. 11.7 Why does it work? 11.8 Bibliographic notes. Problems. 12. Dynamic resource allocation. 12.1 An asset acquisition problem. 12.2 The blood management problem. 12.3 A portfolio optimization problem. 12.4 A general resource allocation problem. 12.5 A fleet management problem. 12.6 A driver management problem. 12.7 Bibliographic references. Problems. 13. Implementation challenges. 13.1 Will ADP work for your problem? 13.2 Designing an ADP algorithm for complex problems. 13.3 Debugging an ADP algorithm. 13.4 Convergence issues. 13.5 Modeling your problem. 13.6 Online vs. offline models. 13.7 If it works, patent it!

PILCO: A Model-Based and Data-Efficient Approach to Policy Search

Abstract: In this paper, we introduce PILCO, a practical, data-efficient model-based policy search method. PILCO reduces model bias, one of the key problems of model-based reinforcement learning, in a principled way. By learning a probabilistic dynamics model and explicitly incorporating model uncertainty into long-term planning, PILCO can cope with very little data and facilitates learning from scratch in only a few trials. Policy evaluation is performed in closed form using state-of-the-art approximate inference. Furthermore, policy gradients are computed analytically for policy improvement. We report unprecedented learning efficiency on challenging and high-dimensional control tasks.

Infinite-Horizon Policy-Gradient Estimation

Abstract: Gradient-based approaches to direct policy search in reinforcement learning have received much recent attention as a means to solve problems of partial observability and to avoid some of the problems associated with policy degradation in value-function methods. In this paper we introduce GPOMDP, a simulation-based algorithm for generating a {\em biased} estimate of the gradient of the {\em average reward} in Partially Observable Markov Decision Processes (POMDPs) controlled by parameterized stochastic policies. A similar algorithm was proposed by Kimura, Yamamura, and Kobayashi (1995). The algorithm's chief advantages are that it requires storage of only twice the number of policy parameters, uses one free parameter $\beta\in [0,1)$ (which has a natural interpretation in terms of bias-variance trade-off), and requires no knowledge of the underlying state. We prove convergence of GPOMDP, and show how the correct choice of the parameter $\beta$ is related to the {\em mixing time} of the controlled POMDP. We briefly describe extensions of GPOMDP to controlled Markov chains, continuous state, observation and control spaces, multiple-agents, higher-order derivatives, and a version for training stochastic policies with internal states. In a companion paper (Baxter, Bartlett, & Weaver, 2001) we show how the gradient estimates generated by GPOMDP can be used in both a traditional stochastic gradient algorithm and a conjugate-gradient procedure to find local optima of the average reward

A Bayesian Foundation for Individual Learning Under Uncertainty

Abstract: Computational learning models are critical for understanding mechanisms of adaptive behavior. However, the two major current frameworks, reinforcement learning (RL) and Bayesian learning, both have certain limitations. For example, many Bayesian models are agnostic of inter-individual variability and involve complicated integrals, making online learning difficult. Here, we introduce a generic hierarchical Bayesian framework for individual learning under multiple forms of uncertainty (e.g., environmental volatility and perceptual uncertainty). The model assumes Gaussian random walks of states at all but the first level, with the step size determined by the next higher level. The coupling between levels is controlled by parameters that shape the influence of uncertainty on learning in a subject-specific fashion. Using variational Bayes under a mean field approximation and a novel approximation to the posterior energy function, we derive trial-by-trial update equations which (i) are analytical and extremely efficient, enabling real-time learning, (ii) have a natural interpretation in terms of RL, and (iii) contain parameters representing processes which play a key role in current theories of learning, e.g., precision-weighting of prediction error. These parameters allow for the expression of individual differences in learning and may relate to specific neuromodulatory mechanisms in the brain. Our model is very general: it can deal with both discrete and continuous states and equally accounts for deterministic and probabilistic relations between environmental events and perceptual states (i.e., situations with and without perceptual uncertainty). These properties are illustrated by simulations and analyses of empirical time series. Overall, our framework provides a novel foundation for understanding normal and pathological learning that contextualizes RL within a generic Bayesian scheme and thus connects it to principles of optimality from probability theory.

Policy search for motor primitives in robotics

Abstract: Many motor skills in humanoid robotics can be learned using parametrized motor primitives. While successful applications to date have been achieved with imitation learning, most of the interesting motor learning problems are high-dimensional reinforcement learning problems. These problems are often beyond the reach of current reinforcement learning methods. In this paper, we study parametrized policy search methods and apply these to benchmark problems of motor primitive learning in robotics. We show that many well-known parametrized policy search methods can be derived from a general, common framework. This framework yields both policy gradient methods and expectation-maximization (EM) inspired algorithms. We introduce a novel EM-inspired algorithm for policy learning that is particularly well-suited for dynamical system motor primitives. We compare this algorithm, both in simulation and on a real robot, to several well-known parametrized policy search methods such as episodic REINFORCE, `Vanilla' Policy Gradients with optimal baselines, episodic Natural Actor Critic, and episodic Reward-Weighted Regression. We show that the proposed method out-performs them on an empirical benchmark of learning dynamical system motor primitives both in simulation and on a real robot. We apply it in the context of motor learning and show that it can learn a complex Ball-in-a-Cup task on a real Barrett WAM™ robot arm.

Horde: a scalable real-time architecture for learning knowledge from unsupervised sensorimotor interaction

Abstract: Maintaining accurate world knowledge in a complex and changing environment is a perennial problem for robots and other artificial intelligence systems. Our architecture for addressing this problem, called Horde, consists of a large number of independent reinforcement learning sub-agents, or demons. Each demon is responsible for answering a single predictive or goal-oriented question about the world, thereby contributing in a factored, modular way to the system's overall knowledge. The questions are in the form of a value function, but each demon has its own policy, reward function, termination function, and terminal-reward function unrelated to those of the base problem. Learning proceeds in parallel by all demons simultaneously so as to extract the maximal training information from whatever actions are taken by the system as a whole. Gradient-based temporal-difference learning methods are used to learn efficiently and reliably with function approximation in this off-policy setting. Horde runs in constant time and memory per time step, and is thus suitable for learning online in real-time applications such as robotics. We present results using Horde on a multi-sensored mobile robot to successfully learn goal-oriented behaviors and long-term predictions from off-policy experience. Horde is a significant incremental step towards a real-time architecture for efficient learning of general knowledge from unsupervised sensorimotor interaction.

Reinforcement Learning for Partially Observable Dynamic Processes: Adaptive Dynamic Programming Using Measured Output Data

Abstract: Approximate dynamic programming (ADP) is a class of reinforcement learning methods that have shown their importance in a variety of applications, including feedback control of dynamical systems. ADP generally requires full information about the system internal states, which is usually not available in practical situations. In this paper, we show how to implement ADP methods using only measured input/output data from the system. Linear dynamical systems with deterministic behavior are considered herein, which are systems of great interest in the control system community. In control system theory, these types of methods are referred to as output feedback (OPFB). The stochastic equivalent of the systems dealt with in this paper is a class of partially observable Markov decision processes. We develop both policy iteration and value iteration algorithms that converge to an optimal controller that requires only OPFB. It is shown that, similar to Q-learning, the new methods have the important advantage that knowledge of the system dynamics is not needed for the implementation of these learning algorithms or for the OPFB control. Only the order of the system, as well as an upper bound on its "observability index," must be known. The learned OPFB controller is in the form of a polynomial autoregressive moving-average controller that has equivalent performance with the optimal state variable feedback gain.

Cooperative Adaptive Cruise Control: A Reinforcement Learning Approach

Abstract: Recently, improvements in sensing, communicating, and computing technologies have led to the development of driver-assistance systems (DASs). Such systems aim at helping drivers by either providing a warning to reduce crashes or doing some of the control tasks to relieve a driver from repetitive and boring tasks. Thus, for example, adaptive cruise control (ACC) aims at relieving a driver from manually adjusting his/her speed to maintain a constant speed or a safe distance from the vehicle in front of him/her. Currently, ACC can be improved through vehicle-to-vehicle communication, where the current speed and acceleration of a vehicle can be transmitted to the following vehicles by intervehicle communication. This way, vehicle-to-vehicle communication with ACC can be combined in one single system called cooperative adaptive cruise control (CACC). This paper investigates CACC by proposing a novel approach for the design of autonomous vehicle controllers based on modern machine-learning techniques. More specifically, this paper shows how a reinforcement-learning approach can be used to develop controllers for the secure longitudinal following of a front vehicle. This approach uses function approximation techniques along with gradient-descent learning algorithms as a means of directly modifying a control policy to optimize its performance. The experimental results, through simulation, show that this design approach can result in efficient behavior for CACC.

Speed/accuracy trade-off between the habitual and the goal-directed processes.

Abstract: Instrumental responses are hypothesized to be of two kinds: habitual and goal-directed, mediated by the sensorimotor and the associative cortico-basal ganglia circuits, respectively. The existence of the two heterogeneous associative learning mechanisms can be hypothesized to arise from the comparative advantages that they have at different stages of learning. In this paper, we assume that the goal-directed system is behaviourally flexible, but slow in choice selection. The habitual system, in contrast, is fast in responding, but inflexible in adapting its behavioural strategy to new conditions. Based on these assumptions and using the computational theory of reinforcement learning, we propose a normative model for arbitration between the two processes that makes an approximately optimal balance between search-time and accuracy in decision making. Behaviourally, the model can explain experimental evidence on behavioural sensitivity to outcome at the early stages of learning, but insensitivity at the later stages. It also explains that when two choices with equal incentive values are available concurrently, the behaviour remains outcome-sensitive, even after extensive training. Moreover, the model can explain choice reaction time variations during the course of learning, as well as the experimental observation that as the number of choices increases, the reaction time also increases. Neurobiologically, by assuming that phasic and tonic activities of midbrain dopamine neurons carry the reward prediction error and the average reward signals used by the model, respectively, the model predicts that whereas phasic dopamine indirectly affects behaviour through reinforcing stimulus-response associations, tonic dopamine can directly affect behaviour through manipulating the competition between the habitual and the goal-directed systems and thus, affect reaction time.

An anti-jamming stochastic game for cognitive radio networks

Abstract: Various spectrum management schemes have been proposed in recent years to improve the spectrum utilization in cognitive radio networks. However, few of them have considered the existence of cognitive attackers who can adapt their attacking strategy to the time-varying spectrum environment and the secondary users' strategy. In this paper, we investigate the security mechanism when secondary users are facing the jamming attack, and propose a stochastic game framework for anti-jamming defense. At each stage of the game, secondary users observe the spectrum availability, the channel quality, and the attackers' strategy from the status of jammed channels. According to this observation, they will decide how many channels they should reserve for transmitting control and data messages and how to switch between the different channels. Using the minimax-Q learning, secondary users can gradually learn the optimal policy, which maximizes the expected sum of discounted payoffs defined as the spectrum-efficient throughput. The proposed stationary policy in the anti-jamming game is shown to achieve much better performance than the policy obtained from myopic learning, which only maximizes each stage's payoff, and a random defense strategy, since it successfully accommodates the environment dynamics and the strategic behavior of the cognitive attackers.

Information Theory of Decisions and Actions

Abstract: The perception–action cycle is often defined as “the circular flow of information between an organism and its environment in the course of a sensory guided sequence of actions towards a goal” (Fuster, Neuron 30:319–333, 2001; International Journal of Psychophysiology 60(2):125–132, 2006). The question we address in this chapter is in what sense this “flow of information” can be described by Shannon’s measures of information introduced in his mathematical theory of communication. We provide an affirmative answer to this question using an intriguing analogy between Shannon’s classical model of communication and the perception–action cycle. In particular, decision and action sequences turn out to be directly analogous to codes in communication, and their complexity – the minimal number of (binary) decisions required for reaching a goal – directly bounded by information measures, as in communication. This analogy allows us to extend the standard reinforcement learning framework. The latter considers the future expected reward in the course of a behaviour sequence towards a goal (value-to-go). Here, we additionally incorporate a measure of information associated with this sequence: the cumulated information processing cost or bandwidth required to specify the future decision and action sequence (information-to-go). Using a graphical model, we derive a recursive Bellman optimality equation for information measures, in analogy to reinforcement learning; from this, we obtain new algorithms for calculating the optimal trade-off between the value-to-go and the required information-to-go, unifying the ideas behind the Bellman and the Blahut–Arimoto iterations. This trade-off between value-to-go and information-to-go provides a complete analogy with the compression–distortion trade-off in source coding. The present new formulation connects seemingly unrelated optimization problems. The algorithm is demonstrated on grid world examples.

Learning variable impedance control

Abstract: One of the hallmarks of the performance, versatility, and robustness of biological motor control is the ability to adapt the impedance of the overall biomechanical system to different task requirements and stochastic disturbances. A transfer of this principle to robotics is desirable, for instance to enable robots to work robustly and safely in everyday human environments. It is, however, not trivial to derive variable impedance controllers for practical high degree-of-freedom (DOF) robotic tasks. In this contribution, we accomplish such variable impedance control with the reinforcement learning (RL) algorithm PI2 (Policy Improvement with Path Integrals). PI2 is a model-free, sampling-based learning method derived from first principles of stochastic optimal control. The PI 2 algorithm requires no tuning of algorithmic parameters besides the exploration noise. The designer can thus fully focus on the cost function design to specify the task. From the viewpoint of robotics, a particular useful property of PI2 is that it can scale to problems of many DOFs, so that reinforcement learning on real robotic systems becomes feasible. We sketch the PI2 algorithm and its theoretical properties, and how it is applied to gain scheduling for variable impedance control. We evaluate our approach by presenting results on several simulated and real robots. We consider tasks involving accurate tracking through via points, and manipulation tasks requiring physical contact with the environment. In these tasks, the optimal strategy requires both tuning of a reference trajectory and the impedance of the end-effector. The results show that we can use path integral based reinforcement learning not only for planning but also to derive variable gain feedback controllers in realistic scenarios. Thus, the power of variable impedance control is made available to a wide variety of robotic systems and practical applications.

Adaptive Dynamic Programming for Finite-Horizon Optimal Control of Discrete-Time Nonlinear Systems With $\varepsilon$ -Error Bound

Abstract: In this paper, we study the finite-horizon optimal control problem for discrete-time nonlinear systems using the adaptive dynamic programming (ADP) approach. The idea is to use an iterative ADP algorithm to obtain the optimal control law which makes the performance index function close to the greatest lower bound of all performance indices within an -error bound. The optimal number of control steps can also be obtained by the proposed ADP algorithms. A convergence analysis of the proposed ADP algorithms in terms of performance index function and control policy is made. In order to facilitate the implementation of the iterative ADP algorithms, neural networks are used for approximating the performance index function, computing the optimal control policy, and modeling the nonlinear system. Finally, two simulation examples are employed to illustrate the applicability of the proposed method.

Reinforcement Learning With Function Approximation for Traffic Signal Control

Abstract: We propose, for the first time, a reinforcement learning (RL) algorithm with function approximation for traffic signal control. Our algorithm incorporates state-action features and is easily implementable in high-dimensional settings. Prior work, e.g., the work of Abdulhai , on the application of RL to traffic signal control requires full-state representations and cannot be implemented, even in moderate-sized road networks, because the computational complexity exponentially grows in the numbers of lanes and junctions. We tackle this problem of the curse of dimensionality by effectively using feature-based state representations that use a broad characterization of the level of congestion as low, medium, or high. One advantage of our algorithm is that, unlike prior work based on RL, it does not require precise information on queue lengths and elapsed times at each lane but instead works with the aforementioned described features. The number of features that our algorithm requires is linear to the number of signaled lanes, thereby leading to several orders of magnitude reduction in the computational complexity. We perform implementations of our algorithm on various settings and show performance comparisons with other algorithms in the literature, including the works of Abdulhai and Cools , as well as the fixed-timing and the longest queue algorithms. For comparison, we also develop an RL algorithm that uses full-state representation and incorporates prioritization of traffic, unlike the work of Abdulhai We observe that our algorithm outperforms all the other algorithms on all the road network settings that we consider.

Empirical evaluation methods for multiobjective reinforcement learning algorithms

Abstract: While a number of algorithms for multiobjective reinforcement learning have been proposed, and a small number of applications developed, there has been very little rigorous empirical evaluation of the performance and limitations of these algorithms. This paper proposes standard methods for such empirical evaluation, to act as a foundation for future comparative studies. Two classes of multiobjective reinforcement learning algorithms are identified, and appropriate evaluation metrics and methodologies are proposed for each class. A suite of benchmark problems with known Pareto fronts is described, and future extensions and implementations of this benchmark suite are discussed. The utility of the proposed evaluation methods are demonstrated via an empirical comparison of two example learning algorithms.

Learning to Control a Low-Cost Manipulator using Data-Efficient Reinforcement Learning

Abstract: Over the last years, there has been substantial progress in robust manipulation in unstructured environments. The long-term goal of our work is to get away from precise, but very expensive robotic systems and to develop affordable, potentially imprecise, self-adaptive manipulator systems that can interactively perform tasks such as playing with children. In this paper, we demonstrate how a low-cost off-the-shelf robotic system can learn closed-loop policies for a stacking task in only a handful of trials-from scratch. Our manipulator is inaccurate and provides no pose feedback. For learning a controller in the work space of a Kinect-style depth camera, we use a model-based reinforcement learning technique. Our learning method is data efficient, reduces model bias, and deals with several noise sources in a principled way during long-term planning. We present a way of incorporating state-space constraints into the learning process and analyze the learning gain by exploiting the sequential structure of the stacking task.

Skill learning and task outcome prediction for manipulation

Abstract: Learning complex motor skills for real world tasks is a hard problem in robotic manipulation that often requires painstaking manual tuning and design by a human expert. In this work, we present a Reinforcement Learning based approach to acquiring new motor skills from demonstration. Our approach allows the robot to learn fine manipulation skills and significantly improve its success rate and skill level starting from a possibly coarse demonstration. Our approach aims to incorporate task domain knowledge, where appropriate, by working in a space consistent with the constraints of a specific task. In addition, we also present an approach to using sensor feedback to learn a predictive model of the task outcome. This allows our system to learn the proprioceptive sensor feedback needed to monitor subsequent executions of the task online and abort execution in the event of predicted failure. We illustrate our approach using two example tasks executed with the PR2 dual-arm robot: a straight and accurate pool stroke and a box flipping task using two chopsticks as tools.

Ventral Striatum and Orbitofrontal Cortex Are Both Required for Model-Based, But Not Model-Free, Reinforcement Learning

Abstract: In many cases, learning is thought to be driven by differences between the value of rewards we expect and rewards we actually receive. Yet learning can also occur when the identity of the reward we receive is not as expected, even if its value remains unchanged. Learning from changes in reward identity implies access to an internal model of the environment, from which information about the identity of the expected reward can be derived. As a result, such learning is not easily accounted for by model-free reinforcement learning theories such as temporal difference reinforcement learning (TDRL), which predicate learning on changes in reward value, but not identity. Here, we used unblocking procedures to assess learning driven by value- versus identity-based prediction errors. Rats were trained to associate distinct visual cues with different food quantities and identities. These cues were subsequently presented in compound with novel auditory cues and the reward quantity or identity was selectively changed. Unblocking was assessed by presenting the auditory cues alone in a probe test. Consistent with neural implementations of TDRL models, we found that the ventral striatum was necessary for learning in response to changes in reward value. However, this area, along with orbitofrontal cortex, was also required for learning driven by changes in reward identity. This observation requires that existing models of TDRL in the ventral striatum be modified to include information about the specific features of expected outcomes derived from model-based representations, and that the role of orbitofrontal cortex in these models be clearly delineated.

Risk, Unexpected Uncertainty, and Estimation Uncertainty: Bayesian Learning in Unstable Settings

Abstract: Recently, evidence has emerged that humans approach learning using Bayesian updating rather than (model-free) reinforcement algorithms in a six-arm restless bandit problem. Here, we investigate what this implies for human appreciation of uncertainty. In our task, a Bayesian learner distinguishes three equally salient levels of uncertainty. First, the Bayesian perceives irreducible uncertainty or risk: even knowing the payoff probabilities of a given arm, the outcome remains uncertain. Second, there is (parameter) estimation uncertainty or ambiguity: payoff probabilities are unknown and need to be estimated. Third, the outcome probabilities of the arms change: the sudden jumps are referred to as unexpected uncertainty. We document how the three levels of uncertainty evolved during the course of our experiment and how it affected the learning rate. We then zoom in on estimation uncertainty, which has been suggested to be a driving force in exploration, in spite of evidence of widespread aversion to ambiguity. Our data corroborate the latter. We discuss neural evidence that foreshadowed the ability of humans to distinguish between the three levels of uncertainty. Finally, we investigate the boundaries of human capacity to implement Bayesian learning. We repeat the experiment with different instructions, reflecting varying levels of structural uncertainty. Under this fourth notion of uncertainty, choices were no better explained by Bayesian updating than by (model-free) reinforcement learning. Exit questionnaires revealed that participants remained unaware of the presence of unexpected uncertainty and failed to acquire the right model with which to implement Bayesian updating.

Sensor Management: Past, Present, and Future

Abstract: Sensor systems typically operate under resource constraints that prevent the simultaneous use of all resources all of the time. Sensor management becomes relevant when the sensing system has the capability of actively managing these resources; i.e., changing its operating configuration during deployment in reaction to previous measurements. Examples of systems in which sensor management is currently used or is likely to be used in the near future include autonomous robots, surveillance and reconnaissance networks, and waveform-agile radars. This paper provides an overview of the theory, algorithms, and applications of sensor management as it has developed over the past decades and as it stands today.

Reinforcement learning in feedback control

Abstract: Technical process control is a highly interesting area of application serving a high practical impact. Since classical controller design is, in general, a demanding job, this area constitutes a highly attractive domain for the application of learning approaches--in particular, reinforcement learning (RL) methods. RL provides concepts for learning controllers that, by cleverly exploiting information from interactions with the process, can acquire high-quality control behaviour from scratch. This article focuses on the presentation of four typical benchmark problems whilst highlighting important and challenging aspects of technical process control: nonlinear dynamics; varying set-points; long-term dynamic effects; influence of external variables; and the primacy of precision. We propose performance measures for controller quality that apply both to classical control design and learning controllers, measuring precision, speed, and stability of the controller. A second set of key-figures describes the performance from the perspective of a learning approach while providing information about the efficiency of the method with respect to the learning effort needed. For all four benchmark problems, extensive and detailed information is provided with which to carry out the evaluations outlined in this article. A close evaluation of our own RL learning scheme, NFQCA (Neural Fitted Q Iteration with Continuous Actions), in acordance with the proposed scheme on all four benchmarks, thereby provides performance figures on both control quality and learning behavior.

Knows what it knows: a framework for self-aware learning

Abstract: We introduce a learning framework that combines elements of the well-known PAC and mistake-bound models. The KWIK (knows what it knows) framework was designed particularly for its utility in learning settings where active exploration can impact the training examples the learner is exposed to, as is true in reinforcement-learning and active-learning problems. We catalog several KWIK-learnable classes as well as open problems, and demonstrate their applications in experience-efficient reinforcement learning.

Informing sequential clinical decision-making through reinforcement learning: an empirical study

Abstract: This paper highlights the role that reinforcement learning can play in the optimization of treatment policies for chronic illnesses. Before applying any off-the-shelf reinforcement learning methods in this setting, we must first tackle a number of challenges. We outline some of these challenges and present methods for overcoming them. First, we describe a multiple imputation approach to overcome the problem of missing data. Second, we discuss the use of function approximation in the context of a highly variable observation set. Finally, we discuss approaches to summarizing the evidence in the data for recommending a particular action and quantifying the uncertainty around the Q-function of the recommended policy. We present the results of applying these methods to real clinical trial data of patients with schizophrenia.

Learning force control policies for compliant manipulation

Abstract: Developing robots capable of fine manipulation skills is of major importance in order to build truly assistive robots. These robots need to be compliant in their actuation and control in order to operate safely in human environments. Manipulation tasks imply complex contact interactions with the external world, and involve reasoning about the forces and torques to be applied. Planning under contact conditions is usually impractical due to computational complexity, and a lack of precise dynamics models of the environment. We present an approach to acquiring manipulation skills on compliant robots through reinforcement learning. The initial position control policy for manipulation is initialized through kinesthetic demonstration. We augment this policy with a force/torque profile to be controlled in combination with the position trajectories. We use the Policy Improvement with Path Integrals (PI2) algorithm to learn these force/torque profiles by optimizing a cost function that measures task success. We demonstrate our approach on the Barrett WAM robot arm equipped with a 6-DOF force/torque sensor on two different manipulation tasks: opening a door with a lever door handle, and picking up a pen off the table. We show that the learnt force control policies allow successful, robust execution of the tasks.

Online human training of a myoelectric prosthesis controller via actor-critic reinforcement learning

Abstract: As a contribution toward the goal of adaptable, intelligent artificial limbs, this work introduces a continuous actor-critic reinforcement learning method for optimizing the control of multi-function myoelectric devices. Using a simulated upper-arm robotic prosthesis, we demonstrate how it is possible to derive successful limb controllers from myoelectric data using only a sparse human-delivered training signal, without requiring detailed knowledge about the task domain. This reinforcement-based machine learning framework is well suited for use by both patients and clinical staff, and may be easily adapted to different application domains and the needs of individual amputees. To our knowledge, this is the first my-oelectric control approach that facilitates the online learning of new amputee-specific motions based only on a one-dimensional (scalar) feedback signal provided by the user of the prosthesis.

Gradient temporal-difference learning algorithms

Abstract: We present a new family of gradient temporal-difference (TD) learning methods with function approximation whose complexity, both in terms of memory and per-time-step computation, scales linearly with the number of learning parameters. TD methods are powerful prediction techniques, and with function approximation form a core part of modern reinforcement learning (RL). However, the most popular TD methods, such as TD(λ), Q-learning and Sarsa, may become unstable and diverge when combined with function approximation. In particular, convergence cannot be guaranteed for these methods when they are used with off-policy training. Off-policy training—training on data from one policy in order to learn the value of another—is useful in dealing with the exploration-exploitation tradeoff. As function approximation is needed for large-scale applications, this stability problem is a key impediment to extending TD methods to real-world large-scale problems. The new family of TD algorithms, also called gradient-TD methods, are based on stochastic gradient-descent in a Bellman error objective function. We provide convergence proofs for general settings, including off-policy learning with unrestricted features, and nonlinear function approximation. Gradient-TD algorithms are on-line, incremental, and extend conventional TD methods to off-policy learning while retaining a convergence guarantee and only doubling computational requirements. Our empirical results suggest that many members of the gradient-TD algorithms may be slower than conventional TD on the subset of training cases in which conventional TD methods are sound. Our latest gradient-TD algorithms are “hybrid” in that they become equivalent to conventional TD—in terms of asymptotic rate of convergence—in on-policy problems.

Neural Correlates of Forward Planning in a Spatial Decision Task in Humans

Abstract: Although reinforcement learning (RL) theories have been influential in characterizing the mechanisms for reward-guided choice in the brain, the predominant temporal difference (TD) algorithm cannot explain many flexible or goal-directed actions that have been demonstrated behaviorally. We investigate such actions by contrasting an RL algorithm that is model based, in that it relies on learning a map or model of the task and planning within it, to traditional model-free TD learning. To distinguish these approaches in humans, we used functional magnetic resonance imaging in a continuous spatial navigation task, in which frequent changes to the layout of the maze forced subjects continually to relearn their favored routes, thereby exposing the RL mechanisms used. We sought evidence for the neural substrates of such mechanisms by comparing choice behavior and blood oxygen level-dependent (BOLD) signals to decision variables extracted from simulations of either algorithm. Both choices and value-related BOLD signals in striatum, although most often associated with TD learning, were better explained by the model-based theory. Furthermore, predecessor quantities for the model-based value computation were correlated with BOLD signals in the medial temporal lobe and frontal cortex. These results point to a significant extension of both the computational and anatomical substrates for RL in the brain.