This paper surveys the field of reinforcement learning from a computer-science perspective. It is written to be accessible to researchers familiar with machine learning. Both the historical basis of the field and a broad selection of current work are summarized. Reinforcement learning is the problem faced by an agent that learns behavior through trial-and-error interactions with a dynamic environment. The work described here has a resemblance to work in psychology, but differs considerably in the details and in the use of the word "reinforcement." The paper discusses central issues of reinforcement learning, including trading off exploration and exploitation, establishing the foundations of the field via Markov decision theory, learning from delayed reinforcement, constructing empirical models to accelerate learning, making use of generalization and hierarchy, and coping with hidden state. It concludes with a survey of some implemented systems and an assessment of the practical utility of current methods for reinforcement learning.

Abstract not found

Planning under uncertainty is a central problem in the study of automated sequential decision making, and has been addressed by researchers in many different fields, including AI planning, decision analysis, operations research, control theory and economics. While the assumptions and perspectives adopted in these areas often differ in substantial ways, many planning problems of interest to researchers in these fields can be modeled as Markov decision processes (MDPs) and analyzed using the techniques of decision theory. This paper presents an overview and synthesis of MDP-related methods, showing how they provide a unifying framework for modeling many classes of planning problems studied in AI. It also describes structural properties of MDPs that, when exhibited by particular classes of problems, can be exploited in the construction of optimal or approximately optimal policies or plans. Planning problems commonly possess structure in the reward and value functions used to de...

Markov decision processes (MDPs) have recently been applied to the problem of modeling decisiontheoretic planning. While traditional methods for solving MDPs are often practical for small states spaces, their effectiveness for large AI planning problems is questionable. We present an algorithm, called structured policy iteration (SPI), that constructs optimal policies without explicit enumeration of the state space. The algorithm retains the fundamental computational steps of the commonly used modified policy iteration algorithm, but exploitsthe variable and propositionalindependencies reflected in a temporal Bayesian network representation of MDPs. The principles behind SPI can be applied to any structured representation of stochastic actions, policies and value functions, and the algorithm itself can be used in conjunction with recent approximation methods. 1

Markov decision processes(MDPs) have proven to be popular models for decision-theoretic planning, but standard dynamic programming algorithms for solving MDPs rely on explicit, state-based specifications and computations. To alleviate the combinatorial problems associated with such methods, we propose new representational and computational techniques for MDPs that exploit certain types of problem structure. We use dynamic Bayesian networks (with decision trees representing the local families of conditional probability distributions) to represent stochastic actions in an MDP, together with a decision-tree representation of rewards. Based on this representation, we develop versions of standard dynamic programming algorithms that directly manipulate decision-tree representations of policies and value functions. This generally obviates the need for state-by-state computation, aggregating states at the leaves of these trees and requiring computations only for each aggregate state. The key to these algorithms is a decision-theoretic generalization of classic regression analysis, in which we determine the features relevant to predicting expected value. We demonstrate the method empirically on several planning problems,

Markov decision problems (MDPs) provide the foundations for a number of problems of interest to AI researchers studying automated planning and reinforcement learning. In this paper, we summarize results regarding the complexity of solving MDPs and the running time of MDP solution algorithms. We argue that, although MDPs can be solved efficiently in theory, more study is needed to reveal practical algorithms for solving large problems quickly. To encourage future research, we sketch some alternative methods of analysis that rely on the structure of MDPs.

There has been a growing interest in AI in the design of multiagent systems, especially in multiagent cooperative planning. In this paper, we investigate the extent to which methods from single-agent planning and learning can be applied in multiagent settings. We survey a number of different techniques from decision-theoretic planning and reinforcement learning and describe a number of interesting issues that arise with regard to coordinating the policies of individual agents. To this end, we describe multiagent Markov decision processes as a general model in which to frame this discussion. These are special n-person cooperative games in which agents share the same utility function. We discuss coordination mechanisms based on imposed conventions (or social laws) as well as learning methods for coordination. Our focus is on the decomposition of sequential decision processes so that coordination can be learned (or imposed) locally, at the level of individual states. We also discuss the use of structured problem representations and their role in the generalization of learned conventions and in approximation. 1

ion and Approximate Decision Theoretic Planning Richard Dearden and Craig Boutilier y Department of Computer Science University of British Columbia Vancouver, British Columbia CANADA, V6T 1Z4 email: dearden,cebly@cs.ubc.ca Abstract Markov decision processes (MDPs) have recently been proposed as useful conceptual models for understanding decision-theoretic planning. However, the utility of the associated computational methods remains open to question: most algorithms for computing optimal policies require explicit enumeration of the state space of the planning problem. We propose an abstraction technique for MDPs that allows approximately optimal solutions to be computed quickly. Abstractions are generated automatically, using an intensional representation of the planning problem (probabilistic strips rules) to determine the most relevant problem features and optimally solving a reduced problem based on these relevant features. The key features of our method are: abstractions can ...

This dissertation is about building learning control architectures for agents embedded in finite, stationary, and Markovian environments. Such architectures give embedded agents the ability to improve autonomously the efficiency with which they can achieve goals. Machine learning researchers have developed reinforcement learning (RL) algorithms based on dynamic programming (DP) that use the agent's experience in its environment to improve its decision policy incrementally. This is achieved by adapting an evaluation function in such a way that the decision policy that is "greedy" with respect to it improves with experience. This dissertation focuses on finite, stationary and Markovian environments for two reasons: it allows the develop...

This dissertation is submitted for consideration for the dwree of Doctor' of Philosophy at the Uziver'sity of Cambr'idge Summary This thesis is concerned with practical issues surrounding the application of reinforcement lear'ning techniques to tasks that take place in high dimensional continuous state-space environments. In particular, the extension of on-line updating methods is considered, where the term implies systems that learn as each experience arrives, rather than storing the experiences for use in a separate off-line learning phase. Firstly, the use of alternative update rules in place of standard Q-learning (Watkins 1989) is examined to provide faster convergence rates. Secondly, the use of multi-layer perceptton (MLP) neural networks (Rumelhart, Hinton and Williams 1986) is investigated to provide suitable generalising function approximators. Finally, consideration is given to the combination of Adaptive Heuristic Critic (AHC) methods and Q-learning to produce systems combining the benefits of real-valued actions and discrete switching