Learning, planning, and representing knowledge at multiple levels of temporal abstraction are key, longstanding challenges for AI. In this paper we consider how these challenges can be addressed within the mathematical framework of reinforcement learning and Markov decision processes (MDPs). We extend the usual notion of action in this framework to include options---closed-loop policies for taking action over a period of time. Examples of options include picking up an object, going to lunch, and traveling to a distant city, as well as primitive actions such as muscle twitches and joint torques. Overall, we show that options enable temporally abstract knowledge and action to be included in the reinforcement learning framework in a natural and general way. In particular, we show that options may be used interchangeably with primitive actions in planning methods such as dynamic programming and in learning methods such as Q-learning.

This article presents a completely automated approach to generating abstractions for planning. The abstractions are generated using a tractable, domain-independent algorithm whose only input is the definition of a problem to be solved and whose output is an abstraction hierarchy that is tailored to the particular problem. The algorithm generates abstraction hierarchies by dropping literals from the original problem definition. It forms abstractions that satisfy the ordered monotonicity property, which guarantees that the structure of an abstract solution is not changed in the process of refining it. The algorithm for generating abstractions is implemented in a system called alpine, which generates abstractions for a hierarchical version of the prodigy problem solver. The abstractions generated by alpine are tested in multiple domains on large problem sets and are shown to produce shorter solutions with significantly less search than planning without using abstraction. 1 1 ...

A preliminary unedited version of this paper was incorrectly published as part of Volume

Clustering is often used for discovering structure in data. Clustering systems differ in the objective function used to evaluate clustering quality and the control strategy used to search the space of clusterings. Ideally, the search strategy should consistently construct clusterings of high quality, but be computationally inexpensive as well. In general, we cannot have it both ways, but we can partition the search so that a system inexpensively constructs a `tentative' clustering for initial examination, followed by iterative optimization, which continues to search in background for improved clusterings. Given this motivation, we evaluate an inexpensive strategy for creating initial clusterings, coupled with several control strategies for iterative optimization, each of which repeatedly modifies an initial clustering in search of a better one. One of these methods appears novel as an iterative optimization strategy in clustering contexts. Once a clustering has been construct...

This paper presents a method by which a reinforcement learning agent can automatically discover certain types of subgoals online. By creating useful new subgoals while learning, the agent is able to accelerate learning on the current task and to transfer its expertise to other, related tasks through the reuse of its ability to attain subgoals. The agent discovers subgoals based on commonalities across multiple paths to a solution. We cast the task of finding these commonalities as a multiple-instance learning problem and use the concept of diverse density to find solutions. We illustrate this approach using several gridworld tasks. 1.

Reinforcement learning (RL) algorithms have traditionally been thought of as trial and error learning methods that use actual control experience to incrementally improve a control policy. Sutton's DYNA architecture demonstrated that RL algorithms can work as well using simulated experience from an environment model, and that the resulting computation was similar to doing one-step lookahead planning. Inspired by the literature on hierarchical planning, I propose learning a hierarchy of models of the environment that abstract temporal detail as a means of improving the scalability of RL algorithms. I present H-DYNA (Hierarchical DYNA), an extension to Sutton's DYNA architecture that is able to learn such a hierarchy of abstract models. H-DYNA differs from hierarchical planners in two ways: first, the abstract models are learned using experience gained while...

Decision making usually involves choosing among different courses of action over a broad range of time scales. For instance, a person planning a trip to a distant location makes high-level decisions regarding what means of transportation to use, but also chooses low-level actions, such as the movements for getting into a car. The problem of picking an appropriate time scale for reasoning and learning has been explored in artificial intelligence, control theory and robotics. In this dissertation we develop a framework that allows novel solutions to this problem, in the context of Markov Decision Processes (MDPs) and reinforcement learning. In this dissertation, we present a general framework for prediction, control and learning at multipl...

Soar is an architecture for a system that is intended to be capable of general intelligence. Chunking, a simple experience-based learning mechanism, is Soar's only learning mechanism. Chunking creates new items of information, called chunks, based on the results of problem-solving and stores them in the knowledge base. These chunks are accessed and used in appropriate later situations to avoid the problem-solving required to determine them. It is already well-established that chunking improves performance in Soar when viewed in terms of the subproblems required and the number of steps within a subproblem. However, despite the reduction in number of steps, sometimes there may be a severe degradation in the total run time. This problem arises due to expensive chunks, i.e., chunks that require a large amount of effort in accessing them from the knowledge base. They pose a major problem for Soar, since in their presence, no guarantees can be given about Soar's performance.

Proceedings of the Eighth International Workshop on Machine Learning (pp. 203-207). Evanston, IL: Morgan Kaufmann. We discuss an approach to creating new terms during the induction of Horn clauses. The new terms enable a selective form of look-ahead during hill-climbing search. This look-ahead is needed because a conjunction of literals may be useful while each literal individually may not appear to be useful. We exploit knowledge of common patterns of conjunctions to avoid the necessity of testing all pairs of conjunctions. 1

Learning, planning, and representing knowledge at multiple levels of temporal abstraction are key challenges for AI. In this paper we develop an approach to these problems based on the mathematical framework of reinforcement learning and Markov decision processes (MDPs). We extend the usual notion of action to include options—whole courses of behavior that may be temporally extended, stochastic, and contingent on events. Examples of options include picking up an object, going to lunch, and traveling to a distant city, as well as primitive actions such as muscle twitches and joint torques. Options may be given a priori, learned by experience, or both. They may be used interchangeably with actions in a variety of planning and learning methods. The theory of semi-Markov decision processes (SMDPs) can be applied to model the consequences of options and as a basis for planning and learning methods using them. In this paper we develop these connections, building on prior work by Bradtke and Duff (1995), Parr (in prep.) and others. Our main novel results concern the interface between the MDP and SMDP levels of analysis. We show how a set of options can be altered by changing only their termination conditions