Pat Langley

Systematic and nonsystematic search strategies

Abstract In this paper we compare the relative costs of a systematic problem-solving method – depth-first search – and a nonsystematic method – iterative sampling. An average-case analysis reveals that, for a well-specified class of domains, depth-first search always requires less effort when there exists a single solution, and is generally superior on tasks with shallow solution paths and few solutions. In contrast, iterative sampling is superior on tasks with deep solution paths and many solutions. Depth-first search scales better to cases with high branching factors if the number of solutions remains the same, but random search scales better when the density of solutions remains constant. The search costs predicted by the analysis closely fit the costs observed in experiments with artificial search tasks. We also relate iterative sampling to other methods, including iterative broadening, which is based on similar intuitions.

Concept formation in structured domains

Publisher Summary This chapter describes a system that learns concepts in structured domains. Most recent work on unsupervised concept learning has been limited to unstructured domains, in which instances are described by fixed sets of attribute-value pairs. Many domains can be described in this simple language. Frequently, however, instances have some natural structure; objects have components and relations among those components. In such domains, an attribute-value language is inadequate. The chapter describes Labyrinth, an implemented system that induces concepts from structured objects. The Labyrinth is viewed as an approach to incremental concept formation. The goal of incremental concept formation is to find concepts that allow useful predictions from partial information. The Labyrinth can make effective generalizations by using a more powerful structured representation language. It carries out incremental, unsupervised concept learning in structured domains. It learns probabilistic concepts and uses them to make predictions of missing attribute values, components, and relations. It also decomposes objects into sets of components to constrain matching.

A design for the ICARUS architecture

We describe our designs for ICARUS, an integrated architecture for controlling an intelligent agent in a complex physical environment. By navigating between locations and manipulating other objects, the agent influences the world and achieves its goals. The architecture includes components for perceiving the environment, for generating plans to solve problems, and for executing the plans generated. A fourth component manages the agents long-term memory. Our assessment of the design suggests that it is general, versatile, scales well to larger problems, and is consistent with a variety of psychological results.

A Unified Framework for Planning and Learning.

Abstract : In this report, we present a computational framework for planning and learning that is constrained by knowledge of human behavior. We first describe DAEDALUS, a planning system that learns from successful problem-solving traces. The model stores plan knowledge in a probabilistic concept hierarchy, retrieves relevant operators through a process of heuristic classification, organizes search using a flexible version of means-ends analysis, and stores plan knowledge through an incremental process of concept formation. We report experimental studies of DAEDALUS behavior that show learning improves solution quality and reduces search, but that also reveal increased retrieval cost and fewer solved problems. In addition, we find that the model accounts for a variety of qualitative phenomena observed in human problem solving. After this, we present our current designs for ICARUS, an integrated architecture for intelligent agents that extends the ideas in DAEDALUS. This architecture will store entire problem-solving traces in memory, which should support a number of additional capabilities, including the unification of search control knowledge and macro-operators, the interleaving of planning and execution, and the integration of closed-loop and open-loop processing. (AN)

Editorial: Advice to Machine Learning Authors

This editorial contains suggestions to authors of papers in the area of machine learning, although much of it applies to the broader field of artificial intelligence. I have distilled these comments from my five-year experience as an editor of Machine Learning, focusing on problems that tended to recur in different papers. Many comments are slanted toward papers that describe running systems, but others will be useful for different types of papers. Authors should focus on those suggestions relevant to their own research emphasis. I have divided the suggestions into a number of categories, which should be self-explanatory. I expect most readers will agree with many of the points, but undoubtedly some will be more controversial. Despite this, I believe that listing them explicitly in this manner will at least encourage authors to think about the issues before drafting their papers, and thus reduce the need for revisions at later dates.