Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies

ISBN-13: 9780898382136 / Angielski / Twarda / 1986 / 314 str.

John Laird; Paul Rosenbloom; Allen Newell

Opis

Rarely do research paths diverge and converge as neatly and productively as the paths exemplified by the two efforts contained in this book. The story behind these researches is worth recounting. The story, as far as I'm concerned, starts back in the Fall of1976, when John Laird and Paul Rosenbloom, as new graduate students in computer science at Carnegie-Mellon University, joined the Instructible Production System (IPS) project (Rychener, Forgy, Langley, McDermott, Newell, Ramakrishna, 1977; Rychener & Newell, 1978). In those days, production systems were either small or special or both (Newell, 1973; Shortliffe, 1976). Mike Rychener had just completed his thesis (Rychener, 1976), showing how production systems could effectively and perspicuously program the full array of artificial intelligence (AI) systems, by creating versions of Studellt (done in an earlier study, Rychener 1975), EPAM, GPS, King-Pawn-King endgames, a toy-blocks problem solver, and a natural-language input system that connected to the blocks-world system.

Spis treści

I. Universal Subgoaling.- 1. Introduction.- 1.1. A Universal Weak Method.- 1.2. Requirements for Universal Subgoaling.- 1.3. Problem Spaces.- 2. The Soar Achitecture.- 2.1. Architecture Description.- 2.2. Soar as a Production System.- 2.3. Universal Subgoaling in Soar.- 2.4. The Universal Weak Method of Soar.- 2.5. The Rest of Soar.- 2.6. Review of Soar.- 3. Empirical Demonstration.- 3.1. Deliberate Subgoals.- 3.2. Universal Subgoaling.- 3.3. The Weak Methods.- 4. Discussion.- 4.1. Goals.- 4.2. Memory Management.- 4.3. Preferences.- 4.4. Production Systems.- 4.5. Future Work.- 5. Conclusion.- Acknowledgment.- References.- Appendix A. Universal Weak Method.- Appendix B. Weak Method Search Control.- II. The Chunking of Goal Hierarchies.- 1. Introduction.- 2. Practice.- 2.1. The Power Law of Practice.- 2.2. The Chunking Theory of Learning.- 2.3. The Results of Chunking in One Task.- 3. Stimulus-Response Compatibility.- 3.1. The Phenomena.- 3.2. Existing Stimulus-Response Compatibility Theory.- 3.3. The Algorithmic Model of Stimulus-Response Compatibility.- 3.4. Other Subphenomena and Experiments.- 4. Goal-Structured Models.- 4.1. The Basics of Goal Hierarchies.- 4.2. Chunking on Goal Hierarchies.- 4.3. A Revised Analysis of the Chunking Curve.- 5. The Xaps3 Architecture.- 5.1. Working Memory.- 5.2. Production Memory.- 5.3. The Cycle of Execution.- 5.4. Goal Processing.- 5.5. Chunking.- 6. Simulation Results.- 6.1. Update on the Seibel (1963) Task.- 6.2. The Compatibility Hierarchies.- 6.3. Compatibility and Practice.- 7. Discussion.- 7.1. On Choosing a Set of Model Operators.- 7.2. The Rate of Learning.- 7.3. Errors.- 7.4. Other Reaction-Time Phenomena.- 7.5. Relation to Previous Work on Learning Mechanisms.- 7.6. Chunking in More Complex Tasks.- 7.7. Chunking, Learning, and Problem-Space Search.- 8. Conclusion.- Acknowledgment.- References.- III. Towards Chunking As A General Learning Mechanism.- 1. Introduction.- 2. Soar—A General Problem-Solving Architecture.- 3. Chunking in Soar.- 4. Demonstration.- 4.1. Eight Puzzle.- 4.2. Tic-Tac-Toe.- 4.3. R1.- 4.4. Over-generalization.- 5. Conclusion.- Acknowledgment.- References.- Author Index.- I.- II.- III.