Invited Talks

In this talk I will discuss the role of common domains in the performance and evaluation of research in AI. Many of my examples will come from the search literature, where common domains include Chess, the Fifteen Puzzle, the N-Queens Problem, the Travelling Salesman Problem, and random boolean satisfiability problems. I will argue that the primary benefits of common domains are that they facilitate reproducible results, the comparison of results obtained by different researchers, and provide an overall measure of progress in the field, or at least progress in individual domains. The important properties of a good domain are that it be simple and compelling, easy to implement, yet difficult to solve. An ideal domain should be complex enough to study the issues of interest, but no more complex. Usually, this involves clever abstraction of a real-world problem down to its essential difficulties. Common domains should not be artificially imposed from the outside, but should arise from the research community, and gain acceptance on their own merits, by researchers choosing to adopt them for their experiments. The introduction of a new domain that catches on is often a very significant research accomplishment in its own right. We suggest several new candidates for common research domains. Finally, we propose that the class of NP-complete problems represent a large and fertile collection of common domains that are easily implemented, yet capture the essence of many computationally difficult tasks. Simplicity~ Design, and Learning Systems Steven N. Minton Recom Technologies NASA Ames Research Center Mail Stop 269-2 Moffett Field, CA 94035 [email protected] Abstract It is a myth that real world problems are necessarily complex. While it is true that current AI techniques are limited in scope and often scale poorly, it is nevertheless also true that simple techniques are sufficient for many purposes. However, even if a simple technique will solve a problem, it may be difficult to recognize this! In some cases, even experts may not see it. One example is the ttubble Space Telescope Long-Term Scheduling problem where an extremely simple iterative repair technique turned out to be a good solution, but it took some work before we realized this. In other cases, it may be apparent to even non-experts that simple solutions are possible, but frustratingly, it may not be equally "apparent" to our AI systems. For instance, in many toy domains like the blocksworld, simple algorithms are sufficient, but most planners can be easily lead astray if the "wrong" problem representation is used. In this talk I’ll discuss methods for designing systems so that simple solutions, if they exist, can be recognized. For instance, the Multi-TAC constraint satisfaction system is designed so that "small" search control rules are sufficient for many problems; thus, a brute-force smallest-first learning method produces good results. In fact, Multi-TAC embodies the idea of "designing for simplicity" in several respects. I will describe both the general idea and the variety of ways that it manifests itself in Multi-TAC. From: AAAI Technical Report FS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.It is a myth that real world problems are necessarily complex. While it is true that current AI techniques are limited in scope and often scale poorly, it is nevertheless also true that simple techniques are sufficient for many purposes. However, even if a simple technique will solve a problem, it may be difficult to recognize this! In some cases, even experts may not see it. One example is the ttubble Space Telescope Long-Term Scheduling problem where an extremely simple iterative repair technique turned out to be a good solution, but it took some work before we realized this. In other cases, it may be apparent to even non-experts that simple solutions are possible, but frustratingly, it may not be equally "apparent" to our AI systems. For instance, in many toy domains like the blocksworld, simple algorithms are sufficient, but most planners can be easily lead astray if the "wrong" problem representation is used. In this talk I’ll discuss methods for designing systems so that simple solutions, if they exist, can be recognized. For instance, the Multi-TAC constraint satisfaction system is designed so that "small" search control rules are sufficient for many problems; thus, a brute-force smallest-first learning method produces good results. In fact, Multi-TAC embodies the idea of "designing for simplicity" in several respects. I will describe both the general idea and the variety of ways that it manifests itself in Multi-TAC. From: AAAI Technical Report FS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.