One of the most frustrating and at the same time tantalizing things about working in Artificial Intelligence is that you are never done. It is even difficult after working for a long time on a book and finally completing it, to have a sense of accomplishment. It is always clear that there are many problems with what one has written.
It is often less clear just where those problems are. But, it is always certain that as the ideas proposed are worked into computer programs, only their bare bones will remain the same. The substance of one’s idea is quite likely to continue to change as the experimentation points up the initial problems.
When Bob Abe Ison and I finished Scripts, Plans, Goals, and Understanding, we knew that the work was seriously incomplete. Our choice was to wait for it to be complete, which might have taken a very long time, or to publish it then and take the consequences. The consequences of publishing it when we did were quite pleasant actually. Many workers in psychology and other disciplines took our book in the vein that it was intended, as a set of suggestions for how experimentation might proceed.
This current work is intended in a similar vein. It had its origins in the incompleteness of Schank and Abelson. We knew, when we w’rote that book, that we had not provided adequate definitions for some of the notions we were developing. In particular, we knew that the concept of a script, while intuitively attractive, needed a great deal more consideration than we had given it.
The point of that work was that such knowledge structures were a natural part of language understanding. That having been shown, we left the myriad residual issues that followed from that hypothesis to others.
One of those residual issues was memory. We really did not consider the consequences of our claims with respect to memory. Fortunately, others did. I had a number of conversations with Gordon Bower and Ed Smith, two psychologists who worried about the claims we were making. At the same time, I began to get increasingly concerned about the nature of the theory underlying the programs in our own laboratory It seemed that no one really had a good idea about what the theoretical limits to a script were. If it was convenient to do so, a structure was labeled a script.
While these concerns about scripts troubled me, their consequences were not grave. What was more important to me was that a number of the programs that we w’ere developing at Yale seemed to have the problem of being too cumbersome to be of great use Wc had always argued that there are no easy shortcuts in natural language processing.
Understanding requires a great deal of knowledge, but humans do not crumble from the weight of their own information store Automated data bases on the other hand, do get overwhelmed when they know too much We wanted to make sure that any understanding system we developed would be capable of acquiring, storing, and using its knowledge with some ease.
The problems we were developing were encumbered by their knowledge, however. They had no easy means of acquiring it, and storage and retrieval created massive problems Moreover, the programs had no real sense of what they knew. They were like a librarian who can read but has no memory for what he read after a very shot! interval. In other words, they never got smarter.
A program should not be cast in stone, immutable and impossible to affect. It should, if it is to be a good model of a person, be changed by what it reads It should be unpredictable in the sense that it might just know more than you thought it knew. So, in our laboratory, static programs such as FRUMP, SAM, PAM, and POLITICS began to give way to more dynamic programs, such as CYRUS and IPP Our emphasis was on letting the purpose of the program guide what we wanted to get out of that program IPP and CYRUS could surprise us in ways that the others could not. Of course, as any computer scientist knows, such surprises are not always pleasant.
Size: 4 mb
Free download Dynamic Memory : A theory of reminding and learning in computers and people by ROGER C. SCHANK