Draft for ACM SIGPLAN Patterns (Language Trends) 1996 Why GAWK for AI? Ronald P. Loui Most people are surprised when I tell them what language we use in our undergraduate AI programming class. That's understandable. We use GAWK. GAWK, Gnu's version of Aho, Weinberger, and Kernighan's old pattern scanning language isn't even viewed as a programming language by most people. Like PERL and TCL, most prefer to view it as a "scripting language." It has no objects; it is not functional; it does no built-in logic programming. Their surprise turns to puzzlement when I confide that (a) while the students are allowed to use any language they want; (b) with a single exception, the best work consistently results from those working in GAWK. (footnote: The exception was a PASCAL programmer who is now an NSF graduate fellow getting a Ph.D. in mathematics at Harvard.) Programmers in C, C++, and LISP haven't even been close (we have not seen work in PROLOG or JAVA). Why GAWK? There are some quick answers that have to do with the pragmatics of undergraduate programming. Then there are more instructive answers that might be valuable to those who debate programming paradigms or to those who study the history of AI languages. And there are some deep philosophical answers that expose the nature of reasoning and symbolic AI. I think the answers, especially the last ones, can be even more surprising than the observed effectiveness of GAWK for AI. First it must be confessed that PERL programmers can cobble together AI projects well, too. Most of GAWK's attractiveness is reproduced in PERL, and the success of PERL forebodes some of the success of GAWK. Both are powerful string-processing languages that allow the programmer to exploit many of the features of a UNIX environment. Both provide powerful constructions for manipulating a wide variety of data in reasonably efficient ways. Both are interpreted, which can reduce development time. Both have short learning curves. The GAWK manual can be consumed in a single lab session and the language can be mastered by the next morning by the average student. GAWK's automatic initialization, implicit coercion, I/O support and lack of pointers forgive many of the mistakes that young programmers are likely to make. Those who have seen C but not mastered it are happy to see that GAWK retains some of the same sensibilities while adding what must be regarded as spoonsful of syntactic sugar. Some will argue that PERL has superior functionality, but for quick AI applications, the additional functionality is rarely missed. In fact, PERL's terse syntax is not friendly when regular expressions begin to proliferate and strings contain fragments of HTML, WWW addresses, or shell commands. PERL provides new ways of doing things, but not necessarily ways of doing new things. In the end, despite minor difference, both PERL and GAWK minimize programmer time. Neither really provides the programmer the setting in which to worry about minimizing run-time. There are further simple answers. Probably the best is the fact that increasingly, undergraduate AI programming is involving the Web. Oren Etzioni (University of Washington, Seattle) has for a while been arguing that the "softbot" is replacing the mechanical engineers' robot as the most glamorous AI testbed. If the artifact whose behavior needs to be controlled in an intelligent way is the software agent, then a language that is well-suited to controlling the software environment is the appropriate language. That would imply a scripting language. If the robot is KAREL, then the right language is "turn left; turn right." If the robot is Netscape, then the right language is something that can generate "netscape -remote 'openURL(http://cs.wustl.edu/~loui)'" with elan. Of course, there are deeper answers. Jon Bentley found two pearls in GAWK: its regular expressions and its associative arrays. GAWK asks the programmer to use the file system for data organization and the operating system for debugging tools and subroutine libraries. There is no issue of user-interface. This forces the programmer to return to the question of what the program does, not how it looks. There is no time spent programming a binsort when the data can be shipped to /bin/sort in no time. (footnote: I am reminded of my IBM colleague Ben Grosof's advice for Palo Alto: Don't worry about whether it's highway 101 or 280. Don't worry if you have to head south for an entrance to go north. Just get on the highway as quickly as possible.) There are some similarities between GAWK and LISP that are illuminating. Both provided a powerful uniform data structure (the associative array implemented as a hash table for GAWK and the S-expression, or list of lists, for LISP). Both were well-supported in their environments (GAWK being a child of UNIX, and LISP being the heart of lisp machines). Both have trivial syntax and find their power in the programmer's willingness to use the simple blocks to build a complex approach. Deeper still, is the nature of AI programming. AI is about functionality and exploratory programming. It is about bottom-up design and the building of ambitions as greater behaviors can be demonstrated. Woe be to the top-down AI programmer who finds that the bottom-level refinements, "this subroutine parses the sentence," cannot actually be implemented. Woe be to the programmer who perfects the data structures for that heapsort when the whole approach to the high-level problem needs to be rethought, and the code is sent to the junkheap the next day. AI programming requires high-level thinking. There have always been a few gifted programmers who can write high-level programs in assembly language. Most however need the ambient abstraction to have a higher floor. Now for the surprising philosophical answers. First, AI has discovered that brute-force combinatorics, as an approach to generating intelligent behavior, does not often provide the solution. Chess, neural nets, and genetic programming show the limits of brute computation. The alternative is clever program organization. (footnote: One might add that the former are the AI approaches that work, but that is easily dismissed: those are the AI approaches that work in general, precisely because cleverness is problem-specific.) So AI programmers always want to maximize the content of their program, not optimize the efficiency of an approach. They want minds, not insects. Instead of enumerating large search spaces, they define ways of reducing search, ways of bringing different knowledge to the task. A language that maximizes what the programmer can attempt rather than one that provides tremendous control over how to attempt it, will be the AI choice in the end. Second, inference is merely the expansion of notation. No matter whether the logic that underlies an AI program is fuzzy, probabilistic, deontic, defeasible, or deductive, the logic merely defines how strings can be transformed into other strings. A language that provides the best support for string processing in the end provides the best support for logic, for the exploration of various logics, and for most forms of symbolic processing that AI might choose to call "reasoning" instead of "logic." The implication is that PROLOG, which saves the AI programmer from having to write a unifier, saves perhaps two dozen lines of GAWK code at the expense of strongly biasing the logic and representational expressiveness of any approach. I view these last two points as news not only to the programming language community, but also to much of the AI community that has not reflected on the past decade's lessons. In the puny language, GAWK, which Aho, Weinberger, and Kernighan thought not much more important than grep or sed, I find lessons in AI's trends, AI's history, and the foundations of AI. What I have found not only surprising but also hopeful, is that when I have approached the AI people who still enjoy programming, some of them are not the least bit surprised. R. Loui (loui@ai.wustl.edu) is Associate Professor of Computer Science, at Washington University in St. Louis. He has published in AI Journal, Computational Intelligence, ACM SIGART, AI Magazine, AI and Law, the ACM Computing Surveys Symposium on AI, Cognitive Science, Minds and Machines, Journal of Philosophy, and is on this year's program committees for AAAI (National AI conference) and KR (Knowledge Representation and Reasoning).