Sunday, November 10, 2013

My Interest in Artificial Intelligence and Expert Systems

It's funny how long I have been interested in some things. I started questioning, reading, and investigating the purpose of life in 8th grade, 1967. It was the same year I took my first computer programming class(1). And it was the year I wrote my first AI "conversational" computer program(2). I was heavily influenced by my love of science fiction, which was replete with robots. No doubt I was subliminally influenced, too, by my father's work and his co-workers at Carnegie-Mellon (then called Carnegie Tech). He and Herbert Simon published the book Organizations in 1958. I was four years old.

My interests were further fed and nurtured by the work being done at Stanford. I was there from 1976 to 1979 getting a bachelor's degree in economics and a master's degree in statistics. I also took many courses in the electrical engineering and computer science departments. I even turned down a job interview with Zilog Corporation after I toured their facilities with my digital electronics classmates. In those days, if you could say "555 chip," you were hired.

It wasn't until several years later, after graduating and spending a couple years forecasting California electric and gas consumption and inflation rates, that I got a job working for Crocker National Bank. I was the second statistician hired (Ron Osborn was the first) in a department that had only started a few years before at the insistence and under the guidance of Muzhir Gailani, Crocker's General Auditor.

I spent a few years learning the loan business and estimating lending risk, predicting loses, and recommending funding of the loan loss reserve account with monthly reports to the Board of Directors. It was my frustration at the failure of statistical modeling that pushed me into expert systems. Although statistical analysis was very successful at predicting consumer loan portfolio performance (a la Fair Isaac aka FICO score), the same methods failed miserably in the commercial loan portfolio. I searched for other methodologies and found expert systems.

At the same time, I became fascinated with a then-new non-parametric statistical technique called CART (Classification and Regression Trees) developed by Berkeley and Stanford professors Breiman, Friedman, Olshen, and Stone. I decided to merge the two technologies, expert systems and CART, and did a study of loan-making decisions that used CART on a data-set of decisions made by acknowledged loan experts. From the CART results, I built an expert system for evaluating business loan risk for "middle-market" companies with sales between $10 million and $500 million per year.

The work was so successful that we were planning on using it to do more statistical decision-making modeling, but fate had something else in store for me. Crocker Bank was bought by Wells Fargo. The work I was doing (along with the seven people I was supervising) was not of any value to Wells. I was fired and Portfolio Analysis was disbanded(3).

So that's when I figured I was being given a not-so-friendly push to go work for myself. I started a business with Wayne Johnson: Wally Industries dba WJM Technologies. We built expert systems for the banking industry. Though most of our great ideas fell on deaf ears, one, fraud detection of new customer accounts, was bought, first by Security Pacific Bank, then by none other than Wells Fargo!

Within four years our specialized hybrid expert system, known as Early Warning, was as good as anyone with 2-year's of fraud detection experience. Within six years our model was beating most of the experts. By 12 years, our system had been integrated into 90% of all risk decisions concerning new customer accounts. Early Warning had woven itself into the fabric of bank new account risk evaluation. But we long ago stopped calling it an expert system. It was a powerful, very intelligent, very fast, very knowledgeable tool that augmented the ability of its human handlers to catch crooks.

Through all this time, now 40-plus years, I have pondered the question of why we are here and what, if anything, that has to do with intelligence. Here are some questions I still find interesting: Is human intelligence the only kind of intelligence? Is intelligence generic? Is it stable and recognizable across cultures, planets, galaxies? Is there a definition of intelligence that would be universal in the same way as the laws of physics and chemistry? If there is such a universal definition of intelligence, is there any reason to believe that the human brain would be the best way to be intelligent? Is there any reason to believe that the human brain is even very good at intelligence? Wouldn't it be more likely that there are better implementations of intelligence than the one we call the human brain, in the same way that machines have replaced most physical labor?

Don't worry about right answers. The questions are where the action is! And don't forget to have fun. Play is a very successful way to increase innovation. Play reduces the perceived risk of failure and its associated perceived cost of being wrong. (Beware of ego über alles!)

Notes

(1) We were told it was the first time in the country that 8th-graders were being taught computer programming, something thought to be perhaps too complicated for 13-year-olds. We started on an Olivetti 101 and moved up to programming in Quicktran using a Teletype Model 33 ASR teleprinter, with tape reader and punch, transmitting to a remote time-share computer at 100 baud, a blazing 10 characters per second! Got my mug, along with my "El Scepter" class teammates, Howard Royster and Bob Davis on the front page of the Daily Pilot, too!

(2) Don't get excited. It was just a bunch of "if-then-else" statements driven by answers to questions asked by the computer so that it seemed like the computer was having a conversation with you. The program included the computer asking "you got any pot?" Hey, it was the '60's.

(3) This was in stark contrast to the promotion and "blank check" I was given when Crocker was earlier rescued by Midland Bank of England. When Midland came in and saw how our forecasts of Crocker's losses were the closest of any other forecasts in the bank (even though we only forecasted half the losses Crocker experienced), I was promoted to Vice President and Manager of Portfolio Analysis and given four more positions to fill, doubling the size of our team.