WHEN HE WAS 28 years old, Joseph J. Heinen was given the awesome, almost bizarre, job of trying to clone two of the top foreign exchange traders at the Security Pacific National Bank.
Using artificial intelligence methods, Mr. Heinen -- now 30 years old -- tried to capture the expertise of the two traders, convert it into computer code, and make that expertise available to many more people throughout the organization. Thus, Security Pacific's two top foreign exchange traders would have been duplicated many times over by ordinary people turned into experts by using their computer's artificial intelligence.
If it had worked, it would have been a major coup for Security Pacific. It would have helped the bank cope with the severe worldwide shortage of ace foreign exchange traders and, in the process, would have driven up the bank's foreign exchange profits, driven down the high cost of expert personnel, and would have made the bank less dependent on a handful of individuals.
But it didn't work. Mr. Heinen found that artificial intelligence had its limits -- at least for the present and at least for foreign exchange. And Mr. Heinsen's experience puts into doubt some of the claims being made about the efficacy of artificial intelligence in the financial field, particularly in trading activities.
Mr. Heinen said the foreign exchange area was chosen initially because it met two critical criteria:
* First, there was a scarcity of "experts" in the field. Foreign exchange was unlike consumer or mortgage lending, for example, where thousands of competent practitioners are available, and where it is relatively easy to develop a large number of people into experts. Foreign exchange traders, of course, trade one currency against another, a fast-moving activity that depends on a seemingly infinite array of political, economic, and climatic developments.
* Second, foreign exchange trading is an area in which large profits could be made on each transaction. In choosing foreign exchange trading as a start, "we felt we could make a lot of money," said Mr. Heinen.
But it did not work out that way. "We found it was impossible because the experts themselves were wrong half the time," said Mr. Heinen, a vice president who is project manager of the expert systems unit within Security Pacific's advanced technology division. Despite their high error rate, the foreign exchange experts make money for the bank because they lose a lot less on their incorrect decisions than they earn on their correct ones, Mr. Heinen said.
Using the artificial intelligence system that Security Pacific developed, ordinary traders failed to match the experts' 50-50 odds of being correct. Instead, they turned out to be right only a third of the time. "I don't think we could get better than that," Mr. Heinen said.
The failure in cloning the foreign exchange experts tells much about artificial intelligence itself, and raises the questions of whether it actually is "intelligence."
It depends, of course, on how "intelligence" is defined. What makes human thought processes different from computer processes is that humans can make huge leaps in their thinking.
Take the task of capitalizing words. With little conscious effort most people can quickly deal with capitalization and can recognize a word even if the capitalization is incorrect. If a person sees the word "father" spelled "fatHer," he quickly realizes the error and understands the word.
For the computer it is different. A computer can be programmed to deal with a problem of that sort, but it requires a substantial amount of time to instruct it to do so. And if that type of problem were multiplied many times, such a system would become extremely costly to produce -- if it could be produced at all. As the problem becomes more complex the computer becomes increasingly limited, and less and …