Accessible Technology?
In Barry Grushkin’s May 15, 2000 column (“More for Less”), he makes the technology sound extremely promising. Obviously he works in a research environment. How practical would it be for industry IS people to concoct similar equipment and software setups? Grushkin certainly sounds as if the resulting applications could yield potentially valuable results in any number of arenas. Still, I was perplexed by Figure 2. Did someone switch the horizontal scale so that it read 100 to 0 percent instead of 0 to 100 percent?
If Grushkin has any sort of mailing list for this type of information, I’d like to be on it. I’ve had a feeling for some time that neural nets were likely to develop a new life and his column seems to confirm my hunch.
Has he thought of any possible ways to combine such neural nets with fuzzy logic? The combination might be useful in maneuvering robots through complex real-world physical enviroments.
It also strikes me that this kind of technology might be useful in predictive modeling of the collective behavior of markets (or countries) if we could derive sufficient relevant data to feed the system.
Christian Bain
New York
Barry Grushkin responds:
I think we have seen only the beginning of the application of these complex decision-making methods.
Filling in information and finding organization in what we have created is one of the key problems of perception and spatial negotiation, as well as structured and unstructured data analysis.
The technology represents a solution to this general type of information-processing problem as it arises in such areas as computer vision, knowledge management, and voice recognition. People are already applying a related learning system with academic origins called Artmap (which uses fuzzy relations) in complex 3D-part retrieval, face recognition, ECG wave recognition, chemical analysis, Chinese character recognition, remote sensing, and musical analysis.
My work had its origin in some financial market forecasting software I developed called Reason Technologies. The software calculated the right fuzzy relations of each financial instrument to millions of groupings of other financial instruments. Deviations from the resulting tracking indexes indicated trading opportunities.
Still, we define fuzzy relations essentially via a linear function. One day I asked myself: What if the forecast of, for example, a given stock is in fact a very complex function of all the variables on which it is potentially dependent? Thus arose the idea of Intuition Technologies, in which sets of neural nets could find more complex relations between a stock and all the variables on which it depended. But there was a limit to how many variables we could input. So as in the column “More for Less,” I broke the problem into parts, and then recombined them, using still more neural nets to fashion a more complete solution.
The setup is basically straightforward, though there are a lot of iterations. We could do most of the processes using commercial neural nets. I had reasonable success with off-the-shelf versions of SPSS’s Clementine when I used them with an Oracle database. We can use this idea of using neural nets iteratively to advantage, even for smaller problems. Often, forecasting some of our variables before performing further analysis offers improvements over merely using the raw data.
By the way, we wacky quants sometimes reverse the direction of an axis scale to make a point. In Figure 2 (“More for Less”), we start on the left with 100 percent of connections. Moving to the right (as we read in English) represents the process I performed of hacking away until Igot to no connections whatsoever.
Speaking of McBee Cards...
I’d like to tell a story of how a McBee card (“Tools and Nostalgia,” Intelligent Enterprise, April 28, 2000) gave me a career and profoundly changed my life.
In the early summer of 1945 I was an infantry private in what was to be the U.S. Army of Occupation in Germany. The war in Europe was winding down and it looked like we wouldn’t be needed in the Pacific, either. One afternoon I was called to the orderly tent and introduced to a sergeant from the 36th Machine Records Unit, whatever that was. The sergeant said his unit was passing through our area en route to Heidelberg for permanent station. Redeployment home had begun, and because everyone in his unit had served since the earliest days of the War, they needed replacements. I met the requirements of high school graduate and bookkeeping training.
When he asked if I wanted to transfer, he couldn’t finish the sentence before I agreed.
Arriving at the new unit was my introduction to IBM punched card systems. I liked the “technology” so much that after earning my degree in accounting I dropped the idea of becoming a CPA. Instead, the CIA hired me to convert its payroll and accounting systems to punched cards. Later, working at General Electric, I was part of the team that programmed the first-ever computerized payroll and accounting systems in 1954 (Univac I).
It all started with a McBee card that afternoon so long ago. The sergeant was holding a large card with perforations and notches around its edges as he explained how I was selected. Every unit in the army had a card describing the background of each soldier in the unit. An array of long needles was arranged so that cards notched for high school and bookkeeping training, in my case, fell out when the needles were run through the stack.
I never saw the sergeant again, as he was on his way home when I arrived at the 36th MRU. Wherever he is, he has my thanks.
Kirk Bennett
Forth Worth, Tex.
Correction
In the May 15 Product Review (“EJB Execution”), we inadvertently printed an error in pricing. The correct pricing is as follows:Tentative U.S. Product List Prices: Unify E-Wave Engine Desktop—$300; Unify E-Wave Engine Professional—$8,000 per CPU; Unify E-Wave Engine Enterprise—$10,000 per CPU. (The Unify E-Wave Engine Desktop version, list price $300, is available right now for free at www.ewavecommerce.com.)
