|
Acting with street smarts and intuition…quickly noticing when things have changed…combining many sources of information…disambiguating complex, jumbled, and ambiguous information…generalizing from a limited number of experiences—these are signs of real intelligence. How do people do these things, and how could our decision-support systems do the same? Jumbled and ambiguous? My article “Artifacts of the Future” (Decision Support, Feb. 9, 2000) argued much of the information we receive is actually this way. Generalizing from limited experiences? In “Connect the Dots” (Decision Support, Mar. 1, 2000), I discussed how an entity can make predictions about events never seen before. Now I focus on certain kinds of learning laws that can help us begin to solve the rest of these questions. I want to introduce the idea of a keyhole diagram, which will help us understand how an electronic system could come to recognize and then combine essential distinctions.
The Retina
Let us start by looking at one disambiguating circuit designed by the grand genetic algorithm we call evolution — a basic retinal complex called “center on, surround off.” In fact, this solution to the problem of interpreting the visual field has been discovered by nature more than once. It evolved separately in more than one evolutionary branch. We have this circuit, as do squid. It seems this function is an essential first computation in recognizing any object at all.
This circuit disambiguates dots from larger patches; a central cell sends a stimulus signal while surrounding cells send inhibitory signals when exposed to light. These computations result in a way of selecting a specific subset of all objects (here, small dots centered on the center “on” cell in the visual field) as organized by two main types of variables (center intensity and surround intensity). (See Figure 1, page 26.) Appropriately placed on the axes in this diagram are differing exemplary shapes.
FIGURE 1 To see at all, the eye needs first to disambiguate dots from non-dots. A basic solution is found in the “center on, surround off” retinal circuit.
The dotted box represents the keyhole signature of this unit. It represents, in other words, the particular range of shapes that this circuit recognizes. Outside of the box, little firing occurs. Thus, these keyhole diagrams offer a graphical way of displaying a circuit’s I/O response characteristics. For the retinal circuit, these I/O characteristics are mainly fixed during the lifetime of an organism. There is only one kind of forecast. It either is or is not a dot. There is no external feedback or accommodation that makes this signature change. (There are, however, top-down feedback circuits from higher level presuppositions that affect this circuit’s response, but this discussion will have to wait for another time.)
Geometric Logic — Combining Trading Rules
Now let us look at how these diagrams can be used to design algorithms to solve business problems. A very common problem is how to combine multiple indicators into a single action plan. For this objective, the feedback and accommodation to real-world data becomes essential.
Some people who want to weight multiple action indicators say, “just add them.” Some optimizers look for the right linear combination. But the best way to combine indicators can be far subtler.
Consider a simple trading example. Suppose our research has developed two “buy” indicators for a security: x > 0 and y > 0. The green area in Figure 2 geometrically indicates the simple compost rule, buy when x > 0 AND y > 0. The blue area represents x > 0 OR y > 0. This type of illustration, as you can see, becomes a geometric way of representing truth tables. The red area is what a linear optimizer might suggest (with “just add them” as a special case). The area bounded by the heavy black dotted lines is what a decision tree might suggest. (Decision trees require boundaries to be parallel to the axes.)
FIGURE 2 Combining indicators or multiple sources of information can be accomplished with differing degrees of success through logic, linear optimizers, decision trees, or neural nets.
In this diagram, the marks represent a record of feedback information — what actually happened next to the security in the context of the various values for the indicators—“1” for up, “2” for down, and “0” for stayed the same. Given this feedback, you can see a yellow angled ellipse defines the most useful trade area — a better solution than derived by any of the techniques mentioned earlier.
One question you can ask is, what kinds of computations or algorithms would discover this kind of solution. You can further ask, as new data arrives, as the world changes, what would be required for the algorithm to be able to learn more appropriate action rules? Though these sorts of problems can be solved with carefully constructed artificial neural nets, the concept of artificial neural nets is so wide open that you really need to concretely understand the issues in the given context domain to set them up for success. The choice of the learning law is critical for success, not just the architecture of the implementation.
Many groups are attempting to use neural nets to forecast markets. (See Resources for some examples.) But given my knowledge of how easy it is in this domain to perfectly fit past data, display a stellar but artificial historic track record, and yet still have absolutely no forecast power whatsoever, I am keeping my wallet in my pocket. To succeed, you need to have the right methods of finding key trade areas and the right ways of balancing change and stability.
Lobsters vs. Bugs
So how might a learning law be able to adapt to new incoming information? You need a dynamic response system. You need the right modeling form and the right timely re-estimation of parameters.
In “Connect the Dots,” I mentioned the story of two-year-old Lucy, who on seeing lobsters for the first time exclaimed, “Bugs!” Her father’s explanations to her indicated that lobsters are much bigger than bugs. In a keyhole diagram (Figure 3), which obviously simplifies the landscape greatly, we can display size on the y-axis and number of appendages on the x-axis, with some of the various creatures Lucy might have learned about appropriately mapped.
For a child attempting to make sense of this new wiggling, multi-legged creature that daddy is about to put in the pot, in terms of what she has seen before, a very reasonable conclusion would be to “connect the dots” and generalize. Everything to the right of the green line that she remembers is a bug, so why not this lobster?
FIGURE 3 From limited experience a child classifies lobsters as bugs. With more data and resulting new distinctions, a child’s usage begins to converge with adults’.
With feedback from parents and teachers, it may not take very many examples for a child to recognize there must be some boundary and to create a trial boundary using essentially size as the dividing issue (indicated by the dotted red line). This is the mark of a very adaptive learning algorithm. And it is exactly this sort of rapidly accommodating modeling system we need if we are going to successfully quantify, understand, and forecast the rapidly changing market, business, and information environments in which we live.
Maybe it is that parental teachings are given high credibility, with differing sources of information instinctually given different weights. If information we fully believe contradicts a part of our model, we know immediate revision is required — even if we are not sure of the exact implementation. Perhaps additionally it may be the human mind is set to presume opposing language terms exist to identify essential differences. When a new term is used in opposition to an old one, the mind sets out, perhaps even on a pre-conscious level, to look for real, observable correlates to the distinction that language marks. Out of the mouths of babes may come methods for solving big problems.
Removing Meaningless Variables
I’ll briefly mention one example of a generalizing rule: Remove variables that offer no information. More is not necessarily better. For instance, in a noisy bar, we want to be able to ignore the sound of the TV so we can listen to our friends. These are sound differences that make no difference to us. In many statistical problems, the addition of noisy but irrelevant variables can turn your study into a muddle.
There are well understood modulating circuits even in simpler animals that can remove a stimulus from consideration if no meaningful feedback is associated with it over a period of time. Infants at first are very reactive to thuds and thumps, but rapidly learn what they can ignore.
Here are some business applications: A number of interesting knowledge management products can keep track of the kinds of questions an expert has answered or the kind of information a person has looked at before and use this as a context to help individuals later make better contacts, find the right answers, or retrieve the right information. Appropriately implemented systems that keep track of digital personas can greatly advance the development, use, management, and maintenance of corporate memories and knowledge bases.
In a related way, e-commerce companies might keep track of customer statistics so that the right ads and promotions go to the right consumers. This saves a lot of advertising dollars and increases customer responsiveness.
But people change, they gain more expertise, change interests, grow. Will these electronic identities be responsive to change? Or will you never be able to grow out of the persona that you still had when you were “young and foolish?”
Let us suppose John has been an active buyer of products at an online retailer. Up until now the products he has purchased have been for 20- to 35-year-olds, so a marketing algorithm has given him a young yuppie standing. But suppose he gets married to a woman who already has a child and elderly parents. All of a sudden his online account is getting orders for life insurance, strollers, and Geritol. A smart algorithm will rapidly figure there has been a major lifestyle change. A dumb one will take years to average in the new data and incrementally change. Immediately the age range variable (and clearly any gender variable too) is no longer useful in defining John’s purchasing patterns. It even confuses matters and can get in the way of a good analysis with other, now more relevant, measures.
Simplification has tremendous value. If you are making decisions based on invalid, meaningless, or irrelevant actionable indicators, especially ones that can vary rapidly, such as with market conditions, you are going to waste a lot of time and money unnecessarily changing your trading, business, or marketing direction.
So where do street smarts and intuition come from? Perhaps with a great deal of experience, people comprehend many subtle distinctions, with just the right ways of updating and combining them. With this ability, knowing exactly what to expect from limited data becomes far less difficult. I hope I have illustrated in the last two articles that making a forecast model is not a straightforward process. Often you cannot derive a pure answer without substantial experience with the domain in question. Careful reflection on the complexities of limited data and the evolution of the domain over time is necessary for coming up with the right modeling rules.
Perhaps we are even on the cusp of a revolution in ways to automate these far more sophisticated modeling methods as we learn more about the miraculous capabilities of the human mind and their implementation in the brain’s circuitry.
Barry Grushkin (bgrushkin@dsslab.com) is the senior lab researcher at the DSS Lab (www.dsslab.com) founded by Erik Thomsen in Cambridge, Mass.
RESOURCESBUSINESS OPTIMIZATIONDecisioneering Inc.: www.decisioneering.com Hyperwave Information Management Inc.:www.hyperwave.com MCOR-PS Ltd.: www.mcor-ps.com/showcase.html DIGITAL PERSONAS Orbital Software Group Ltd.:www.orbitalsw.com Math Works Inc.:www.mathworks.com/products/finprod/finprev4.jhtml FINANCIAL NEURAL NETS Attrasoft Inc.:www.attrasoft.com/decision Equity Analytics Ltd.:www.e-analytics.com/softdi/soft10d.htm Ward SystemsGroup Inc.:www.wardsystems.com/index.html Lederman, Jess. Virtual Trading: How Any Trader with a PC Can Use the Power of Neural Nets & Expert Systems (McGraw-Hill, 1994) |
|
|
|
|
|