Focus on Knowledge
Knowledge-Enhanced Electronic Logic Programming
Knowledge-Enhanced Electronic Logic Programming
Valley Forge, Pennsylvania earned its reputation for nasty winters during the Revolutionary War. By the time we roll into February around here, the phrase”warm and fuzzy” elicits thoughts of dry mittens, cozy fleece jogging suits, and slippers being heated by the fireplace. During the 2000 U.S. Presidential election, the term “fuzzy” was used as a pejorative to describe the “fuzzy math” used by one of the candidates to predict the effect of tax cuts on the economy — a complex system with many inputs, some of which are hidden or counter-intuitive. While “traditional math” is well-suited for simple or “tame” problems that have a “correct” answer, it is often inadequate for the modeling of “wicked” problems whose “best available” answers lie in a gray area between good and bad.
Photo by Paul Skorupskas on Unsplash
Wicked problems were studied by Professor Horst Rittel at the University of California (UC) — Berkeley in the 1970s, where he developed the Issue-Based Information System (IBIS) to augment the vocabulary of traditional math. IBIS decomposes complex problems into the branches of a decision tree having pro and con arguments as inputs for the coordination and planning of political decisions. When formulated, an IBIS tree looks similar to the layout of an artificial neural network (ANN); the leaves serve as inputs and the final decision as the output. The solution of both the IBIS tree and the ANN requires the inputs to be weighted as a function of their relative importance to each other and to the final decision. The weights connecting the neurons of an ANN are commonly adjusted by presenting the network with a set of matched inputs and outputs in conjunction with a supervised or unsupervised algorithm that trains it to “learn” the relationships between each neuron. The final values of the trained ANN weights are typically incidental as long as the network performs correctly. This training approach does not consider that the relative importance (weights) of issues is often known by the human decision makers. While it is difficult for a human expert to explain exactly why they “feel” one issue is of greater importance over another, they “know” it to be true given their education, experience and knowledge of the situation. A more efficient training algorithm would incorporate this expert heuristic knowledge of the weights.
One method used to connect a collection of known inputs to a known output using heuristically-known weights is the application of classical logic. The weights take the form of True (100%) or False (0%), and their combinations are processed using IF/THEN statements coupled with the logical operators of AND, OR, NOT and XOR. This results in the creation of a von Neumann-style expert system produced by traditional programming languages; however, problems arise when the differences between True and False are not clear or sharp. Professor Lofti Zadeh, also at UC Berkeley, introduced “fuzzy” logic (FL) in 1965, enabling the use of real-valued weights in IF/THEN statements. For instance, if an employee misses one day of work in a 30-day month, an FL system can use the weight of “1/30th-True” for the input named “MISSED WORK.” This allows the system to work with inputs that blur the line between True and False, making them “fuzzy.” While FL continues to be an important paradigm for processing real-valued inputs, the design of the FL algorithms can be tedious. A graph of geometric shapes having known angles and widths must be adjusted to tune each FL algorithm. Human experts often envision the weights of competing inputs to a decision as just that — weights on a balanced scale that tips toward the more important factor, not a complex geometric pattern.
With the desire to utilize the ability of FL to incorporate expert heuristic-knowledge of competing inputs into the creation of layered IBIS decision trees, Compsim has developed a dynamic graphical language called Knowledge Enhanced Electronic Logic (KEEL) that permits the developers to define the number and identity of inputs to a complex decision and, more importantly, adjust the normalized weight of each input by moving a virtual element between the values of 0–100 percent (as in Figure 1). Using KEEL technology, the expert human can supervise the education of the IBIS through direct manipulation of the weights, rather than handing the process over to the blind learning algorithm of an ANN. When a satisfactory IBIS functionality is obtained, all of the settings can be exported to an XML file to be used as documentation for an audit trail, the creation of reusable components, and for importing the values into applications written in traditional programming languages for incorporation into devices and controllers. KEEL technology enables the expert user to focus on the known importance of the input factors to a decision, and to teach an automated system to use the same reasoning. Knowing that an automated controller is using the identical reasoning as my way of thinking gives me a warm and fuzzy feeling.
This material originally appeared as a Contributed Editorial in Scientific Computing 23:3 February 2006, pg. 12.
William L. Weaver is an Associate Professor in the Department of Integrated Science, Business, and Technology at La Salle University in Philadelphia, PA USA. He holds a B.S. Degree with Double Majors in Chemistry and Physics and earned his Ph.D. in Analytical Chemistry with expertise in Ultrafast LASER Spectroscopy. He teaches, writes, and speaks on the application of Systems Thinking to the development of New Products and Innovation.