Seeing the Elephant

As our field evolves, the tools available to us seem to change at a rate faster than the environmental problems we are trying to understand. One of these exciting new tools is a modeling approach based on fuzzy logic: artificial neural networks. Originally devised in the 1970s, programs spawned from this research area have crept into our daily lives and are operating around us continuously. Every time your video camera removes the vibration from a home movie or when you scan text into a word processing program on your computer, a pattern recognition tool based on fuzzy logic is working for you. Yet, the application of these tools to problems we face as environmental engineers seems to be lagging.

I often speak to colleagues and students who do not understand what an artificial neural network is or how they could apply this method in their work. I offer the following analogy for those unfamiliar with these modeling tools. Consider the old Indian story of the four blind men describing an elephant. Each man describes what he can touch directly in front of him; a rope tail, a tree-trunk leg, a slab side, or a snake-like trunk. Each man is correct in his view of the elephant, but each assessment is incomplete. The picture of the elephant is only complete after blending all of the perceptions. If you think of each man as a separate model and the elephant as the environment, then you have a sense of how limited conventional models can be in their description. The models often accurately produce results in a limited manner but do not address the larger picture. Artificial neural networks create a matrix of equations that illustrate the whole picture, in essence, blending the multiple perceptions of the blind men into a whole.

One of the problems that I sense limiting the quick adoption of fuzzy logic tools into the field with which I am most familiar, public health microbiology and surface water quality control, is our reliance upon strict numbers for action. For example, if water quality in Michigan exceeds 300 E. coli $\mathrm{cfu}∕100\mathrm{mL}$ on a single day, beaches will be closed. Yet, anyone who has measured bacteria in surface water knows that the numbers produced by the current analytical techniques are not consistent and have great inherent variability. As noted by the American Public Health Association in Standard Methods for the Examination of Water and Wastewater, microbial counts are not absolute numbers. So why do engineers continue to treat them as such and apply models that require these counts to have precision that they do not have? Perhaps it is time to rethink our entire approach to improving surface water quality.

Instead of relying on a single number for control or the presence of a single organism, we must begin to look for new ways to approach our data that will work with the inherent fuzziness in environmental systems. I think, based upon my experience over the past few years, that the use of artificial neural networks is one way we could have a more complete understanding of the water quality data that we collect. Research intuition led me to believe that there were patterns and trends in collected data, but traditional modeling and statistical methods of analysis fail to clearly illuminate the cause and effect events. Too many of the environmental data have interrelated input variables, skewed data distributions, large ranges and measurement errors, and discontinuous functions that confounded traditional modeling tools. The only modeling tool I found that could begin to address these issues while maintaining pertinent inputs is artificial neural networks; a tool that harkens back to the days of empirical modeling.

Although many deride artificial neural network modeling as a “black box,” engineering has a long tradition of applying empirical findings from “black box” experiments for the betterment of society. Engineers are not scientists; they can not wait for the fundamental relationships to unfold before they take action and often are devising solutions based upon incomplete knowledge. Engineers were treating potable water with chlorine to remove the health risk from waterborne pathogens long before the mechanisms of disinfection for each individual organism were clearly understood. I admit that I do not know the exact underlying fundamental relationships between the multiple inputs fed into the neural network program for the watershed that I am studying. However, I am beginning to understand the multiple combinations of events that create detrimental situations for water quality and, by disassembling the box, understand how to design and implement effective controls. I readily admit that the fuzzy models that I create for my local watershed most likely will not translate to another watershed, but I believe that I can improve local water quality, and as an engineer, I am satisfied with that utility.

So, I urge my fellow engineers and academics not to shy away from a promising approach, just because it is derided as a “black box” by some of their peers. I urge them to add fuzzy logic programming to their tool belts and then take a step back from their narrow perspective on environmental problems until they can start to see the whole elephant. Maybe then we can find new and better ways to improve the quality of our surface water and environment.