Using Machine Learning to Predict the Effects of Global Warming
Story posted April 26, 2000
One of the most important skills students should take away from Bowdoin is the ability to think critically about what they read, Computer Science Professor Eric Chown told his audience at a recent Faculty Seminar Series lecture, "Using Machine Learning to Predict the Effects of Global Warming."
Global warming may be an especially important topic to think critically about, since the media regularly report conflicting viewpoints on the basic facts of global climate change: if it's happening, how fast, and what’s causing it.
"The battle is far from won," in terms of convincing the general reading public that global warming is happening and that it will have increasingly apparent consequences, according to Chown.
"Computers Model World's Climate, But How Well?" questioned one recent New York Times headline.
When not teaching students how to think critically, Chown spends time teaching computers how to think.
Chown has used two types of computer modeling to predict how various ecosystems will react to global climate change. The type known as the Empirical Model almost teaches the computer to think. Its system of neural networks is based on the structure of the human brain. As the program crunches training data during test runs, it picks up patterns among correlated data.
But the empirical model has some weaknesses when compared to another type, called process models. Chown worked with a process model at Oregon State University called MAPSS: Mapped Atmosphere Plant Soil System.
There are two yardsticks for evaluating the competing models: how accurately does the model describe current climate conditions when it is tested with current data? And how consistent are its predictions of future climate change?
By the second measure, the process models are better. In scenario after scenario, they produce remarkably consistent predictions.
On the accuracy of describing current conditions, process models actually score lower than empirical models, but that will change. As data inputs improve, process models can theoretically achieve accuracy rates of 90%. Empirical models will probably never be accurate more than 75% of the time.
Just because one model predicts a mild set of consequences from global warming doesn't mean it won’t have dire consequences.
"You have to be careful about the use" of various models, Chown told his audience. The models deal with the incredibly complex systems that make up the global environment, and the ways that these systems interact and change one another are incompletely understood. Changing the data parameters of a model even slightly can lead to widely divergent predictions.
"But just because they're imperfect doesn’t mean they’re not useful."
The tide is turning in the public education battle. There is little debate in academia over whether policy makers should respond to possible effects of global climate change, and that consensus is becoming more apparent in the media.
Some naysayers have begun to change their arguments from "It's not really happening" to "It's not happening as fast we thought," or "It's happening but it will have some good consequences."
Like more beachfront property?
"Sure," says Chown."But where is the beachfront going to be?" may be the more important question.
The Faculty Seminar Series is held every Wednesday while classes are in session, at 12:30pm in Main Lounge in the Moulton Union. Lunch is available for $3. The next scheduled lecturer is Katherine Watson, Director Emerita of the Bowdoin College Museum of Art. On Wednesday, May 3, she will discuss the Brunswick architecture of Felix Arnold Burton.
« Back | Campus News | Academic Spotlight | | Subscribe to Bowdoin News by Email