Extended Abstract (180K)P1.1
Teaching Artificial Intelligence to Meteorology Undergraduates
George S. Young, Penn State University, University Park, PA; and S. E. Haupt
Advanced Artificial Intelligence (AI) forecast systems have shown superiority to traditional regression models in a number of research projects. Yet operational use of AI-based forecast systems remains limited in comparison with older regression-based statistical forecast systems. Likewise, development of AI forecasts systems by operational forecasters remains uncommon. Given the potential of AI in operational forecasting why has adoption not been more widespread?
A major factor restricting the adoption of AI systems in operational meteorology is the lack of a pool of forecasters trained in the areas required to use the new systems wisely. The supply of forecasters and researchers trained in the theory and tools required to develop new AI-based forecast systems is even more limited. In order to develop a large enough pool of appropriately trained operational meteorologists one must teach AI at the undergraduate level.
The goal of this project was to develop an undergraduate course that would prepare students to be both users and developers of AI-based forecast systems. Successful users of such systems must understand the strengths and limitations of the various AI methodologies and understand verification well enough to be able to test for themselves the skill of AI methods versus their own forecasts. As the development tools improve system development is opened up to a much wider group of meteorologists. To use these tools wisely the students must know enough theory to select the set of methods that are likely to work well on the problem at hand. Likewise, to fit system development into an already busy schedule operational meteorologists require both training and hands-on experience in a tool set that lets them develop AI forecast systems efficiently and within the limits of their knowledge.
Teaching AI to undergraduate meteorologists is challenging because students at this level are not fluent with all of the tools used by AI researchers. They have generally had calculus and differential equations but lack linear algebra. They are comfortable with descriptive statistics and regression but have little statistical theory. Their computer skills are highly varied, often including programming ability in Fortran, C, or Matlab and familiarity with Excel. They lack knowledge of Prolog and other logic languages. Today’s undergraduates are more comfortable using GUI applications than writing their own code.
These constraints dictate the instructional methods used to prepare our students. The text must cover the AI theory required by users and system developers, but should avoid the depth and complexity required by method developers. The hands on part of the course should revolve around a GUI tool that lets students experiment with a broad range of AI methods to develop their own forecast systems. The tool should take full advantage of, and make high demands on, the students’ knowledge of AI theory to achieve optimal system designs. The lectures component of the course covers AI theory to just the level required to make wise and efficient use of the tool and the forecast systems it produces. The lab / homework projects exercise the students’ knowledge of theory and show them that application of that knowledge results in more successful forecast systems.
Poster Session 1, Poster
Tuesday, 11 January 2005, 9:45 AM-9:45 AM
Next paper