How Good is Bad Weather?

Accurately identifying key parameters in complex systems demands sufficient excitation, so that the resulting data will be informative enough to reveal hidden parameter values. In many situations, however, users choose inputs that attempt to optimize the system response, not necessarily those that yield more informative data. This leads to the classic trade-off between exploitation and exploration in learning problems. Farmers face a similar issue. Although they would like to identify key soil parameters affecting the growth of their crops, market pressures force them to manage their product to maximize yield, resulting in less informative data. This suggests that weather, and bad weather in particular, may play a critically important role in creating informative data for crop systems by driving them into low-yield regimes that no farmer would otherwise choose to explore. This paper investigates these issues using a standard computational model for corn and real weather data. Two model-based measures characterizing any year's weather pattern are introduced. The first measure characterizes how well a particular year's weather pattern produces corn, according to the model. The second measure characterizes how well a particular year's weather pattern distinguishes the way different soil types affect corn growth. We then use these measures to show that, from the perspective of corn, bad weather can indeed be very good for distinguishing soil type.