A Decision Algorithm to Improve Data Collection in Stochastic Environments
Open Access
- Author:
- Stefik, Jason Brian
- Graduate Program:
- Meteorology
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 14, 2010
- Committee Members:
- Johannes Verlinde, Thesis Advisor/Co-Advisor
Johannes Verlinde, Thesis Advisor/Co-Advisor
Arthur A Small Iii, Thesis Advisor/Co-Advisor - Keywords:
- optimized decision making
self-organizing maps
dynamic programming
resource allocation
field experiments - Abstract:
- Field experiments are conducted to collect observations of a particular meteorological condition or phenomena in order to enhance scientific understanding. In some cases, the decision to deploy costly data-collecting resources must be made ahead of time, when the presence of the sought after condition is unknown, resulting in decisions being made from imperfect forecasts. Traditionally, forecasters are used to predict the likelihood of good data-collecting conditions existing, while a group of scientists use this information to make resource deployment decisions. A new method for resource deployment decisions is presented that shifts the emphasis from the forecasts to the decision-making aspect of the problem. The performance of this new method is evaluated through using the Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) RACORO field campaign, which sought to obtain aircraft measurement of boundary layer clouds (BLCs). Dynamic programming is used to quantify the expected number of successful deployments yet to be launched for any combination of days and resources remaining in the field season. For a given forecast of boundary layer clouds existing a decision can be made which maximizes the expected number of successful deployments. To create BLC forecasts Self-Organizing Maps (SOMs) are used to cluster days according to relative humidity profile. Using cloud data available from the experiment site, the likelihood of BLCs existing for each cluster is determined. A numerical weather prediction model predicts the cluster that the relative humidity profile will belong to, from which the probability that good data-collecting conditions will exist is derived. If the presented, alternative method had been implemented for RACORO, 34 successful flights would have been launched, as opposed to the 28 successful flights that were actually launched. In addition to this 20% increase in the amount of data collected, the new method can operate efficiently in real time, significantly saving scientists time spent on decision-making and eliminating forecasting costs.