Jonathan How, Nicholas Roy, and Jim Hansen Massachusetts Institute of Technology
Accurate predictions of natural phenomena rely on accurate physical models, good computer simulations, and accurate estimates of the current state of the system. For example, predicting storms over California several days in advance requires good models of the weather development and precise and accurate knowledge of weather variables such as temperature and pressure in distributed regions of the Pacific. Acquiring this precise knowledge of the current system state typically requires a large number of sensor measurements, but the quality of the subsequent prediction is a strong function of both the total sensor resources available (number and kind of sensors), and also how the sensors are used (the spatio-temporal distribution of the sensors). Current sensor systems fly a few manned planes with long time horizons, selecting from a handful of possible flight paths. While useful, the resulting measurement strategies are essentially uncoordinated, preprogrammed and not easily modified. There are often inconsistencies between the assumptions in the data assimilation, ensemble construction (Monte Carlo sampling), and targeting methodologies, which lead to sub-optimal plans for data gathering. UAVs have been developed to reduce the costs of manned flights, and we envisage a scenario with multiple UAVs using distributed sensing techniques to significantly reduce the planner response times. Our research will lead to a new framework for coordinating this team of mobile sensing assets that provides more efficient measurement strategies and a more accurate means of capturing spatial correlations in weather system dynamics. We will also demonstrate the importance of using consistent strategies for the data assimilation, ensemble construction, and targeting. The key step in this work will be to exploit the structure inherent in weather system dynamics to develop a unique hybrid central/distributed planning strategy that uses both multi-resolution optimization techniques to solve the global task assignment and reinforcement learning to solve the local task planning problems. The planning algorithms resulting from this work will be applicable to a wide range of systems that also exhibit strong coupling through both the information (measurements taken by one influence the world models of all) and the tasks (standard conflict avoidance).
This research will facilitate a strong linkage between the nonlinear weather prediction and planning/control communities. By synergistically combining technologies from two diverse fields, our research effort within the DDDAS program will develop a new measurement strategy for weather prediction that closely links model predictions and adaptive observations. Our work will extend and apply new coordination techniques for teams of UAVs that have been recently developed for other (mainly DoD) applications. This represents a paradigm shift in active weather sensing models, in that this project will be the first to show how a network of mobile sensors can be used efficiently to optimize the data gathering that drives predictions of natural phenomena such as weather. Our results could lead to substantially better weather predictions, which in the future could save lives and money.
