Information Theoretic Control

Seeking Information using Decentralized Optimization and Particle Filters

Mobile sensor networks can be deployed to efficiently acquire information about the world, with the ability to make simultaneous measurements from multiple vantage points. To automate such a sensor network, the control objective is to search for information quickly, safely, and reliably. We are developing methods to meet this control objective using a probabilistic foundation. Given a particular configuration of sensors, these techniques exploit the structure of the probability distributions of the target state and of the sensor measurements to compute control inputs leading to future observations that minimize, in expectation, the future uncertainty of the target state.

These methods directly use particle filters, enabling them to more accurately capturing the potentially available information than existing linear and Gaussian-based mutual information methods. A new approximation is used that allows the full network's information objective function to be computed in time that is polynomial in the number of sensors. Thus, as more vehicles are added to the network, the algorithm is able to optimize the use of the additional resources.

Simulation results for three sensor types follow, using bearings-only, range-only, and rescue beacon sensors. In all cases, the mobile sensors are modeled as quadrotor helicopters, such as the STARMAC vehicle shown at right, although the techniques are not limited to any one type of vehicle. More details can be found in "Mobile Sensor Network Control using Mutual Information Methods and Particle Filters".

Control with Bearings-Only Sensors

Bearings-only sensors, such as cameras and directional antennas, measure the direction of the target. The particle filter, handles the nonlinear mapping between measurements and potential locations of the search target. The decentralized control algorithm allows the optimization to run online. The target is known to be in a 40 by 40 square search region. The green X is the MMSE estimate of the location from the particle filter.

Control with Range-Only Sensors

Range-only sensors, using techniques such as signal strength measurements and time-of-flight measurements, measure the distance to the target. The particle filter aptly handles the nonlinearity of the sensors, and the decentralized optimization uses the model of available information computed directly from the particle set.

Control with Rescue Beacon Sensors

To rescue a victim buried in snow due to an avalanche, neither bearing nor range measurements are typically available. Rather, when the victim is carrying a standard avalanche rescue beacon, which contains a modulated magnetic dipole that can be measured using beacon receivers. The magnetic field penetrates snow and water to maximize range, but is absorbed by rock, to prevent measurements of reflections. The particle filter handles the estimation problem, estimating both position and orientation, over a periodic space with high uncertainty. The information seeking algorithm uses the resulting particle filter distribution to extract information from the surroundings as quickly as possible.