Lead Investigator:

Miodrag Potkonjak

Overview

Our goal is development of conceptually new approach for computational sensing. Computational sensing is emerging scientific and engineering field that can be defined as process of extraction, analysis and use of knowledge about the instrumented environment and sensed phenomena. Examples include event and anomaly detection, trends tracking, reverse engineering of environment and physical, chemical, and biological laws, identification of actual and virtual sources of excitation, hypothesis checking-driven deployment, causality identification, and generalization of knowledge obtain in multiple environments. Sensing is not just the ultimate objective of embedded sensor networks, but also powerful enabler and facilitator for many sensor network tasks such as deployment and sampling rate determination.

We are currently building foundations for addressing sensing problems by coordinated formulation and use of experiments, modeling, simulation, and decision making procedures. Coordinated development and use of non-parametric statistical models, combinatorial and continuous optimization algorithms, and mathematical logic will provide techniques and tools for creating and evaluating testing sensing procedures. Our immediate specific targets are rain and wind identification and prediction, hypothesis checking-driven deployment, and soil structure reverse engineering.

Approach

We propose introduction of embedded sensing models where each sensor in mapped to a node of a graph embedded in N-dimensional space in such a way that the distance between to nodes is proportional to ability to predict corresponding sensors from each other. The graph is dynamic in a sense that as prediction accuracy is changing, the structure and topology of the embedding is altered. We plan to conduct embedding using our already developed location discovery procedures. The number of used dimensions, N, will be selected in such way that discrepancy between the measured and imposed distances that correspond to prediction errors is minimized. Note that we can use information about sensors of different modalities and can even incorporate the information about the quality of multi-hop wireless communication. The embedded graph can rapidly answer many questions including ones about obstacles and location of virtual sources of excitations and correlation between sensors of different modalities. We plan to develop algorithms and software that will answer common and important queries about sensed phenomenon and the instrumented environment.

The traditional way to identify important variables (that correspond to sensors in embedded sensor networks) is the application of principal component analysis. However, principal component analysis is looking for linear relationship between variables and minimizes a specific error norm. In order to overcome these limitations, more recently proposed independent component analysis (ICA) has been proposed. Nevertheless, ICA is still subject to strong set of assumption. In order to develop a non-parametric for identifying essential variables and discovery mutual relationship between the variables, we propose to employ powerful optimal and heuristic techniques. Specifically, we plan to create combinatorial independent component analysis (CICA) approach for discovery of essential variables and most accurate relationship between variables.

Finally, we have been developing ideal types-based simulation. The idea is to conduct a large number of simulations of different soil structures. Each structure will be subject to many different initial conditions and patterns of watering. Using statistical methods, our goal is to find a small number of structures and initial conditions such that any other soil structure and initial conditions are behaving similarly under all patterns of watering. These idea types of structures are consequently used as starting pints for identifying actual structure by comparing their measured values with the corresponding values in simulations of the initial ideal candidates (types).

Systems/Experiments

The main task is to build link between simulation and a few already collected or currently collected data sets. Specific target is soil contamination project. The practical logistic objective is to enable and demonstrate quantitatively better deployment of CENS systems developed by other projects.

Accomplishments

We have developed several new classification, regression, density estimation techniques that satisfy constraints such as monotonicity, unimodularity, symmetry and transitivity. We have also developed several robust algorithms for optimization in presence of different forms of uncertainty.

Future Directions

We plan to address the following computational sensing problems.

• Virtual Sources Identification. We aim to develop mobility and trajectory models for common sources of excitation such as sun and surrounding environment.
• Global and Local Fusion and Predictions. The goal is to combine global information about global environment that is available through the Internet and locally measure data of the same or different modality to predict events in the instrumented world that depend on global behavior of phenomenon of interest.
• Semantic Sensor Compression where the objective is to lossy compress data in such a way to no specific error is minimized, but the information of events of interest is maximally preserved.
• Hypothesis Checking-based Deployment. The goal is to figure out how to deploy sensors in such a way that a specific hypothesis about phenomenon or events is best evaluated.
• Causality Identification aim to identify under which conditions summer event will happen. For this task we plan to combine mathematical logic and statistical models.
• Extracting Knowledge from multiple experiments (datasets). Our final model is to extract knowledge about events and phenomena that are common to different environments that are instrumented in different ways.

People

Faculty

• Miodrag Potkonjak, Computer Science Department, UCLA

Graduate Students:

• Jessica Feng, Computer Science Department, UCLA
• Vishwa Goudar, Computer Science Department, UCLA
• Jen-Ru Lai, Computer Science Department, UCLA
• Kannika Sikangwan, Computer Science Department, UCLA
• Jennifer L. Wong, Computer Science Department, UCLA

External Research

• Dr. Lewis Girod, MIT (current)
• Dr. Negar Kiyavash, Univeristy of Illinois, Urbana-Champaign  (current)
• Prof. Farinaz Koushanfar, Rice University  (current)
• Prof. Seapahn Megerian, University of Wisconsin  (current)
• Dr. Nina Taft, Intel Research Lab  (current)
• Prof. Jennifer L. Wong, SUNY Stony Brook  (current)