Research Interests
My main research interests are in the
areas of wireless sensor networks, distributed and intelligent systems,
nonlinear optimization, bioinformatics, and information analysis.
Wireless sensor networks and
applications
The emerging field of wireless sensor networks (WSNs) offers new computation and communication platforms
that enable novel, low cost, high volume applications. Significant efforts
have been made in recent years in building large-scale wireless sensor networks
from small devices with integrated sensing, computation, and wireless
communication capabilities. There are exciting opportunities and challenges in
developing system support that optimally manages the tremendous hardware
resources in WSNs. In our research, we have been
developing new methods for node and target localization, network and system deployment,
data-centric routing, topology and density control, resource management, mobile
sensor networks, and various applications.
As an example, we have
developed several state-of-the-art methods distributed and adaptive localization methods for ad hoc and wireless
sensor networks. In applying multidimensional
scaling (MDS) algorithms to localization, we have investigated several
variants of MDS and compare the results of various metric scaling and
non-metric scaling methods. In addition, we have systematically
investigated representative localization methods, including Ad-hoc Positioning
System (APS), multidimensional scaling based (MDS-MAP), and semidefinite
programming based (SDP) and studied important factors that affect their localization
results. We have proposed several
algorithms that represent different approaches to adaptive localization. We have developed a network-aware positioning
(NAP) algorithm that can adaptive at network level and a map-based adaptive positioning
(MAP) algorithm that can adapt to different space partitions.
We have developed WSNs for various applications. We applied WSNs with intermittent connectivity to applications in
large geographical areas, such as using WSNs for
search and rescue of lost hikers in trails. A key issue we addressed is the
optimal placement of sensors and access points such that the cost of search and
rescue is minimized. The problem is
divided into how to identify the lost hiker position as accurately as possible
and how to search efficiently in trail segments for different trail topologies
and search agent capabilities. We have
formulated several theoretical models that consider both efficiency and
accuracy criteria and developed efficient heuristic algorithms for various
trail topologies. After access
point deployment is decided, the actual cost of search in individual trail
segment can be computed. We have analyzed
different types of search and rescue agents, presented algorithms to find the
optimal search paths for each one of them, and computed their search costs. In another project, we have applied WSNs to improve the performance of pursuers in
pursuit-evasion games. We have
developed several WSN-assisted pursuer systems and investigated pursuer
performance under different conditions. We have developed several new methods to
address issues such as how to identify evader moving patterns, how to predict
the evader locations using different evader moving models, and how to determine
the most efficient pursuit strategies.
Recently, we are actively
working on intelligent systems based on WSNs for traffic
monitoring and real-time adaptive control. Improving the 
transportation
system’s efficiency is a national priority in the face of energy shortage
and global warming. In intelligent transportation
systems, many solutions have been proposed to solve traffic jams. Low-cost wireless traffic sensors have
the potential to establish a new era of traffic monitoring and control. In this
project, we investigate real-time adaptive traffic light control using WSNs to improve traffic flow and developed optimal sensor
deployment and selection methods to maximum sensor and network lifetime while
achieving accurate traffic information. We develop decentralized traffic light
control methods to take advantage of real-time data of nearby sensor nodes and
offline traffic models. We develop distributed coordinated traffic light
control methods to further improve the traffic over larger geographic regions.
Many graduate and
undergraduate students have been involved in our research on wireless sensor
networks. Here are examples of student projects.
·
Monitoring people traffic. This project aims at
developing a human traffic monitoring system using wireless sensor networks. This system 
is able to count the number
as well as the direction of movement of people and has many potential
applications such as reporting room occupancy and usage. Panasonic passive infrared (PIR) sensors
are used for motion detection because of its low energy consumption. By mounting two PIR sensors in a row and
connecting them to a MicaZ mote through the 51-pin
connector, the motion direction is estimated based on the time difference of
detections. Several methods have
been developed to eliminate the interference among different ADC channels of MicaZ and to reduce missing and false detection. The TPSN synchronization algorithm is implemented
to synchronize the clocks of different motes and the base station. Detection events are logged into the
local memory of a mote. Aggregated
data such as the count for each direction within a certain time interval is
sent to the base station at a low frequency to reduce communication energy
consumption. Experiments with the
system show that it can accurately count people traffic and estimate their moving
directions.
·
Bridging
the gap between motes and Aibo. Despite advances in the field of WSNs, the development of most applications currently
requires 
sensor-specific
programming techniques. The
research community, as well as industry, needs an efficient, convenient method
for accessing WSN data through existing infrastructure such as intranets and
the Internet. In this project, we have developed TRI, the TinyOS
Robot Integration server. TRI is a
multithreaded server that provides developers with WSN data management and
agent-agent communication channels through a TCP/IP connection and a human-readable
message protocol. The TRI server
hides the details of retrieving data from and managing a WSN. Thus, developers with standard TCP/IP
socket experience can incorporate WSNs into their
projects. We have also developed
TRI applications executing on a Sony Aibo that
responds to its environment by its onboard sensors and the extra sensory data
from a WSN.
·
Web-based access to real-time sensor data.
This project aims at developing use-friendly Web-based interface for accessing
real-time and 
historical sensor data. A
geo-centric Web interface based on Google Map is developed to visualizing
spatial and geographic data. A
small-scale wireless sensor network of Mica2 nodes is deployed in the Wireless
Sensor Networks Lab in the Engineering Build West on the MU campus to monitor the
light, temperature, and noise level. Sensor data is retrieved to a gateway
computer via TinyDB. A Java application then uploads the data to a remote
MySQL database.
User queries issued through the Web interface are transformed into SQL
queries to the MySQL database. The query results are transformed into
appropriate formats, such as tables or graphs, based on the user’s
preference and displayed inside the user’s browser.
Bioinformatics
On bioinformatics, we have been focusing on
novel models, scoring schemes, and techniques based on the mini-threading
approach for protein structure 
prediction. The
protein folding problem is considered a grand challenge and, in 2005, Science
named it one of the 125 biggest unsolved problems in science.
We are developing new methods to improve the statistical models and
computational methods for identifying useful fragments in PDB for a query
protein and to formulate spatial restraints derived from the alignments between
a query sequence and its fragment hits of known structures. We have
investigated new optimization techniques to build coarse-grain structural
models and developed effective methods to select the top protein structures
from a large pool of predicted structures.
·
Statistical models and computational methods for identifying useful fragments
in PDB. Built upon our results on efficiently identifying
remotely compatible fragments in PDB for a query protein sequence, we use
secondary structures of the query protein predicted from our in-house tool
MUPRED to guide the selection of the compatible fragments. Our results indicate
that predicted secondary structures improve the accuracy of the tertiary
structure prediction in almost all proteins that we tested. We study the
relationships between RMSD and the alignment parameters (e.g., E-value, length
of the alignment, gap size, etc.) for the compatible fragments, which served as
guidance for better selecting compatible fragments. We also search compatible
fragments based purely on (predicted) secondary structure matches for protein
structure construction.
·
Deriving spatial restraints from compatible fragments and building
coarse-grain structural models. We explore various functions
for representing spatial restraints (e.g., non-harmonic objective function and non-local
distance function) and methods for optimizing the structures using the
restraints. In our general framework, our goal is to generate diverse
structural models first and then use independent methods to assess the quality
of models for selecting the final structures as prediction. Our refinement
using different objective function, such as harmonic, non-harmonic, or non-local
distance function can help generate diverse population of structural models. Our
methods improve the prediction accuracy for most of the test cases.
·
Evaluating quality of coarse-grain structural models. We have developed several successful methods for evaluating the quality
of structural models. One method is to check how the spatial restraints are
satisfied. If the spatial restraints are not satisfied well, the quality of the
structural model is generally poor and we can filter it out. We also developed
a decision-tree approach for this purpose. We found the radius distribution of
C-alpha atom can indicate the quality of a structural model. Furthermore, we
performed
Information Analysis and
Recommendation on the Web
User information on
the Web is increasingly becoming an important component of business
intelligence. Product recommender systems are effective tools used by online
business to sale relevant products to interested customers. For example, in the online movie rental
business, Netflix’s success heavily depends on
their movie recommender systems. In 2006, Netflix announced a 1-million dollar contest
to the public aimed at producing a better recommender system. For the Netflix
Prize contest, Netflix has made available a large, anonymous subset of its
ratings database, totaling around 100 million ratings made by 500,000 customers
against 18,000 movies.
Problems similar to the
Netflix Prize problem have been tackled successfully by a class of methods called
collaborative filtering. Given
a known history of customer preference ratings on a set of “items”
in a database (items could be products for sale, documents in a website, etc.),
how can we predict the rating of a customer for an item that he or she has not
yet rated? Several approaches exist, including clustering, content-based
systems which use information about items drawn from a database to inform their
predictions, memory-based systems which correlate each item’s ratings
with all others (or ratings made by each user with those made by all others),
and model-based systems such as singular value decomposition (SVD). We have
developed several methods and software code to tackle the challenge problem and
our results have been ranked in the top 50 in the world.
For students at MU,
if you are interested and like to do research in any of these areas, feel free
to contact me via e-mail or come to my office in EBW 125.