My research interests are in Artificial Intelligence. In particular I am interested in the following topics:
1.
Anomaly detection and diagnosis:
Anomaly detection procedure
is necessary in many systems: physical systems, software, autonomous
systems, networks, cyber systems, etc. Once an anomaly has been detected a
further diagnosis task is essential to identify the root cause of the
anomaly. The diagnosis enables a decision support system to recommend either
repairing a certain component or replacing it or even tolerate the fault.
In my research I am interested mainly in Model-Based Diagnosis (MBD)
techniques, but also data driven techniques. MBD relies on a model of the
diagnosed system, which is utilized to simulate the behavior of the system
given the inputs. The outputs are compared to the actual behavior to detect
discrepancies indicating failures. The model can then be used to pinpoint
possible failing components within the system. To address diagnosis
problems, I established the Fault Detection and Diagnosis Lab at BGU, which
promotes research with the government and leading corporations such as
General Motors, Mekorot and IBM.. For example, diagnosis of faulty
communication components in networks, identifying faulty components in
distributed systems, or diagnose faults in UAVs.
In addition, I am interested in using
MBD for new theoretical problems, such as using SAT solvers for MBD, the
differences between intermittent and non-intermittent faults and the
influence of different parameters of the systems on MBD algorithms.
In the last years I started to work on cyber security
research. Cyber security tries to identify anomalous events, users communication
etc. This is related to my background of trying to identify faults and to reason
about their root cause. In this research I use especially machine learning
techniques. For example, anomaly detection using temporal pattern recognition
for SCADA systems, anomalous user detection by user profiling, and threat
sharing models and representations.
My new research using machine learning techniques on the one hand, and my deep background in model-based diagnosis on the other hand, brought me to the conclusion that the combination of these two techniques is a key factor for efficient and strong new tools for the diagnosis problem. Following this conclusion, in the last years, I started to work on this hybrid method. I used it in the automated debugging area and now I implement it in some new projects dealing with faults in water main systems, rolling systems, and cyber-physical systems.
2.
Voting procedures to
minimize preference elicitation:
Voting is an essential
decision-making mechanism in multi-agent systems (MAS) that allows multiple
agents to rank possible candidates and chooses a winner that reflects their
joint preferences.
Previous work has typically assumed that the voters
provide a complete set of preferences to the center. However, in practice,
especially in distributed settings, it may not be feasible for all agents to
submit a complete set of preferences. First, in real-world applications, it may
be impractical to expect individuals to provide all their preferences for a
large number of candidates, from the perspective of the human interface and due
to the need to interrupt the human as less as possible. For example, if a group
of friends wish to choose a movie to watch together, it would be impractical to
ask all the members to individually rate all current movies, i.e., the candidate
set.
A second motivating scenario for submitting an incomplete set of
preferences requirement may be seen in telecommunications. In this domain, a
basic requirement is to save on communications in order to reduce bandwidth
overhead. For example, consider a meeting schedule involving prospective
participants using their PDAs to vote on meeting times. A user would not specify
preferences for all possible slots since there are hundreds a month.
Furthermore, it would be costly to communicate all preferences to the center
conducting the vote.
In practice, it is possible to determine a winner by
specifically requesting agents for certain preferences rather than for their
whole set of preferences. A key question is what partial information is
essential for determining a winner? In my research I address this question