Research Interests and Current Work
|
I am currently a post-doctoral research associate under the joint supervision
of Jude W. Shavlik and
C. David Page.
At present, I am also working with
Sriraam Natarajan, Richard Maclin
and my former advisor, Kristin P. Bennett.
Bootstrap Learning.
(read more)
BL (see
here) is a new machine learning paradigm that focuses on
learning progressively more complicated concepts through a `ladder` of lessons;
lower (earlier) rungs of the ladder teach simpler concepts needed to learn the
concepts at the higher rungs. The emphasis in BL is on efficient communication between
a helpful teacher and the student-agent; the main
motivation of BL is it may be easier for a teacher to instruct rather than to program
an agent to perform a certain task, much as a human coach would do, via natural language
instructions.
In the BL paradigm, the student-agent learns
from a variety of modalities of teacher input, including from pedagogical examples,
from being told, from noticing, and from experimentation and feedback.
Our focus is currently on discriminative learning from examples using Inductive Logic
Programming (ILP) and Support Vector Machines. (SVMs)
Online Learning With Prior Knowledge.
(read more)
We are interested in incorporating prior knowledge
into online support-vector-based learning algorithms. In the online setting, the learning algorithm
receives data points sequentially rather that all at once (as in batch learning) and predicts the
label at each round. After the prediction, the algorithm receives feedback indicating the correct label,
using which the decision function may be updated. The goal is to successively update the decision
function taking into account prior knowledge in the form of soft polyhedral advice (knowledge-based
support-vector machines, see here)
so as to make increasingly accurate predictions on subsequent rounds. The advice
helps speed up and bias learning so that better generalization can be obtained with less data.
Reinforcement Learning.
(read more)
RL methods have difficulty
scaling to large, complex problems. One approach
that has proven effective for scaling RL is
to make use of advice provided by a human. The goal of this research is
to extend a recent advice-giving technique, called
Knowledge-Based Kernel Regression (KBKR), to
RL in both batch and online settings. These approaches can then
be evaluated approach on the KeepAway
subtask of the RoboCup soccer simulator. The potential of advice-giving
techniques such as KBKR for RL, and the design decisions involved in
employing support-vector regression in RL are prime motivations for this study.
Bilevel Optimization for Model Selection.
(read more)
A key step in many statistical
learning methods used in
machine learning involves solving a convex optimization problem
containing one or more hyper-parameters that must be selected by the
users. While cross validation is a commonly employed and widely
accepted method for selecting these parameters, its implementation
by a grid-search procedure in the parameter space effectively limits
the desirable number of hyper-parameters in a model, due to the
combinatorial explosion of grid points in high dimensions.
Bilevel optimization approach is used to formulate
a unifying framework within which issues such as model
selection can be addressed. The machine learning problem is formulated as a bilevel
program--a mathematical program that has constraints which are
functions of optimal solutions of another mathematical program
called the inner-level program. The non-convexity of bilevel programs is a
problem that must be addressed through optimization techniques. The latter
must also be fast and be capable of handling large data sets in order to be
effective for real-world problems.
|
