Yale University
Department of Statistics
Seminar

Monday, September 30, 2002

Two Estimation Problems in Machine Learning

Peter Bartlett
BIOwulf Technologies and Australian National University
(visiting UC Berkeley)
 
In this talk, we consider estimation problems from
two areas of machine learning: pattern classification
and reinforcement learning.  Many successful pattern
classification algorithms (including adaboost, support
vector machines, neural network classifiers, and logistic
regression) can be viewed as large margin classifiers:
they use thresholded real-valued functions, and choose
these functions by minimizing a cost function involving the
margin - the amount by which the function's value lies to
the correct side of the threshold.  The first part of the
talk will present recent results on statistical properties
(such as convergence of the risk, or misclassification
probability) of large margin classifiers.
 
In reinforcement learning problems, an agent chooses
actions to take in some environment, aiming to maximize a
reward function.  Many control, scheduling, optimization
and game-playing tasks can be formulated in this way.
Policy gradient methods consider agents that are restricted
to some set of policies, and aim to move through policy
space so as to improve the performance of the agent.
The central problem for such methods is that of estimating
performance gradients. The second part of the talk will
describe algorithms for this problem, and show how their
performance depends on properties of the controlled system,
such as mixing times.


Seminar to be held in Room 107, 24 Hillhouse Avenue at 4:15 pm