CIS 732
Sunday, 08 October 2001
Due:
Thursday, 25 October 2001
(before
midnight Friday 26 October 2001)
This
problem set is designed to apply your theoretical understanding of supervised
learning using connectionist – i.e., artificial neural network (ANN) and
Bayesian belief network (BBN) – models.
Refer
to the course intro handout for guidelines on working with other students.
Note: Remember to submit your solutions in electronic form using hwsubmit and produce them only from your
personal notes (not common work or sources other than the textbook or properly
cited references).
Problems
1.
(30 points) Maximum A Posteriori / Maximum Likelihood
(MAP/ML). Problem 6.3, Mitchell.
2.
(20 points) Bayes Optimal Classification. Explain how a stochastic iterative
improvement algorithm such as simulated
annealing can be used to approximate the Bayes
optimal classifier. What theoretical and
practical limitations (specifically regarding convergence) do such Markov chain
Monte Carlo (MCMC) techniques face?
3.
(25 points) Bayesian Belief
Networks (BBNs).
Problem 6.6, Mitchell. Fill in
the CPT for the node Humidity instead
of Wind.
4.
(25 points) Neural Computation
(NC). Problem 19.3, Russell and Norvig
Extra credit
a)
(5 points) Try the MATLAB
Neural Network toolkit on Sleep1
and report the same results for a feedforward ANN
(specifically, a multi-layer perceptron) trained with
backprop. This
package can be found on the KDD Core systems, including a Windows version
installed on the Hobbits.
b)
(5 points) TBA (SNNS).
c)
(5 points) TBA (Hugin).