Introduction to Artificial Intelligence
Due: Thursday, 11/9
Learning Boolean Functions with Back-Propagation: Implement
back-propagation as described in Section 19.4 of our text. Also implement
the momentum variant described in class. Use the sigmoid function
as your activation function. No extra thresholding units will be
necessary. An implementation of a BooleanFunction class and an interface
for your neural network implementation is provided here.
Javadoc documentation can be found here.
Details of what you should implement are given in the comments of interface
MultilayerFeedForwardNeuralNetwork. Training of the network can,
of course, terminate early if the network has converged to correct weights.*
Varying arity: Create simple neural networks with n inputs,
2^n hidden units, and 1 output unit. Such networks are capable
of learning any boolean function. Your first experiments will seek
to gain understanding of how many training epochs are necessary to learn
random boolean functions. Do not use momentum for this experimentation,
and use a learning rate of 0.1. Do simple analysis of the epoch
data (mean, median) and plot a histogram with 4-10 intervals for three
different arity: n=3,4,5. Be sure to set an epoch limit of
at least 1000 in case. What do you observe?
Varying learning rate: Continuing with the previous network configuration,
let n=3. Vary the learning rate and repeat the analysis of
the previous question.
At least learning rate values (.25, .05, .1, .2, .4). What do
Varying momentum: Continuing with the previous network, use a fixed
learning rate of 0.1. Vary the momentum constant and repeat the analysis
of the previous question. Test at least four values between 0.0 (no
momentum) and 1.0. What do you observe?
Independent experimentation: Design an experiment of your
own to gain further understanding of some aspect of back-propagation.
This experiment can be a refinement of one of the previous three experiments,
with a range of values providing a clearer picture. You might instead
seek to vary two parameters to see if there's a dependency. You might
think of something completely different. This is an important opportunity
to think as a researcher: (1) pose an interesting question, (2) design
an experiment to gain insight, (3) execute the experiment, (4) analyze
the results, and (5) share your findings.
* NOTE: What it means for your neural network to converge to correct
weights: If, for all boolean true/false inputs converted to 1.0/0.0, the
network's output is >= 0.5 if and only if the correct output is true, then
the network has converged. In other words the boolean expression
(output >= 0.5) should be equivalent to the desired boolean output.
However, this does not mean that the network output should
be rounded to the nearest integer. 0.0 and 1.0 are still the target
values and the non-target values are still needed for error. A
network can converge and still have error.