Unformatted text preview: can leave your solution expressed as a
product, and it may help us assign partial credit. 9 Part C (10 points)
Your friend is building a device to unlock a door when it hears a secret knock pattern, and realizes that
a neural network could do it, if only it had a sense of timing. You suggest feeding the output of one
neuron into the input of one of its ancestors, and thereby get a dependence on timing.
1. You think about training the network by standard back-propagation, but decide that you can't. Why? The solution is clear: Genetic Algorithms! You'll set up a population of identical neural networks with
random weights, you discretize your input every 100 milliseconds into a sequence k1...kn of 0 if silence
and 1 if a knock was heard, ensuring that k1 is always 1, and timing out eventually. You'll choose the
fittest few neural networks at each step. Your friend jots down a few ideas for fitness functions:
A. Whether the full knock pattern was correctly classified
B. The length of the subsequence k1 … kt that is correctly cl...
View Full Document
- Fall '10
- Artificial Intelligence, Artificial neural network, neural network, Shaun, decision function