quiz3_2009

The length of the subsequence k1 kt that is correctly

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: assified C. The length of the longest subsequence ki … kj that is correctly classified D. The number of ki correctly classified E. The number of knock subsequences ki … kj that are correctly classified 2. Select all of these fitness functions that one cannot evaluate using the neural net as a black box: 3. Select all of the fitness functions that will immediately trap the genetic algorithm in a fitness plateau: 4. Select all of the fitness functions that do not correlate with the actual fitness: 5. Having selected a fitness function, you decide to mutate weights randomly, and choose about half of the weights from each parent for crossover. Your friend uses the GA to train an NN on the example knock sequence, and it consistently says true for that knock sequence. Excited, he installs it, goes outside, waits for it to lock, someone runs by, and the door opens. What was missing from his training data? 6. Having added that, he retrains the system on all the training data, and it's classifying things perfectly, and he goes outside and waits for it to lock, knocks the secret pattern, trying again and again, but you 10 eventually have to let him back in. What happened? Tear-off sheet A B T wA 0 0 -1 0 1 -1 1 0 -1 1 1 -1 wB 0 wT 0 ∑ 1 y y* 1 1 0 1 Perceptron update: if |y*-y| > 0: for each wi: w'i = wi + r(y*-y)xi 11 -1 x y w11 w12 w21 w1T w31 -1 w w32 w22 w2T -1 For your entertainment after the quiz: http://www.youtube.com/watch?v=zE5PGeh2K9k 12 3T...
View Full Document

This document was uploaded on 03/17/2014 for the course EECS 6.034 at MIT.

Ask a homework question - tutors are online