Final Review

Question 1

casewrite()

Answer 1

writes a frozen video frame to a file.

Question 2

caseread() –

Answer 2

reads a frozen video frame from a file. The file also contains the “supervisor’s” choice of a correct move.

Question 3

readweights()

Answer 3

reads in the weights data from a file. This is the “brain.” There are two weights arrays formed. One is 3 dimensional and the second is 2 dimensional.

Question 4

Writeweights()

Answer 4

records the current weights files (the “brain”) to a file.

Question 5

applynet()

Answer 5

uses the input node values and the weights arrays to calculate the firing values at the hidden nodes and then at the output nodes.

Question 6

backpropagate()

Answer 6

where the learning happens. Working backward from output values to hidden values to input values and the current weights, the network choice for an output node and the correct output node value from the supervisor, this function calculates and applies changes to the weights to reduce the overall network error.

Question 7

steer()

Answer 7

send the correct commands to the robot motors to steer the robot and go forward

Question 8

normalize()

Answer 8

adjusts input values to produce a consistent contrast to reduce the differences caused by different light levels

Question 9

sigmoid() brief

Answer 9

our “threshold” function to make all hidden and output node firing values be between 0 and 1.

Question 10

Capture functions

Answer 10

normalize() and casewrite()

Question 11

capture

Answer 11

cam on, robot off; record case w/ steering move

Question 12

Initialize functions

Answer 12

writeweights()

Question 13

Initialize

Answer 13

builds starting wts file; creates brain that knows nothing

Question 14

Train functions

Answer 14

ApplyNet(), Backpropogate(), readwts, writewts, caseread, sigmoit, normalize

Question 15

train

Answer 15

read in current wts and training, use applynet to find firing vals at hidden and output nodes, use supervisor’s input and do backprop() to correct all wts

Question 16

RobotRun functions

Answer 16

readwts(), normalize(), applynet(), sigmoid(), steer()

Question 17

RobotRun

Answer 17

runs robot; neural network driving

Question 18

how to apply steering move

Answer 18

after applynet finds all the hidden and then output node values, we choose that output node with the highest value. (-2, -1, 0, 1, 2)

Question 19

tic tac toe recursion

Answer 19

minimax is double recursion with min and max calling each other;

Question 20

supervised learning

Answer 20

learns by comp calc w/supervisors; hardest part is getting training cases

Question 21

sigmoid funct and x

Answer 21

1/(1+e^-x);

Question 22

sigmoid role

Answer 22

use instead of square wave for differentiability; is used to connect every layer to the next layer. For each, the system needs an array of real number weights.

The final layer is called the output layer. Each node represents an action for the agent. The agent chooses that action with the largest firing value.