1 Part I Artificial Neural Networks Sofia Nikitaki.

1

Part I Artificial Neural Networks

Sofia Nikitaki

2

What is a Neural Network

An information processing paradigm inspired by the biological nervous systems

Key information processing system Large number of highly

interconnected elements Learn by example

5

How Neural Networks works Basic features:

Construction of network

Computation functions of network

Training of network

6

Inside the Neural Network Pi : inputs Σ : sum for each

neuron f : function of the

transfer output

f(ai) = Σ ( Wi,i x Pi)+bi bi : threshold of each node.

7

Inside the Neural Network (cnt’d)

8

Training of the network Supervised

learning Output target exist Backpropagation

technique

Unsupervised learning Output target

unknown The network adjust

similar output for similar inputs

9

Backpropagation technique

1. Computes the total weight input Xi

2. Calculates the activity yi using the transfer function

3. Computes the error E

10

Backpropagation technique (cnt’d) How fast the E changes1. Activity of an output unit is changed

Error derivative (EA) Yi actual activity di the desired activity

2. As the total input received by an output unit is changed

Quality (EI) is the answer from step 1 multiplied by the rate at which the output of a unit changes as its total input is changed.

3. As a weight on the connection into an output unit is changed

4. As the activity of a unit in the previous layer is changed

11

Transfer functions

12

Gradient descent method

Algorithm step searchs locally to find the optimal value the optimal direction of the weight change

Training function – TrainSCG updates weight, bias values

Adaption Learning Function – LearnGDM updates weight, bias learning function

13

Performance Function - MSE

Network's performance the mean of squared errors

Measures the average of the square of the error

14

Part IINeural Networks and Location Sensing for indoor environments

15

CLS – Neural Networks

Input Deciles of Signal Strength values Signal Strength measurements

16

Neural Network Input : Deciles of Signal Strength values

2 layers

Input: 80 SS values per cell

Total: 8 APs

85 neurons

Tan-sigmoid transfer function

POSITION Output:2 valuesx,y coordinates

Linear transfer function

17

Results - Location Error

Median 2,6meter

18

Neural Network Input : Signal Strength values

Input: 480 SS values per cell

Total: 8 APs

2 layers 100 neurons

Tan-sigmoid transfer function

POSITION Output:2 valuesx,y coordinates

Linear transfer function

19

Results - Location Error

Median 1.8 meter

20

CLS – Comparison of all methods

21