01.11.2014 Views

MACHINE LEARNING TECHNIQUES - LASA

MACHINE LEARNING TECHNIQUES - LASA

MACHINE LEARNING TECHNIQUES - LASA

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

148<br />

6.8.1.1 Algorithm for Kohonen's Self Organizing Map<br />

Assume output nodes are connected in an array (usually 1 or 2 dimensional). Assume that the<br />

network is fully connected - all nodes in input layer are connected to all nodes in output layer.<br />

Use the competitive learning algorithm as follows:<br />

1. Initialize the weights – the initial weights may be random values or values<br />

assigned to cover the space of inputs.<br />

2. Present an input x r to the net<br />

3. For each output node j , calculate the distance between the input x r and its<br />

weight vector<br />

j<br />

w r , e.g. using the Euclidean distance<br />

n<br />

r r j r r j j<br />

, = − = ∑ i<br />

−<br />

i<br />

i=<br />

1<br />

( )<br />

d x w x w x w<br />

(6.57)<br />

4. Determine the "winning" output node i , for which d is minimal.<br />

r r r r<br />

i<br />

k<br />

w −x ≤ w −x ∀k<br />

(6.58)<br />

Note: the above equation is equivalent to w i x >= w k x only if the weights are<br />

normalized.<br />

5. Update the weights of the winner node and of nodes in a neighborhood using<br />

the rule:<br />

Δ r = ⋅ ⋅ r −<br />

r<br />

( )<br />

i<br />

( ) η ( ) ( ) ( )<br />

k<br />

w t t hki<br />

,<br />

t x w t<br />

where η ( t) ∈ [0,1] is a learning rate or gain and h ( t)<br />

ki ,<br />

r<br />

ki ,<br />

1<br />

( t)<br />

(6.59)<br />

= is the<br />

neighborhood function. It is equal to 1 when i = k and falls off with the distance<br />

r between output nodes i and k. Thus, the closer are the units to the winner,<br />

ki<br />

the larger the update. Note that both the learning rate and the neighborhood vary<br />

with time. It is here that the topological information is supplied. Nearby units<br />

receive similar updates and thus end up responding to nearby input patterns. The<br />

i<br />

above rule drags the weight vector w and the weights of nearby units towards<br />

the input x.<br />

Note that if the output nodes are 1, the Kohonen rule is equivalent to the Outstar<br />

rule, see Section 6.6.2.<br />

Δ w = α⋅x ⋅y −γ<br />

⋅ w<br />

(6.60)<br />

ij i j ij<br />

© A.G.Billard 2004 – Last Update March 2011

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!