MODULE-IV
STATISTICAL METHODS IN ANN
Module 4
Statistical Methods: Boltzmann's Training - Cauchy
training - Artificial specific heat methods - applications
to general non-linear optimization problems
Statistical Methods are used for
Training ANN
Producing output from trained network
Training Methods
Deterministic Methods
Statistical Training Methods
Deterministic Training Method
Follows a step by step procedure.
Weights are changed based on
their current
values of weight.
It
also based on the desired output and the
actual output.
E.g.:-Perceptron Training Algorithm.
Back Propagation Algorithm etc
Statistical Training Methods
Make pseudo random change in the weights
Retains
only those change which results in
improvements.
GENERAL PROCEDURE
( FOR STTISTICAL TRAINING METHOD)
Apply a set of input and compute the resulting
output
Compare the result with target, find the error.
The objective of the training is to minimize the error.
Select a weight in random and adjust it by a small
random amount.
If
the adjustment improves our objective retain
the change
Otherwise
return the weight to the previous
value
Repeat
the procedures until the network is
trained to the desired level
The local minima problem
Objective Function
The objective function minimization problem can get
trapped in poor solution.
A
B
Weight
If the objective function is at A and if the random
weight
changes
are
small
then
the
weight
adjustment will be rejected.
The superior weight setting at point B will never
found and the
system will be trapped in local
minima instead of global minima at point B.
If the random weight changes are large both point
A and B are visited frequently, but so will every
other point.
The weight will change so drastically that it will
never settle at desired point.
Solution & Explanation
Statistical method overcome local minima problem by
a weight adjustment strategy.
Example:
Let
the fig. represents a ball on a surface in a
box.
If the box is shaken violently ,then the ball will
move rapidly from one side to the other side.
The probability to occupy any point on the
surface is equal for all points.
If the violence
of shaking is gradually reduced the ball
will stick to both point A and B.
If the shaking is again reduced it will settle to point B.
The ANN are trained in the same way as through
random weight adjustment.
At first large random adjustment are made.
The weight change that improves the objective
function is retained.
The average step size is hen gradually reduced to
reach global minimum.
Annealing [ Boltzmann Law ]
Annealing:-If a metal is raised to a temperature
above melting point ,the atoms are in violent random
motion. The atoms always tend to reach a minimum
energy state. As the metal is gradually cooled the
atoms enters a minimum possible energy state
corresponds to each temperature.
P (e) exp ( e / kT )
P(e)=probability that the system is in a state with
energy e.,k Boltzmanns constant. T temperature.
Simulated Annealing [Boltzmann Traing]
Define a variable T that represents an artificial
temperature. (Start with T at large value).
Apply a set of input to the network, and calculate
the outputs and objective function.
Make a random change weight and recalculate the
network output.
Calculate new objective function.
If the objective function is reduced, retain the
weight change.
If the weight
change results in an increase in
weight change ,calculate the probability of accepting
the weight change.
P(c) exp
( c / kT )
P(c)=probability of a change of c in the objective
function,k Boltzmanns constant. T temperature.
Select
a random number r from a uniform
distribution between zero and one.
If p(c) is greater than r, retain the change otherwise
return the weight to previous value.
This allows the system to take a step in a
direction that worsen the objective
function,hence
escapes
from
local
minimum.
Repeat the weight change process over each of the
weights in the network, gradually reducing the
temperature T untill an acceptably low value for
objective function is obtained.
How to select weights/artificial Temperature for training
The size of the random weight change is selected by
various methods.
2
2
Eg:- P ( ) exp( w / T )
P(w)=Probability of a weight change of size w.
T=artificial temperature
To achieve global minimum at the earliest the
cooling rate is usually expressed as follows
T0
T (t )
(log(1 t ))
The main disadvantage of Boltzmanns training is very
low cooling rate and hence long computations.
Boltzmanns machine usually takes impractical time
for training.
Cauchy Training
Cauchy training method is more rapid than
Boltzmann training.
Cauchy training substitutes cauchys distribution for
Boltzmann's distribution.
Caushys distribution has longer tails", hence more
probability for larger step size.
The
temperature reduction rate is changed to
inverse linear. (For Boltzmann training it was inverse
logarithmic.)
Cauchy s distribution is
T (t )
P( x)
[T (t ) 2 x 2 ]
The inverse linear relationship
reduction reduces the training time.
T0
T (t )
(1 t )
for
temperature
