Your Federal Quarterly Tax Payments are due April 15th Get Help Now >>

INTRODUCTION TO GENETIC ALGORITHMS by ewghwehws

VIEWS: 4 PAGES: 23

									   INTRODUCTION TO GENETIC
    ALGORITHMS AND NEURAL
          NETWORKS
GENETIC ALGORITHMS WERE
 INTRODUCED IN THE 1970’S
THEY AROSE FROM THE FACT THAT
 SCIENTISTS REALISED THAT
 BIOLOGICAL SYSTEMS CAN
 PERFORM TASKS MORE COMPLEX
 THAN COMPUTER POWER
 INTRODUCTION TO GENETIC
       ALGORITHMS
MOST SCIENTISTS AGREE THAT
 COMPLEX BIOLOGICAL SYSTEMS ARE A
 RESULT OF EVOLUTION
 INTRODUCTION TO GENETIC
       ALGORITHMS
EVOLUTION TAKES PLACE AT THE
 LEVEL OF THE CHROMOSOME
THE ORGANISM (BODY) SERVES AS
 A VESSEL IN WHICH THE GENES
 ARE CARRIED AND PASSED
 ALONG
 INTRODUCTION TO GENETIC
       ALGORITHMS
NATURE TENDS TO MAKE MORE
 COPIES OF CHROMOSOMES THAT
 PRODUCE A MORE “FIT”
 ORGANISM
DIVERSITY NEEDS TO BE
 MAINTAINED IN THE POPULATION
THESE CONCEPTS CAN BE APPLIED
 TO THE WORLD OF COMPUTING
 INTRODUCTION TO GENETIC
       ALGORITHMS
SIMPLE STRINGS OF NUMBERS ENCODED
  IN BINARY STRINGS CAN REPRESENT
  CHROMOSOMES
  INTRODUCTION TO GENETIC
        ALGORITHMS
THE 1ST COMMERCIAL APPLICATION WAS
 EVOLVER IN 1989, AN EXCEL ADD-IN.
 IT IS MORE POWERFUL THAN SOLVER
 IN EXCEL
EXAMPLES FOLLOW:
   EXAMPLE APPLICATIONS

BUDGET PROBLEMS WITH MULTIPLE
 CONSTRAINTS
COMPLEX SORTING PROBLEMS
SCHEDULING TASKS E.G CLASSES
 OR FLOW OF JOBS
MINIMISE TRANSPORTATION TIMES
 BETWEEN SITES
ROUTING PROBLEMS
   EXAMPLE APPLICATIONS

AN OPTIMAL MIX OF NPV PROJECTS
 MAY PROVE MORE PROFITABLE
 THAN A RANGE OF PROJECTS
 CONSIDERED INDIVIDUALLY AS
 THIS MAY ALLOW UNPROFITABLE
 PROJECTS TO BE SHELTERED
 FROM TAX
     NEURAL NETWORKS

BOTH GENETIC ALGORITHMS AND
 NEURAL NETWORKS ARE
 EXAMPLES OF ARTIFICIAL
 INTELLIGENCE TECHNOLOGIES

GA MIMICS DARWINIAN EVOLUTION
 AND NN MIMICS THE BRAIN
     NEURAL NETWORKS

EACH TECHNOLOGY IS SUITED TO
 SOLVING DIFFERENT TYPES OF
 PROBLEM
WITH GA’S WE USUALLY WANT TO
 FIND THE INPUTS TO PUMP
 THROUGH A GIVEN MODEL TO
 FIND THE OUTPUTS
     NEURAL NETWORKS

IN A NEURAL NETWORK PROBLEM
  WE OFTEN HAVE MANY SETS OF
  INPUTS AND THEIR RELATED
  OUTPUTS SO WE SEARCH FOR THE
  MODEL THAT TIES THEM
  TOGETHER (PATTERN
  RECOGNITION)
     NEURAL NETWORKS

STOCK MARKET MODELLING
USED TO DETERMINE VALIDITY OF
 EMH AND HENCE FORECASTING
 SHARE PRICES
         EXAMPLES

GENETIC ALGORITHM

GIVEN THE COMPLEX SALES
 FUNCTION FIND THE LEVEL OF
 SERVICE OFFERED AND
 ADVERTISEMENTS PLACED TO
 ACHIEVE THE HIGHEST PROFIT
      NEURAL NETWORK

GIVEN THE LEVEL OF SERVICE
 OFFERED AND ADVERTISEMENTS
 PLACED TO ACHIEVE THE
 HIGHEST PROFIT FIND THE
 COMPLEX SALES FUNCTION TO
 PREDICT SALES LEVEL
    GENETIC ALGORITHMS

THE GA IS PRIMARILY A SEARCH
  MECHANISM TRYING OUT N
  SOLUTIONS AND EVALUATING THE
  RESULTS
IT EVALUATES HOW GOOD OR BAD
  EACH GUESS IS
      NEURAL NETWORK

THE NEURAL NETWORK IS ALMOST
 THE OPPOSITE IN THAT IT BUILDS
 A BLACK BOX FROM INPUTS AND
 OUTPUTS

IT IS GOOD FOR PREDICTIONS AND
  PATTERN RECOGNITION
       NEURAL NETWORK

A NN RESEMBLES A SIMPLIFIED BRAIN
THE PROCESSING ELEMENTS OR
  NEURONS ARE JOINED BY A PROCESS
  OF INTERCONNECTED WEIGHTS
ONCE A WHOLE NETWORK OF NEURONS
  IS CONNECTED IT CAN RESPOND TO
  THE INPUTS PRESENTED TO IT
       NEURAL NETWORK

INFORMATION FLOWS FROM THE INPUTS
  THROUGH THE NETWORK TO THE
  OUTPUTS
EXPECTED AND ACTUAL OUTPUTS ARE
  COMPARED AND THE WEIGHTING
  ADJUSTED ACCORDINGLY PROVIDIND
  A “FLUID”COMPUTATIONAL SYSTEM
  CAPABLE OF “LEARNING “FROM
  EXPERIENCE
NEURAL NETWORKS
Neural Network Structure:




Source: http://www.bostonspinegroup.org/images/research/artificial-
intelligence.jpg
       NEURAL NETWORK

THE MODEL CREATED IS NOT
  MATHEMATICAL
THE MODEL LEARNS WHICH INPUTS
  INFLUENCE THE OUTPUT
A DANGER IS THAT IT CAN MEMORISE
  SOLUTIONS FROM THE “TRAINING”
  DATA AND HENCE PERFORM POORLY IN
  PRACTICE
AN EXAMPLE OF THE DANGERS

IN AN ATTEMPT TO IDENTIFY GENDER
  THE FOLLOWING CHARACTERISTICS
  WERE DETERMINED IN A PERSON:
VERY SHORT HAIR
100KG WEIGHT
DEEP VOICE
2M TALL
BROWN EYES
         INFLUENCES

WHICH OF THE ABOVE INPUTS ARE
 INFLUENTIAL?
WHICH ARE IRRELEVANT?
ARE THERE ANY COMPLEX
 RELATIONSHIPS?

BLACK BOX CREATED (NON-
 MATHEMATICAL MODEL)
         PROBLEMS

CAN YOU TELL WHETHER THE
 NETWORK HAS “LEARNT” FROM
 THE TRAINING DATA OR IS IT
 JUST MEMORISING IT AND
 THEREFORE MAY FAIL WHEN REAL
 DATA IS INPUT?

								
To top