Evolutionary Computing Genetic Algorithms

1
Evolutionary ComputingGenetic Algorithms

• TMR4115
• Thursday, Oct 30 2003
• Prof. Stein Ove Erikstad

2
Classes of Evolutionary Algorithms and AI

• Genetic Algorithms
• analogous to evolution
• Simulated annealing
• analoguous to the way metal cools into a minimum
  energy crystalline structure
• Neural networks

3
Genetic Algorithms

• Search based on mechanisms from natural selection
  and genetics
• chromosomes
• genes
• alleles
• locus
• genotype
• phenotype
• generation
• population
• survival of the fittest
• The driving force in GA is the combination and
  exchange of chromosome material during breeding

4
Difference from classical optimization

• Population of points
• not a single one
• Use fitness (payoff)
• not derivatives
• Probabilistic transition rules
• not deterministic

5
Design Representation in Genetic Algorithms

• A design description is represented as a string
• This string may be
• a direct binary representation of a design
  variable
• a binary representation of a discrete point in an
  interval
• a combination of different types of varibles,
  e.g. numerical, boolean, lists, etc.

100100100111101010111001011000111
6
Three core processes

• Reproduction
• Crossover
• Mutation

7
Reproduction

• Individuals are selected for mating
• Probability of selection increases with fitness
• survival of the fittest

8
Crossover

• Simple crossover by splitting strings and
  exchanging parts
• Crossover point selected randomly

1 0 1 1
0 0 0 1
1 0 1 1
1 0 1 1 0 0 1 0
0 0 1 0
0 1 1 1 0 0 0 1
0 0 1 0
0 1 1 1
0 1 1 1
0 0 0 1
9
Mutation

• Arbitrary change at single point
• Typically low frequency in both natural and
  artificial systems
• But useful insurance towards premature loss of
  important notions

Mutation Flipping bit position
0 1 1 1 0 1 0 1
0 1 1 1 0 0 0 1
10
A Simple GA Algorithm
We want to determine the value of the design
variable x for which the performance f(x) is
maximized

• Create initial population
• Determine range
• Create random individs
• Create associated bitstring representation
• Select for reproduction
• Determine fitness for each
• Assign propability for reproduction based on
  fitness
• Use spinning wheel to select individs for mating

11
A Simple GA Algorithm 2

• Use crossover to create children
• For each pair, randomly select crossover point
• Exchange bitstrings to create two new children
• Use mutation to insert new genetic material
• For each individ, use mutation probability to
  determine whether to mutate or not
• If mutation, randomly select position to flip bit

12
Example

• Excel

13
Some strategies for improved performance

• Normalized fitness function
• often necessary for survival of the fittest
  mechanism to work
• increases relative fitness difference
• Elitism
• preserve best individs from generation to
  generation
• avoids loosing good solutions
• may lead to premature convergence
• Parent competition
• children must also comete against parents
• quicker convergence poorer search coverage
• Tuning algorithm settings
• population size
• mutation probability

14
Pro Cons

• Computationally intensive
• for complex fitness evaluation
• Gives little insight into design model behaviour
• blackbox approach
• Difficult to know about quality of solution

• Easy to adapt to existing systems
• wrap around existing analysis tools
• only link is fitness calculation (performance)
• No requirements to model behaviour
• Not trapped in local optima