AI Programming Lecture 9 Genetic Algorithms

1
AI ProgrammingLecture 9Genetic Algorithms

• Richard Price Simon Worgan
• School Of Computer Science University Of
  Birmingham
• msc59rmp, msc74sxw_at_cs.bham.ac.uk

2
Last Week

• NLP
• Compositionality.
• Blocksworld.
• Semantics.
• Context.
• Intentionality.

3
Introduction

• Nature inspired.
• Travelling Salesman Problem.
• Genotype.
• Phenotype.
• Selection.
• Operators.
• Replacement.
• Considerations.

4
Nature Inspired

• Darwins Theory
• Reproduce offspring.
• Offspring arent identical.
• Selection Pressure.
• Chromosome
• 0 1 0 1 0 1 1. (an individual).
• Gene
• A single bit.

5
Example Problem

• Travelling Salesman Problem (TSP).
• Given
• A number of cities spread across a landscape.
• Find the shortest complete tour of all cities.

3
5
6
4
2
7
8
1
9
6
Algorithm

• Select representation (genotype).
• Select fitness function (phenotype).
• Initialise population.
• Select parents.
• Generate offspring.
• Form next generation.
• Until stopping criteria go to 1.

7
Genotype

• Genotype defines representation.
• E.g
• Bits String. 0 1 1 1 0.
• Real values. 0.5 1.3 26 4.
• TSP representation.
• Order of cities in tour.

2
3
2
3
1 2 3 4
1 3 4 2
1
4
1
4
8
Phenotype

• Phenotype is observed behaviour or fitness.
• Genotype performance in environment.
• E.g
• Eye colour, height, etc.
• Function values.
• TSP
• Total distance travelled in tour.
• Minimisation problem.

2
4
2
4
1
3
1
3
9
Selection Methods

• Select individuals according to their fitness.
• Tournament Selection.
• Given a subset of individuals.
• Order these individuals dependent on fitness.
• Select individuals dependent on position in
  tournament.
• Either deterministically or stochastically.

10
Selection Methods

• Roulette Wheel Selection
• Individuals are given a proportion of the wheel.
• Size of proportion is dependent upon fitness.
• Larger proportions are more likely to be
  selected.

11
Generating Successors

• Improve from one generation to the next.
• Generate offspring from the selected
• How?
• Needs to be legal.
• Needs to be useful.

12
Mutation

• Legal
• Visit all cities only once.

• Insertion mutation
• Swap mutation

13
Mutation - Algorithm

• Insert mutation
• 1) Copy one of the survivors
• 2) Select a random city from the chromosome (item
  from list)
• 3) Insert in new location
• Swap mutation
• 1) Copy one of the survivors
• 2) Remove a sub-list
• 3) Reverse sub-list
• - rev(sub-list)
• 4) Re-insert list

14
Crossover

• Legal tours

15
Crossover - Algorithm

• 1) Select two of the winners
• 2) Take a subset of cities
• 3) Insert these into a new list, maintaining the
  ordering and location from the first parent
• 4) Take the missing cities from the second parent
• 5) Insert into the gaps in the list, maintaining
  the ordering.
• 6) Make the first parent the second and
  visa-versa
• 7) Repeat once

16
Useful Pop-11 features

• Mutation maplist(list1, p) -gt list2
• maplist(winners, mutate) -gt list2
• Selection syssort(population, fitter) -gt
  ranking
• fitter(E1, E2)
• Selection Mutation random0(population or
  range)
• Iteration fast_for .... end fast_for
• Warning, no type checking.
• Have a look at 'Help Fastprocs' and 'Help
  Efficency' for more advice.

17
Replacement

• Options
• Strong replacement
• Weak replacement
• Spatial replacement
• Generational replacement

18
Problems / Considerations

• Phenotype redundancy
• Plateaus in the search space
• Hamming Cliffs
• Erratic fitness landscape
• Unreachable solutions
• Illegal barriers

19
Extensions

• Species.
• Male / Female.
• Co-evolution
• Predators and Prey
• Baldwin evolution / Local optimisation
• REMEMBER
• Biological analogy Vs. Computational reality

20
Conclusions

• Representation defines the fitness landscape.
• Production of offspring defines the search.
• There are a huge number of different approaches.