4 Machine evolution

1
4 Machine evolution

• Evolutionary computation
• Genetic programming (GP)
• program representation
• GP process
• Wall-following robot
• Discussion

2
4.1 Evolutionary computation

• Evolution
• descendants perform better than ancestors
• something similar to produce programs?
• changes to population, selective survival
• Application areas
• function optimization (Holland -75) with genetic
  algorithms (GA)
• classification problems
• genetic programming

3
4.2 Genetic programming

• Individuals (functional) programs
• expressed as trees presenting the function
• target evolve a wall-following robot
• repeated execution of a function
• primitive functions and/or/not/if
• action functions n/e/s/w
• same notation n, ne, e, ... for sensory inputs
• Note total functions only
• reason all constructions compute something

4
4.2.2 GP Process

• Generation 0
• population of random programs
• parameter size of the population
• Evaluation of individuals
• run program, see how well it manages
• here take 60 steps, count of cells next to
  wall visited (0..32)
• summed over 10 runs from random points
• fitness

5
Constructing the next generation

• Tournament selection (TS)
• select k individuals from population
• choose the one that evaluates best
• Survival
• choose p of current generation with TS
• Crossover
• create (100 - p - m) child programs
• select parents with TS
• crossover subtree replacement

6
Next generation (cont.)

• Mutation
• used (if at all) sparingly
• select m of individuals (TS)
• replace a randomly chosen subtree with a new
  random subtree
• Parameters of GP
• population size, k, p, m
• evaluation function

7
4.3 Wall-following robot

• What does Gen 0 look like?
• (AND (sw)(ne)) does nothing (0)
• (OR (e) (west)) moves if e0 (5)
• uninformed search
• Evolution
• note difficult to read, redundant operations
• 4.5 Gen 0 (92), 4.6 Gen 2 (117)
• 4.7 Gen 6 (163), 4.8 Gen 10 ok!

8
4.3 Discussion

• Successful evolution of S-R agents
• e.g. inverted pendulum
• synthetization of electronic circuits
• Enhancements
• evolution of subroutines
• memory, recursion