Genetic Programming

1
Genetic Programming

• Jelle Kastelein

2
Overview

• Genetic algorithms
• Introduction to genetic programming
• Example Automatically programming a robot
• Critiques
• Overview of real life applications

3
Genetic Algorithms (GAs) 1

• John Holland (1975)
• Loosely inspired by biological evolution
• Typically performs operations on bit strings
  (crossover, reproduction, mutation)
• Hypothesis search space 2n, expensive

4
Genetic Algorithms (GAs) 2

• Example
• Crossover 111000 x 101010 by mask 111100 gives
  111010 and 101000
• Mutation 111000 -gt 101000

5
Genetic Programming (GP)

• John Koza (1992)
• Evolve programs, rather than bit strings
• Motivation
• Tell computer what you want, rather than what to
  do. (Arthur Samuel)
• Easier than writing complex programs(?)

6
3 Step algorithm

• 1 Generate random initial population
• 2 Do until termination criterium has been
  reached
• A Execute each member and assign fitness
• B Create new population according to fitness
  values by
• I Reproduction (Copy)
• II Crossover (Recombine)
• 3 Select best (approximate) solution as final
  result.

7
A practical issue

• Problem Many syntactically useless programs
• Useful LISP
• S-expressions
• (OR (NOT D1) (AND D0 D1))
• (OR (OR D1 (NOT D0)) (AND (NOT D0) (NOT D1)))

8
S-Expressions

• (OR (NOT D1) (AND D0 D1)) (OR
  (OR D1 (NOT D0)) (AND (NOT D0) (NOT D1)))

9
Crossover

• Crossover Fragments

• Resulting Trees
• OR(AND (NOT D0) (NOT D1)) (AND D0 D1))
• OR(OR (D1 (NOT D0)) (NOT D1))

10
5 decisions

• Set of terminals (constants, variables)
• Set of primitive functions
• Fitness function (domain dependent)
• Parameters (population size, generations)
• Termination criterium

11
Example problemBox moving robot

• Robot pushes box towards wall in irregularly
  shaped room
• Turn left (TL), right (TR) (30o), move forward
  (MF) (1ft / timestep)
• No arguments, constant
• 12 sonar sensors, BUMP, STUCK detector
• Nothing is labeled or interpreted a priori
• 2 versions (SS, MF0, TL0, TR0)

12
Ingredients

• Set of terminals (T)
• Input from sensors primitive motor functions
• T S00, S01, ..., S11, SS, (MF), (TL), (TR)
• Set of primitive functions
• F IFSTK, IFBMP, IFLTE, PROGN2

13
Ingredients (continued)

• Fitness
• Distance between wall and closest box corner
  after a maximum of 350 timesteps summed over four
  fitness cases
• Fitness of 0 is best

14
Ingredients (continued) (continued)

• Parameters
• Population size 500 (should reflect complexity)
• Max 51 generations
• Termination criteria
• Program with a fitness of 0
• 51 generations have been run

15
Results (Version 1)(45 gen., 307 p., f0.0)
16
Results (Version 2)(20 gen., 207 p., f0.0)
17
Comparison to Reinforcement Learning

• GAs have been viewed as a form of (exotic) RL
  (Fitness Reward)
• Differences with traditional RL
• RL requires a large nr. of training steps, but...
• GP requires a great deal of (parallel)
  computational force.
• Koza Easier to implement (5 Vs. 13 decision
  steps)

18
RL solution issues

• Reduce to 3 behaviors (find, push, recover from
  stalls)
• Condense sensor info into NEAR/FAR (218 262144
  states)
• Required designing filters to compensate for
  information loss
• Problem specific
• Consolidate states within hamming distance (29
  512)
• Etc...
• The point requires expertise within the domain.

19
GP Critique

• Conventional programs are extremely sensitive to
  random changes.
• Use less fragile representation
• Fitness function hard to define (how bad is a
  wrong answer?)
• Concentrate on application domains where fitness
  is easy to define
• Computationaly intensive

20
So, why use GP?

• Human competitive results can be obtained
• Routineness and generality (domain inspecific)
• Controller -gt antenna design -gt sorting network
• Simply adapt functions and fitness
• Exploiting parallel hardware

21
Applications

• Planning (artificial ants)
• Control
• Boolean function learning
• Classification and pattern recognition
• Sorting Networks
• Designing circuitry
• Discovering inverse kinematic equations

22
Real life applications

• 36 human competitive results
• 21 duplicate patented inventions
• E.g. Tunable, integrated active filter
• Some solve problems in previously unseen ways
• E.g. High-current load circuit
• Some outperform previous patents
• Balun circuit (4-fold improvement)
• Voltage-current conversion circuit (62 of
  original error)
• Transmembrane domain protein identifier (lower
  error than human written classifiers)

23
Questions?