Announcement

What: Resume Writing Workshop
When: 4:10 p.m. - 5:00 p.m. on Friday, January 20th
Where: 218 Roberts
Who: Ruth Kronfuss, Director of Human Resources at Zoot Enterprises
Sponsor: Local ACM Chapter

Chapter 2: GAs in Problem Solving

Section 2.1: Evolving Computer Programs

Evolving Lisp Programs

Genetic Programming or GP

• John Koza
• Parse trees are used to represent Lisp code. For example, (sqrt (+ (* x x) (* y y)))
• Steps
  1. Choose a set of functions and terminals
  2. Generate a syntactically correct initial population of parse trees
  3. Evaluate the fitness of each program
  4. Copy the most fit 10% into the next generation, produce the other 90% by using fitness based selection with single point crossover (no mutation due to large initial population)
  5. Go to step 3

Blocks World

• Koza
• Sensors: CS, TB, NN
• Functions: MS(x), MT(x), DU(exp-1, exp-2), NOT(exp), EQ(exp-1, exp-2)
• 166 fitness cases
• Population size of 300

Evolving Cellular Automata

• Crutchfield, Das, Hraber, Mitchell
• To study how natural evolution creates sytems that enable emergent computation
• To study how to engineer emergent computation in decentralized multiprocessor systems

• 1-D CAs typically "wrap around"
• A rule takes as input the current cell and the other cells in its radius, r
• Space time diagram, Figure 2.6

Task

• Decide within M time steps whether an initial configuration of N bits contains a majority of 1s with r = 3.
• In this case the critical density, p_c = 0.5
• Space time diagrams for the majority rule, Figure 2.7
• A chromosome has a bit length of 2^2r+1 = 2⁷ = 128. Thus, the search space size is 2¹²⁸
• N = 149
• Population size = 100
• Steps
  1. Generate a random initial population
  2. Select 100 new initial configurations over a uniform distribution of density
  3. Calculate the fitness for each rule
  4. Copy the 20 most fit rules into the next generation
  5. Form the other 80 rules by randomly selecting 2 of the 20 selected rules and then performing single point crossover. Mutate each offspring twice.
  6. Go to step 2
• P_N(rule): the percent of 10000 initial configurations generated from an unbiased distribution correctly classified within 2N timesteps
• Measured values of P_N, Table 2.1
• More space time diagrams, Figure 2.8