COSC Quiz1 6367 Evolutionary Programming Tu., Feb. 20, 2007
Your Name: Your SSN:
Problem 1 --- Selection and Survival of the Fittest [11]: Problem 2 --- ES and Rechenberg's Rule [8] Problem 3 --- Crossover and Mutation [9] Problem 4 --- Genetic Programming [8] Problem 5 --- Genetics [8] Problem 6 --- Questions [9]:
[53] :
Grade:
The quiz is "open books and notes" and you have 75 minutes to complete the quiz; it will count approx. 11% towards the overall course grade.
1
1) Selection / Survival of the fittest [11]
a) Assume 3-tournament selection is used in an EC system with a population size of 100. What is the probability that the worst population member is selected for reproduction (assume, no fitness ties)? What is the probability that the best individual is selected for reproduction? Giving the formula is fine![3] With replacement: Worst: Best: (1/100) ^ 3 1 (1 1/100)^3 = 1 (99/100)^3
Without replacement: Worst: 0 Best: 1 (99/100)*(98/99)*(97/98) b) Compare k-tournament selection with roulette wheel selection. What are the major differences between the two selection methods? [4] According to Eiben p. 62, roulette wheel can be used with either a fitness-based or ranking selection distribution (e.g. SUS). A difference between k-tournament and roulette wheel is that roulette wheel requires determination of global population statistics. K-tournament performs a ranking amongst the selected k individuals. Also, ktournament is explicitly rank-based, while roulette wheel is potentially fitnessproportional. Selection pressure can be regulated in k-tournament based upon choice of k, while with roulette wheel manipulation of the distribution is necessary for the same purpose. c) Assume you use a fitness function f(x) (x is the solution that is evaluated by the fitness function) in conjunction with roulette wheel selection, and when running your EC system you observe that the population converges quite prematurely ("too much selective pressure"). What could be done to alleviate the problem? [2] Scaling of the fitness function to reduce the spread between most and least-fit individuals alleviates the problem. Examples: sqrt(f(x)), log(f(x)), etc. d) EC are based on Darwinian evolution. How is "survival of the fittest" simulated in the ES approach? [2] Because all individuals are eligible parents in ES, survival of the fittest is simulated by survivor selection: the killing off of less-fit offspring.
2
2) ES and Rechenberg's Rule [8]
a) Explain how Rechenberg's 1/5 rule works! How does a system behave following the rule if it is near the top of a hill? How does it behave if it is on a plateau (an area were all fitness values are the same)? How does it behave if it is on the foot of a big hill? [5] Eiben3.ppt Slide 6 describes Rechenberg's rule
p = ratio of mutations in which child is fitter than the parent Near the top of the hill On a plateau Foot of big hill p < 1/5 p < 1/5 p > 1/5 decrease decrease increase
b) ES systems change the variance of their employed mutation operator during the evolution process. Why are they doing that? [3] The changing variance of the mutation operator facilitates the self-adaptation ES of systems. The variance adapts to fit the topology of the changing fitness landscape, to direct the search in directions which are more advantageous.
3) Crossover and Mutation [9]
a) What role does a crossover operator play in a GA system? What role does a mutation operator play in a GA system? [4] Eiben3.ppt Slide 19
(summarized) Crossover - exploration Mutation - exploitation b) Assume you have an ES system that solves a 2-dimensional optimization problem, whose initial solutions consist of (0,2), (2,2), (2,0), and (1,1) population size 4 and your system employs whole arithmetical recombination (see page 51 of our textbook; also called arithmetical crossover in the lecture) as its only modification operator. What solutions will be searched, which solutions will not be searched by the described system? [5] The 4 points can be considered as boundary points of a convex hull. Arithmetical crossover explores those points which lie inside this boundary. Solutions not searched are points which are external to the area.
3
4) Genetic Programming [8]
a) What are the major applications for genetic programming? [4] Machine learning (classification, model building, prediction) Symbolic regression Evolving programs b) In genetic programming the size of solutions can change. Is this, in general, desirable? Give a reason for your answer! [2] This is desirable. Because there is no prescribed or known size for good solutions of a genetic programming problem, it is advantageous for the EC system to search through many different solution sizes. c) GP faces the problem of bloating (survival of the fattest). Propose one approach to alleviate the problem of bloating. [2] Bloating can be alleviated by these methods: 1) Introduce a maximal size, prohibit variation operators from exceeding this size for generated offspring. 2) Parsimony pressure: introduce a penalty term in the fitness formula that penalizes being oversized
4
5) Genetics [8]
a) How is reproduction in human beings different from traditional GA systems that, let us say, employ one point crossover on a single-stranded bitstrings? Be specific![5] Human genetic code exists in diploid not haploid form. In humans the combination of genetic material from the 2 parents occurs without crossover (haploids align). Crossover in humans occurs in meiosis. This process differs from one point crossover on single-stranded bitstrings, in the replication of the diploid strands and in that 2 out of 4 haploids exchange genetic material. b) Define phenotype and genotype! Which operators in a GA system operate on the genotype? [3] Phenotype candidate solutions in the original problem context, external characteristics Genotype the genetic representation of a phenotype, in an evolutionary computing system Operators in a GA system on the genotype: crossover, mutation
5
6) Questions [9]
a) EC systems search probabilistically for good solutions. What advantages you see in using probabilistic algorithms over determinic algorithms for solving complex search problems? [5] b) Under which circumstances is it desirable to use a steady state EC approach, in contrast to generating a new population from scratch all the time? [4]
6