Name _____________ CS472/572 Evolutionary Computation First Exam This is a closed note, closed book exam. The abbreviation GA is used for Genetic Algorithm throughout. 1. (15 points) Consider a real valued optimization problem with 50 variables, all in the range -512 to +512. Below are 3 mutation approaches. What are the weaknesses of each approach? a) Change all of the 50 values by either +1 or -1. b) Change one value (selected randomly) by 0.1, if this doesn’t improve the fitness, change the same value by -0.1. If that doesn’t improve fitness, repeat the process, picking a different value to change each time the process is repeated. c) With 50% probability change each of the values (i.e. on average half of the values are changed), the new value for each changed variable is drawn randomly from the full range of -512 to +512. 2. (33 points) Match each of the definitions to the appropriate term or terms from the following list. Note that there are some extra terms. _______ Type of evolutionary algorithm in which the whole population is replaced in every iteration. _______ Type of search algorithm in which the probability of keeping a new, worse solution gets smaller as the algorithm proceeds. _______ Crossover as shown below: 01011100 11100101 01100100 11011101 _______ Crossover as shown below: 01011100 11100101 01000101 11111100 _______ A selection operator based on the relative fitness ranking of the individuals. _______ A selection operator based on the relative fitness values of the individuals. _______ A solution that has no neighbors with a better fitness. _______ Optimization technique with a “population of size 1” that only moves to better solutions. _______ Form of GA mutation in which a small change is made to many of the values in the individual. _______ Form of GA mutation used with permutation representations. Terms: a. uniform crossover b. swap mutation c. insert mutation d. creep mutation e. evolutionary strategies f. GA g. steady-state h. 2-point crossover i. Order crossover j. tournament selection k. roulette wheel selection l. m. n. o. p. q. r. s. t. meta-GA generational 1-point crossover local optima global optima fitness landscape fixed bit simulated annealing hill climbing 3. (10 points) In the 0/1 knapsack problem you have N items, each item has a weight (wi) and a value (vi). The goal is to find the subset of the N items that have the maximum value and total weight less than some maximum (Wmax). For example, if you have the following 9 items: Item 1 2 3 4 5 6 7 8 Num: Weight 5 7 3 2 8 11 3 1 Value 5 6 2 3 11 10 5 1 If you select items 2, 5, and 7 the total weight is 7+8+3 = 18 and the total value is 6+11+5 = 22. Now, consider a 0/1 knapsack problem with N = 1,000, i.e. 1000 items. You would like to solve it using a generational GA with elitism of two (two copies of the best individual are made each generation) and a population size of 1,000 running for 1,000 generations. What is an upper bound on the fraction of all possible solutions will the GA test? Be as exact as possible and explain your answer. 4. (5 points) Draw the landscape of a problem that would be very difficult for hill climbing, but reasonable for a GA. 5. (20 points) A clique is a set of completely interconnected points in a graph (see below). The maximum clique problem is to find the largest clique in a graph. Imagine designing a GA to solve the maximum clique problem for a graph with 100 nodes. 4 7 Sample graph, the maximum 1 clique is shown with dashed 8 lines (vertices: 1,4,5,7,8). 5 Another, non-maximum, clique is shown with dotted lines (vertices: 5, 8, 9). 2 9 6 3 How would you represent potential solutions? I.e. what would an individual/potential solution look like? Why would you use that representation? How would you measure fitness? Why? How would you perform mutation? Why? How would you perform crossover? Why? 6. (5 points) What are the two most important differences between a GA and simulated annealing? 7. (4 points) Describe one practical problem with roulette wheel selection. 8. (8 points) The following is a list of changes that you could make to an evolutionary algorithm. Circle all of the changes that are likely to make the population converge faster (i.e. that are likely to make all of the individuals in the population become more similar faster). Underline all of the changes that are likely to make the population converge slower (i.e. that are likely to make all of the individuals in the population remain different for longer). Some changes may not do either, leave those alone. You may explain your answers. a) Increasing the tournament size when using tournament selection. b) Increasing the probability of mutating each value in an individual. c) Increasing the number of elite individuals. d) Switching from generational to steady-state