1 NEURAL NETWORKS 1. What is the difference between product units and summation units? (2) 2. Name and sketch three activation functions and describe when it would be appropriate to use each of them. (6) 3. A Feedforward Neural Network is one of several types of neural networks. Name and briefly discuss two other types of neural networks and mention the type of problem that can be solved by each. (8) 4. Explain what is meant by the terms overfitting and underfitting. Which steps can be taken to ensure that they do not occur? (5) 5. Discuss four approaches that can be followed to improve the training time and generalization ability of a neural network. (12) 6. Explain how a SOM can be used to find the missing value of an attribute in an input vector after training of the SOM. (2) 7. Describe the main differences between supervised and unsupervised learning rules. (4) [39] 1 2 EVOLUTIONARY COMPUTING 1. What is meant by elitism? What are the advantages and disadvantages of using elitism? (3) 2. Why may it be advantageous to vary the mutation rate of a Genetic Algorithm? (2) 3. Discuss and criticize the following strategy for updating the mutation rate of a Genetic Algorithm: Start evolution with a large mutation rate, and decrease the mutation rate as the number of generations increase. (4) 4. What is the motivation for using Gray coding to represent individuals? (2) 5. Name three possible mutation operators that one could apply when evolving code. Assume that the chromosome representation is a tree structure. (3) 6. What is the main difference between Genetic Algorithms and Evolutionary Programming? (1) 7. Assume a population size of N ≥ 50. Compare the following two approaches to tournament selection: • Tournament size is 0.1N • Tournament size is 0.9N (4) 8. Can Differential Evolution be used to train a Neural Network? Motivate your answer. (5) 9. Give five fundamental differences between Genetic Algorithms and Differential Evolution. (3) [27] 2 3 SWARM INTELLIGENCE 1. What are the main differences between swarm intelligence and evolutionary computing? (3) 2. Consider the following function: f (x1 , x2 ) = axb1 + cxd2 + e Assume that you are given a list of values for x1 , x2 and f (x1 , x2 ) (i.e. a list of inputs and expected outputs of the function). Explain how you would use a particle swarm optimization algorithm (PSO) to approximate the actual values for a, b, c, d and e. (5) 3. Discuss three Social Network structures typically used in PSO. (6) 4. Discuss and criticize the following strategy for updating the dynamic inertia weight: Start with an inertia weight of 3.0 and linearly decrease it to 0.5 as a function of the iteration number. (4) 5. The Vmax constant in PSO is used to prevent too large jumps in the search space by controlling the amount of velocity added to the position of particles. What potential problem exists in using Vmax ? Propose a method to solve this problem that will still ensure that velocity updates do not become too large. (5) 6. Explain how PSO can be used to find niches (i.e. different solutions to the same problem). (5) 7. Without giving any formulas, give a very brief explanation of how a simulated ant colony can be used to cluster data. How can the problem of overfitting (having more clusters than what are necessary) be overcome? (6) [34] ————————————- 3