This question paper consists of 4 printed pages, each of which is identified by the Code Number COMP232001 UNIVERSITY OF LEEDS School of Computing June 2004 AI23: Bio-Inspired Computing Time allowed: 2 hours Answer THREE questions. Question 1 Habituation (or desensitisation) can be described as reduced responsiveness (or disengagement) in the presence of continued stimulation. A minimal model of a habituation gate is a "pulsifier", which generates a brief pulse in response to an arbitrarily long but continuous string of non-zero inputs. a) Give a biological example of habituation. [2 marks] b) Label the neurons, and fully specify weights, thresholds and initial conditions, such that the circuit below functions as a pulsifier. Assume the inputs and outputs are all binary. [6 marks] c) Test your circuit and show that: (i) the output neuron returns zero in response to a zero input; (ii) the output neuron returns a brief positive pulse in response to a continued stimulus, and (iii) the network resets itself when the stimulus is stopped. [4 marks] d) List two different algorithms for training dynamical neural networks. Which would you choose to train a habituation circuit? Justify your answer. [4 marks] e) Could a feed-forward network be used to design the same habituation gate operation? Explain your answer. [4 marks] [Total: 20 marks] TURN OVER 1 Question 2 The Travelling Salesman Problem (TSP) can be solved by 1) 2) 3) 4) a) Hopfield networks (or attractor neural networks) Kohonen networks (or Self Organising Maps) Genetic algorithms Ant algorithms Describe how each method can solve the TSP problem. [8 marks] b) Write out a pseudo-code implementation of a 5-city TSP for one of the above algorithms. Include the initial set up of parameters, training and running stages of the code, as needed. [12 marks] [Total: 20 marks] TURN OVER 2 Question 3 Worker ants follow a simple clustering algorithm for cleaning up their nest. Objects to be cleaned emit an attracting signal which cause ants to pick them up. The more isolated an object, the more likely it is to be picked up. Clusters of objects emit a stronger signal which cause the ants to deposit objects onto them. The larger the cluster, the more attractive it is for further depositions. a) The key principle behind this algorithm is that of stigmergy. Explain what is meant by this term and how stigmergy facilitates clustering in the example above. [3 marks] b) Describe what happens as ants begin to cluster randomly placed objects on a 2-D grid. (i) How might the first cluster begin to form? (ii) What will the object distribution on the grid look like after some time? (iii) In the long run, where on the grid are the clusters most likely to be found? [4 marks] c) Clustering algorithms can straightforwardly generalise to sorting algorithms. How would you generalise the clustering algorithm described above (i.e., what modifications/additions would you make) to obtain a sorting algorithm that would partition multiple classes of objects into distinct clusters. [3 marks] d) Describe a specific real-world application of your sorting algorithm. How does it work? [3 marks] e) Suppose the ants in your algorithm were trained using a fitness function. Specify a suitable fitness function. Justify your answer. [3 marks] f) List and explain the advantages of ant algorithms in real-world applications. [4 marks] [Total: 20 marks] TURN OVER 3 Question 4 a) What role or effect does selection play in the following scenarios? i) ii) iii) iv) Artificial evolution of mice in the mice-demo of BEAST Artificial evolution in the absence of mutation (or with a negligibly slow mutation rate) Biological evolution on a neutral landscape Co-evolution in a predator-prey situation [8 marks] b) List and describe specific biological/natural analogies to three (3) of the above. [6 marks] c) List and describe three (3) different rules for selection for evolutionary and co-evolutionary algorithms. [6 marks] [Total: 20 marks] END 4