Computational Intelligence exam answers PDF

Preview

Summary

Question 1a. Evolutionary algorithms involve an Initialisation, Selection, Recombination or Crossover, Mutation, and finally Replacement. Initialisation involves the process of defining an initial population of n genotypes, where n is a user-specified variable. Usually, this process is stochastic to...

Description

Question 1 a. Evolutionary algorithms involve an Initialisation, Selection, Recombination or Crossover, Mutation, and finally Replacement. Initialisation involves the process of defining an initial population of n genotypes, where n is a user-specified variable. Usually, this process is stochastic to grant a good initial level of diversity; however, the population can be seeded to involve good solutions from the start. Selection regards choosing two or more individuals to be parents to new solutions. Selection is an exploitation operator, as it employs known areas of the search space to formulate better solutions. Although there exist many selection methods, the selection operator is what brings selection pressure into an EA and subsequently usually employs some form of elitism: the best parents can be chosen amongst the population based on their fitness value. Recombination or Crossover is the process of generating offspring: children of the chosen parents involve inherit some part of each parent’s chromosome and constitute a new, different solution, which could have a higher or lower fitness value than their parents, but will maintain the same chromosome length. Crossover is an exploitation operator, as it employs a pair of known genotypes to construct other possible solutions. Mutation is designed to bring a small amount of variation in the population and can be applied to the offspring after Crossover: a gene or a set of genes in the child chromosome get changed either stochastically or, in case of floating point or integer representations, by a value within user-defined bounds. Each child has a very small probability of being mutated. Mutation is an exploration operator, as it introduces new genetic material in the existing population. The Replacement operator determines which members of the population the offspring would replace, if any, to maintain a constant population size throughout generations. Usually, the best individuals will replace the worst existing solutions in the population. The evolutionary process is described by the following diagram: the algorithm is run for a userdefined number of iterations (generations).

1

b. Selection pressure is the pressure amongst the population to be picked as a parent and carry on its genetic material to the next generation. An overly high selection pressure entails individuals with lower fitness values will rarely get picked to be parents, and subsequently their genetic material will never impact the progression of the evolution. This could cause a loss of useful diversity in the population and could lead to premature convergence of solutions around a local optimum. On the other hand, an excessively low pressure could lead to slow evolution times, as individuals with high fitness values will not necessarily be selected as parents or carried on to the next population. A way to adjust selection pressure is to choose an appropriate algorithm for the selection operator so to grant a balanced level of elitism and useful diversity. Similarly, picking an adequate replacement operator can grant the opportunity for non-optimal solutions to be included in the population. Another factor consists in the amount of seeding introduced at the initialisation stage of the algorithm: if a number of solutions are introduced at the start, selection pressure will increase if not regulated by selection and replacement operators.

c. Parent 1 Parent 2 Child

30 100 65

82 75 79

90 42 66

8 28 18

82 82 82

34 58 46

2

83 25 54

53 94 74

39 27 33

17 33 25

Question 4 a. A Crisp set is a set of clearly defined items: an item is either a member of a Crisp set, or it is not. On the other hand, items can be members of multiple Fuzzy sets, as their limits are not clearly defined. Fuzzy sets have membership functions to determine the degree to which an element is a member of such set, which is usually symbolised by a value between 0-1. An example of a Fuzzy set could be height specified as Short, Average or Tall, while an example of a Crisp set would be Fruit, as each category specifically defines an item and there is no intermediate degree of membership (e. g. Apple, Pear, Banana). b. It would be appropriate to define Age in a fuzzy set, as the boundaries between age groups are not clearly defined. An example of a fuzzy Age set would be Young, Adult, and Old. Similarly, a fuzzy Temperature set could be defined by Cold, Ok, Hot as there is no crisp definition of which temperatures are members of each category. Job Title could not be described as a fuzzy set, as professions are clearly defined categories, and it could not be a member of multiple sets. Movie Rating can again be defined as a fuzzy set in the form of Bad, Ok, and Good, as a movie can have a degree of membership in each category. c. i.

ii.

Slow: 0.6 Medium: 0.4 Fast: 0 Soft: 0 Medium: 0.6 Hard: 0.6

d. i.

If Slow Then Slow = 0.6 If Medium and Not Soft Then Fast = min(0.4, 0.4) = 0.4 If Fast Then Slow = 0 If Slow and Hard Then Fast = min(0.6, 0.6) = 0.6

ii.

Fast = 0.6 and Slow = 0.6

e. Fuzzy logic is useful for game control as it can give advantages in movement and in circumventing obstacles, as fuzzy rules provide a more gradual and smoother transition between states and there would be no need to modulate crisp steering forces. Fuzzy logic can also help decision making, as the AI can pick which action to perform based on the estimated fuzzy confidence level of an action. Ultimately, fuzzy logic simulates the human way of reasoning, and does not model the perfect level of confidence of a computer: imperfect decision making will make the game more interesting and credible. The main drawback of using fuzzy logic for decision making is the amount of time and resources required to calculate confidence levels for each rule and take action: as to decide which action to perform an AI would have to calculate confidence levels for every rule, in a table with a large number of rules, the process can result resource consuming and add further complication to the implementation of a system.

3

Question 2 a.

4

b. Information gain refers to how useful an attribute is in the decision-making process: when splitting a branch, maximising information gain is crucial to gain conclusions from the dataset and assess whether the branch should be split further. Particularly, the information gain algorithm calculates the entropy of each attribute before splitting and after splitting the branch; if the entropy is lower after the split and information gain is subsequently higher, said attribute will be chosen to split the branch. c. Overfitting happens when the data used to train a tree does not provide a sufficient number of examples for each possible classification, and the tree is subsequently overly biased towards certain solutions. The tree will not be able to generalise and make conclusions based on unseen data, as it will have learnt the training data as a strict model, and not as a set of rules. d. Firstly, we would need to pre-process the data: the tree will not be able to use some of the attributes as classifiers effectively unless they are formalised in some kind of range or ratio. This is the case with Living Area, which could be summarised in an interval of categories Small, Medium and Large. Some attributes, on the other hand, could be completely excluded from classification: for example, the Unique identifier of each property and the Salesperson Name are unlikely to bring useful conclusions. The second step would be to identify the possible actions decision making could lead to. In this case, the tree could be either used to find correlations amongst the properties – i.e. a house with a conditions of 4 will have a higher price – but it could also be used by an estate agency to find a set of suitable properties for a buyer given a desired attribute. For example, a buyer could specify their budget and the desired number of bedrooms: the tree could decide which property to market based on these two attributes. In both cases, the tree can use a splitting criterion such as information gain to choose which attributes to split the branch in, but the attributes would have different information gains for each scenario. Once the tree has been trained on a large number of examples with properties with diverse attributes in each category, the effectiveness of its decision making can be tested with a training data set.

5...