• Exercise 19.1.

  For each leaf in the decision tree of Figure 19.1 (starting with “use” or “do”), give an example of an application that has the characteristics that would end at that leaf. For example, for the first leaf, give an application where the stakes are low, there is abundant homogeneous data, etc.

  Suggest an application that should not follow the advice of Figure 19.1. Explain why.

  Exercise 19.2.

  Consider one of the following scenarios.

  • (i)

    You are working for a company and have been asked about the feasibility of building a tool for predicting the results of an upcoming election so they can plan appropriately. You will have access to data about the outcome of previous elections, demographic data from the census about the voters, information about the parties and candidates. The goal is to predict the probability of which party or parties will form government. A rival company has proposed solving it by combining hidden Markov models and gradient-boosted trees.

  • (ii)

    A large national coffee and donut shop is looking to modernize their donut production operations. Currently they have a single production machine that bakes and decorates the many kinds of donuts and pastries they make and dumps them onto a single line. The donuts on this line are then sorted by human operators into boxes for display and delivery. They need to fill two types of boxes: boxes where all the donuts are the same type, and boxes containing 12 unique types of donuts (there are more than 12 types of donuts). The company is proposing to replace these human sorters with robots. Your job is to advise them. A rival company has proposed using deep learning with deterministic planning.

  • (iii)

    A biomedical firm approaches your company to develop a drug interaction advisor for doctors to understand the interaction of different drugs on their patients. They have access to a number of large databases of leading research on the effects of different drugs. However, the databases overlap on some drugs and diseases but not others. The databases use different notation for representing drugs, diseases, and their impact. Each database makes probabilistic predictions about drugs curing specific diseases, causing negative side-effects, or having no impact. A rival company has proposed using a mix of ontologies and probabilistic relational models to solve this problem.

  • (a)

    Explain how the problem fits into the abstraction of an agent.

  • (b)

    Explain how the rival company’s solution may work, and explain why they may have chosen the technologies they proposed.

  • (c)

    What is the most challenging part of solving this problem? What would you recommend as a way to solve this? Justify any recommendation made.

Find some current AI applications and classify the state-of-the-art for that application in terms of the dimensions. Does the application automate what Kahneman [2011] calls System 1 or System 2 or neither or both?

Give an argument for and against each of these propositions, about the possibility of a singularity, when computers will be more intelligent than people.

• (a)

  The singularity will never occur as people have common sense that can never be matched by a computer.

• (b)

  The singularity has already occurred; I would trust an answer that Google provides more than I would trust an answer from a random person.

• (c)

  The singularity is meaningless as humans are so tightly integrated with computers that the interaction will always be more intelligent than either one.

• (d)

  The singularity will happen in a few decades and will make humans subservient to computers.