Introduction to Artificial Intelligence Questions and Answers

Answer: C

Reports such as Lighthill's in the United Kingdom concluded that the field had failed to deliver on its promises, particularly on the combinatorial explosion in real problems, and funding was withdrawn. The other options describe events that either did not occur or had no such effect. External assessment of a field against the claims it made for itself is a hazard of promising more than can be delivered, and the pattern repeated a decade later with expert systems.

Answer: D

An expert system captured the if then rules of a human specialist in a narrow field such as medical diagnosis or mineral prospecting, together with an inference engine to apply them. Their commercial success in the nineteen eighties and their subsequent difficulties with maintenance and brittleness shaped the field's later direction. Automatic learning, generality and neural simulation describe other approaches. A technique that succeeds narrowly and fails to generalise is a recurring shape in this subject.

Answer: A

Once a system contains thousands of rules, adding or changing one can have unforeseen consequences elsewhere, and no expert has a global view of the interactions. Knowledge acquisition also proved to be a bottleneck, since experts often cannot articulate what they know. Hardware, cooperation and deployment were secondary difficulties. A system whose behaviour emerges from many local rules is hard to reason about globally, which is a difficulty shared with modern learned systems for different reasons.

Answer: B

Human experts often cannot state the rules they follow, and eliciting them through interviews is slow and error prone, which limited how quickly expert systems could be built. Machine learning was adopted partly as a way around this bottleneck. Reading speed, staffing and storage cost were not the fundamental difficulty. A limitation that lies in the human rather than the machine is one that better computers cannot solve, which is why the response was a change of method.

Answer: C

In a learned system the mapping from input to output is fitted from examples rather than specified by a programmer, which sidesteps knowledge acquisition and copes with problems nobody can articulate rules for. The price is a dependence on data and a loss of transparency. Speed and numerical restriction are not the distinction, and evaluation becomes more rather than less important. Where the behaviour comes from is the deepest division within the field.

Answer: D

The three are nested: artificial intelligence encompasses search, planning, knowledge representation and much else, machine learning is the subset concerned with learning from data, and deep learning is the subset of that using many layered neural networks. Treating them as synonyms obscures large parts of the subject that involve no learning at all. Getting the nesting right prevents the common error of supposing that every artificial intelligence problem is a learning problem.

Answer: A

Logic came from the philosophical tradition running through Aristotle to Boole and Frege, and was formalised by mathematicians, giving the field its representation and inference machinery. Neuroscience contributed the neural metaphor, economics decision theory and utility, and linguistics grammar and parsing. Each contributing discipline brought a different piece, and recognising which came from where explains why the field's vocabulary is so mixed.

Answer: B

Utility theory, expected value and game theory gave the field its formal account of what acting rationally under uncertainty means, which is the foundation of the rational agent view. Grammar came from linguistics, the threshold neuron from neuroscience and search from computer science and operations research. Tracing an idea to the discipline that developed it usually explains its vocabulary and its assumptions, and here it explains why agents are described as maximising something.

Answer: C

Rationality is defined relative to a performance measure, the agent's knowledge and the actions available, so a rational agent does the best it can with what it has rather than achieving perfection. Omniscience is not required and is not achievable. Human resemblance is a different criterion, and errors caused by incomplete information do not make an agent irrational. Defining rationality relative to available information is what makes it a usable engineering criterion.

Answer: D

An agent that crosses a road after looking, and is struck by a falling object, acted rationally though the outcome was bad, because rationality concerns the decision given the information available rather than the outcome. Omniscience would require knowing the actual future, which is unattainable. Confusing the two leads to judging agents by outcomes rather than by decisions, which is unfair and also useless as a design principle.