Introduction to Artificial Intelligence Questions and Answers

Answer: C

Left to itself the performance element would always take the action it currently believes best, and would never discover anything better, so the problem generator deliberately proposes suboptimal but informative actions. It is the architectural expression of the exploration and exploitation trade. Evaluation, goal setting and data collection are separate functions. Building exploration into the design, rather than hoping it happens, is what allows a learning system to improve beyond its starting point.

Answer: D

An agent that always exploits its current knowledge never improves, while one that always explores never uses what it has learned, and the balance between them determines how quickly it converges on good behaviour. It appears throughout reinforcement learning, bandit problems and adaptive control. Memory, methodology and evaluation cost are separate trades. A dilemma that arises purely from acting under uncertainty is one of the most general ideas in the subject.

Answer: A

The symbolic tradition represents facts and rules explicitly and derives conclusions by manipulating them, which makes the reasoning inspectable and the knowledge editable. Its weakness is brittleness and the difficulty of capturing knowledge. Connectionist, evolutionary and probabilistic approaches are the main alternatives. The division between explicit and learned representation is the deepest methodological split in the field and it recurs in every subfield.

Answer: B

Connectionism represents knowledge in the weights of a network of simple units, learned from data rather than stated, giving robustness to noise at the cost of interpretability. It is the tradition that produced modern deep learning. The other options describe system integration, symbolic inference and distributed computing. Recognising that the choice between explicit and distributed representation carries consequences for robustness, data requirements and explainability alike is what makes the distinction worth drawing.

Answer: C

Because the knowledge is explicit, a symbolic system can produce a derivation showing why it reached its conclusion, which matters in medicine, law and any regulated domain. It needs little or no data but a great deal of expert effort, and it is notably brittle at the edges of its knowledge. Training speed is not a meaningful comparison. Explainability is the property that has kept symbolic methods in use despite the successes of learning.

Answer: D

When an action is described, the vast number of facts that remain true must somehow be carried forward without stating each one, and doing so efficiently proved far harder than expected. It became emblematic of the difficulty of commonsense reasoning. Vision, frame representations and search are unrelated despite the shared word. A difficulty that appears trivial to humans and proves intractable to formalise is characteristic of commonsense knowledge generally.

Answer: A

Everyday knowledge is enormous in extent and almost never articulated, because people acquire it through experience rather than instruction, so there is no corpus to draw on. Projects to enumerate it by hand have made slow progress. It is largely shared rather than personal, changes slowly, and storage is not the limitation. A body of knowledge whose very obviousness prevents it being recorded is a genuine obstacle rather than merely a large task.

Answer: B

In search and reasoning the number of states or proofs grows exponentially with depth, so a method that works on a toy problem becomes hopeless on a realistic one. It was the difficulty that ended the early optimism and it motivates heuristics, pruning and approximation throughout the field. Data volume and rule proliferation are related practical problems, and methodological diversity is not a technical one. Recognising exponential growth as the enemy explains most of the techniques the subject has developed.

Answer: C

A heuristic estimates how promising a state is, so search can be directed rather than exhaustive, and it trades a guarantee for tractability. Admissible heuristics preserve optimality while still pruning enormously. Proofs, resource bounds and specifications are separate concepts. Accepting a method that usually works in exchange for one that is computationally impossible is the central bargain of practical artificial intelligence.

Answer: D

The subfield covers parsing, meaning representation, translation, summarisation, dialogue and generation, and it has been transformed by statistical and then neural methods. Programming language design, format conversion and compression are unrelated. Language is a particularly demanding testbed because it requires world knowledge as well as grammar, which is why progress in it has so often been taken as a proxy for progress in the field as a whole.