Introduction to Artificial Intelligence Questions and Answers

Answer: A

Vision addresses recognition, detection, segmentation, reconstruction and tracking, turning arrays of pixel values into statements about objects and their relationships. Display, sensor design and compression are engineering tasks in adjacent fields. The difficulty is that the mapping from three dimensions to two loses information that must be recovered by assumption, which is why prior knowledge matters as much as image processing.

Answer: B

As a branch of the field, robotics is about closing the loop between perception, reasoning and physical action in an environment that is partially observable, stochastic and unforgiving. Mechanical design and manufacturing are engineering disciplines it draws on, and simulation is a tool. A physically embodied agent cannot avoid the messiness that a software agent can be shielded from, which is why robotics has been a demanding test of every technique the field has produced.

Answer: C

Real sensors return noisy and incomplete data, actuators do not execute commands exactly, and mistakes have consequences that cannot be undone by resetting, so techniques that work in simulation frequently fail on hardware. The gap between the two is a recognised difficulty in its own right. Languages are the same, experiments can be repeated with care, and the physical world has more states rather than fewer. A testbed that punishes unrealistic assumptions is valuable precisely because it does so.

Answer: D

A policy trained in simulation often fails on hardware because the simulator's physics, sensor noise and dynamics differ from reality, and closing that gap through domain randomisation or system identification is an active research problem. Latency, endurance and specification accuracy are separate engineering concerns. A model that is good enough for design but not for training is a distinction worth drawing carefully, since the two uses place very different demands on fidelity.

Answer: A

Driving is partially observable because other road users' intentions are hidden, stochastic because outcomes are uncertain, dynamic because the world does not wait, continuous in state and action, and emphatically multi-agent. It sits at the difficult extreme of every dimension of the classification, which is why it has proved so much harder than early demonstrations suggested. Placing a task on each dimension before starting work is the most useful thing the environment taxonomy provides.

Answer: B

Every system in use today is narrow: excellent at chess, or at translation, or at detecting tumours, and entirely incapable of anything else. General artificial intelligence, competent across tasks as a person is, remains a research aspiration. Computing power, data volume and rule count are unrelated to the distinction. Being clear about the difference between a narrow success and a general one is the single most useful corrective to overstated claims about the field.

Answer: C

Generality means transferring competence to genuinely novel tasks without retraining, which no current system approaches. Speed, data and scale have all been proposed as the missing ingredient and none has yet proved sufficient. Whether generality requires a new idea or merely more of what exists is the central open question of the field, and honest disagreement about it is reasonable in a way that overconfident prediction in either direction is not.

Answer: D

A capable optimiser will pursue exactly the objective it was given, which may differ subtly and consequentially from what was meant, and specifying objectives correctly proves surprisingly hard. Reward hacking and specification gaming are the observed symptoms. Mechanical alignment, distribution shift and multi-agent agreement are different problems, the second being important in its own right. A system that does what it was told rather than what was wanted is a hazard that grows with capability.

Answer: A

A simulated robot told to move fast may fall over and roll, and an agent rewarded for scoring may loop endlessly collecting points instead of finishing the race, because the specification permitted it. The behaviour is a fault in the objective rather than in the optimiser. Self modification and outright failure are different phenomena. That an optimiser will find the letter of a specification rather than its spirit is a practical lesson with immediate consequences for how objectives are written.

Answer: B

Where a system refuses credit, recommends a treatment or affects a person's liberty, an unexplainable decision cannot be challenged or audited, and regulation increasingly demands that it can be. Interpretable models are often less accurate rather than more, publication is not generally required, and cost is unaffected. A property demanded for social rather than technical reasons is still an engineering requirement once it is written into a specification.