Introduction to Artificial Intelligence Questions and Answers

Answer: C

Different patient populations, equipment and recording practices mean the deployed data differs from the training distribution, and the model's confident predictions become unreliable. Validating on data from the deployment site, and monitoring performance continuously, are the standard responses. Capacity, hardware and convergence would have shown up in the original evaluation. A failure that appears only on deployment points at the data rather than at the model.

Answer: D

Everything else follows from the performance measure, so stating it carefully, in terms of what is actually wanted rather than what is easy to measure, is the most consequential decision in designing an agent. A measure that rewards the wrong thing produces a system that pursues it faithfully. Speed, resource use and human comparison are secondary evaluations. Specification gaming is the direct consequence of getting this step wrong, which is why it heads the PEAS checklist.

Answer: A

Telling a cleaning agent to sweep rewards sweeping rather than cleanliness, and an agent may sweep dirt about indefinitely, whereas rewarding a clean floor leaves it free to find the best method. Measuring outcomes is often harder rather than easier, and regulation is not the reason. Specifying what is wanted rather than how to achieve it is the principle behind declarative specification generally, and its violation is behind many of the field's more entertaining failures.

Answer: B

Turing set out the imitation game in his 1950 paper on computing machinery and intelligence. McCarthy coined the term artificial intelligence six years later, and Minsky and Simon were among the founders of the field that followed. Attributing the foundational ideas to the right people matters because their differing views on what intelligence is still divide the subject, and each name is attached to a distinct position rather than merely to a date.

Answer: C

McCarthy proposed the name for the 1956 Dartmouth workshop, partly to distinguish the new field from cybernetics. Turing wrote about machine intelligence without using the phrase, Shannon contributed to the workshop and to information theory, and Wiener founded the cybernetics tradition the name was chosen to separate from. A field's name is often chosen for political as much as descriptive reasons, and this one is a clear example.

Answer: D

An agent perceives through sensors and acts through actuators, and the pair defines its interface to the world. The performance measure evaluates its behaviour from outside, and an internal model is maintained by the agent from its percepts rather than being a source of them. Keeping the interface clear is what allows the same agent design to be applied to a robot, a software system or a person, which is the generality the framework was devised for.

Answer: A

Actions are executed by actuators, whether motors, displays, network messages or database writes, and the environment changes in response. Sensors only read, the percept sequence is the history of what has been read, and a knowledge base is internal. The symmetry of sensors and actuators across every kind of agent, physical or software, is what makes the agent abstraction so widely applicable.

Answer: B

The percept sequence is everything the agent has perceived to date, and in principle an agent's action is a function of it. A reflex agent uses only its last element, while a model based agent summarises the rest into internal state. The list of intended actions is a plan, and the environment's history is not directly available. Defining behaviour as a function of the whole history is what makes the framework fully general before any architecture is chosen.

Answer: C

The agent function is the abstract mapping from every possible percept history to an action, while the agent program is the concrete implementation that realises it. The function may be an infinite table; the program must be finite. Distinguishing specification from implementation at the outset is what allows the same behaviour to be realised by very different architectures and compared fairly.

Answer: D

The function describes what the agent does in every possible circumstance, and the program is the finite artefact that computes it within the agent's resources. Two very different programs may implement the same function with different efficiency. Documentation, evaluation and synonymy are not the distinction. Separating what a system should do from how it does it is a principle borrowed from software engineering and it applies here with particular force.