Introduction to Artificial Intelligence Questions and Answers

Answer: C

Because the reasoner is general and the knowledge is data, a domain expert can extend the system by adding sentences rather than by programming, which was the great promise of the knowledge based approach. Speed is usually worse than a hand coded program, memory is not saved, and evaluation remains essential. Making a system extensible by people who are not programmers is an architectural goal that recurs whenever domain expertise and software expertise sit in different heads.

Answer: D

The inference engine applies the rules of the representation language to the knowledge base, deriving conclusions by forward or backward chaining. It is domain independent, which is what allows one engine to serve many knowledge bases. Accelerators, acquisition tools and user interfaces are separate components of a complete system. Naming the parts of the classical architecture correctly makes the modern contrast with end to end learned systems easier to state precisely.

Answer: A

Forward chaining starts from what is known and applies rules to derive everything deducible, which suits situations where new data arrives and its consequences are wanted. Backward chaining starts from a query and is more efficient when only one conclusion matters. State space search and neural propagation are different processes sharing directional vocabulary. Choosing the direction of inference according to whether data or goals drive the system is a practical decision with large efficiency consequences.

Answer: B

Backward chaining works only on the rules relevant to the query, so it avoids deriving the vast number of true but irrelevant conclusions that forward chaining would produce. Continuous data arrival favours forward chaining instead. Explanation is available from either, and a knowledge base without rules supports no chaining at all. Directing computation by what is actually wanted, rather than computing everything and selecting afterwards, is a principle far wider than inference.

Answer: C

A brittle system performs well within its intended envelope and then degrades abruptly rather than gracefully when conditions move slightly outside it, which was the characteristic failure of rule based expert systems. Graceful degradation is the property wanted instead. Maintenance burden, data sensitivity and portability are separate weaknesses. How a system behaves at the edge of its competence matters at least as much as how it behaves at the centre, and it is much less often measured.

Answer: D

Real conditions always eventually depart from those assumed, and a system whose accuracy declines steadily is far safer than one that fails abruptly and without warning. Knowing when it is outside its competence, and saying so, is better still. Safe shutdown, resource use and infallibility are different properties. Designing for the boundary of the operating envelope rather than only its centre is what separates a deployable system from a demonstration.

Answer: A

A shared benchmark makes results comparable and progress measurable, which has driven much of the field's advance, but a method that tops a benchmark has been shown to do exactly that and no more. Overfitting to a benchmark through repeated use is a recognised hazard. Theoretical limits come from analysis rather than from measurement. A measure that becomes a target ceases to be a good measure, and the field has learned that lesson repeatedly.

Answer: B

Repeated evaluation on the same data amounts to fitting the test set through the accumulated choices of the community, so reported gains overstate what would be seen on fresh data. Held out challenge sets and periodic replacement are the usual responses. Hardware, cost and data integrity are practical issues rather than the fundamental one. Selection pressure applied to a fixed target eventually optimises for the target rather than for what it was meant to represent.

Answer: C

The term denotes a system able to learn and perform across the full range of tasks a person can, transferring competence to genuinely novel problems, and nothing existing approaches it. Multi-domain training, integration of narrow systems and general architectures are all steps some believe lead towards it, and none constitutes it. Keeping an aspiration clearly separate from current capability is what allows both to be discussed sensibly.

Answer: D

The field has a long record of expecting general intelligence within a decade and of underestimating specific advances that arrived suddenly, so confident forecasts in either direction have a poor track record. Publication is extensive, measures exist though they are imperfect, and hardware growth has been comparatively predictable. A domain whose history contains both overoptimism and surprise justifies caution about any confident timeline, including a pessimistic one.