Intelligent Agents Questions and Answers

Answer: D

A measure that correctly identifies the desired outcome may still be satisfiable in a way nobody anticipated, and a capable optimiser will find that route. Adding constraints, penalising side effects and testing in simulation are the usual mitigations. Precision is achievable with effort, computability is a separate concern, and measures apply to both task types. The gap between specifying an outcome and specifying an acceptable way of reaching it is where most practical difficulty lives.

Answer: A

The framework's contribution is the discipline of writing down the performance measure, the environment and the interface first, because every later decision depends on them and errors there cannot be recovered by good implementation. Sophistication should be chosen to match the environment rather than maximised, learning is one technique among several, and incremental building is a development strategy rather than a design principle. Specifying the problem before the solution is the whole of it.

Answer: B

Observability, determinism, dynamism and the presence of other agents each demand a capability, and the architecture must supply all of those the environment requires and need supply no more. Language, data and preference are implementation matters. Letting the characteristics of the problem select the architecture, rather than choosing an architecture and hoping the problem suits it, is the practical payoff of the whole classification exercise.

Answer: C

The definition is deliberately broad: anything with sensors and actuators, whether a thermostat, a robot, a program or a person, counts as an agent, and the framework applies to all of them. Autonomy, embodiment and learning are properties some agents have. Adopting a definition wide enough to cover the whole field, and then classifying within it, is what gives the abstraction its reach.

Answer: D

The environment is whatever the agent's percepts come from and its actions affect, which for a software agent may be a file system and a network rather than any physical space. Rooms, platforms and other agents are particular kinds of environment or parts of one. Drawing the boundary between agent and environment is a modelling choice, and drawing it in a different place changes what counts as a percept and what counts as internal state.

Answer: A

The boundary is a modelling decision: a robot's arm can be treated as part of the agent or as part of the environment it must control, and each choice makes different problems easier. Physical surfaces, program boundaries and sensor ranges are conventional places to draw it but none is required. Recognising a modelling choice as a choice, rather than as a fact about the world, is what allows it to be made deliberately.

Answer: B

The task environment is the PEAS specification taken together, and it is the problem to which an agent design is the solution. Deployment setting, task list and test harness are related but different. Naming the whole specification as a single object encourages it to be settled and recorded before design begins, which is precisely what the framework is trying to achieve.

Answer: C

Partial observability demands memory, stochasticity demands probability, dynamism demands timeliness and other agents demand strategy, so the classification is effectively a requirements list for the architecture. Standards, simulation and language choice are unrelated. A taxonomy that translates directly into design requirements earns its keep, whereas one that merely labels does not, and this is a good example of the former.

Answer: D

Discreteness applies separately to the state, to time and to the percepts and actions, and a chess game is discrete in all of them while driving is continuous in most. Sampling regularity, object count and action latency are different notions. The distinction matters because many algorithms assume enumerable states or actions, and a continuous problem must either be discretised or handled by methods designed for it.

Answer: A

The state includes positions and velocities that take any real value, and the steering and throttle inputs are likewise continuous, so no finite enumeration is possible. That forces either discretisation or continuous optimisation. Journey continuity, physical contact and changing conditions describe other properties, the last being dynamism. Identifying which variables are continuous determines which parts of the problem need which class of algorithm.