Intelligent Agents Questions and Answers

Answer: C

When today's choice constrains tomorrow's options, the agent must reason about futures rather than merely respond, which brings in search, planning and the whole apparatus of lookahead. State count is not the criterion, since a large episodic problem can be easy. Percept rate and evaluation timing are practical rather than fundamental. Interaction between decisions is what turns a classification problem into a control problem, and the difficulty rises sharply at that point.

Answer: D

Brooks proposed building a robot from simple behaviours such as avoid obstacles and wander, each connected directly from sensors to actuators, with higher layers suppressing lower ones when appropriate and no central world model at all. It produced robust insect like behaviour and challenged the symbolic tradition directly. Goal decomposition, agent delegation and class hierarchies are different ideas. An architecture that deliberately omits the world model is a sharp test of whether one is necessary.

Answer: A

Brooks argued that maintaining and updating an internal model is expensive and error prone, and that sensing the world directly at each moment is both cheaper and more robust. The resulting robots were reactive and fast but found it hard to pursue extended goals. Learning, biological fidelity and verification are separate positions. A claim that removes a component everyone assumed necessary is valuable even if it is only partly right, because it forces the assumption to be examined.

Answer: B

Without internal state or lookahead, a reactive system cannot plan a route around an obstacle it can no longer see, or carry out a task requiring several steps in a particular order. Its speed and low cost are precisely its advantages. That limitation led to hybrid architectures combining a reactive layer for immediate response with a deliberative layer for planning. Recognising where an approach stops working is what motivates the architecture that succeeds it.

Answer: C

A three layer architecture typically has a fast reactive layer handling immediate hazards, an executive sequencing layer, and a slow deliberative planner, each operating on its own timescale. It is the standard arrangement in serious robotics. Symbolic and neural integration is a separate concern, as are shared embodiment and program switching. Assigning different timescales to different layers is the organising principle, and it resolves the conflict between reacting quickly and thinking carefully.

Answer: D

An obstacle avoidance reflex must act within milliseconds, while route planning may take seconds and can be interrupted, so separating them lets each run at its natural rate without the slower blocking the faster. Memory, hardware and regulation are implementation details rather than the reason. Matching the response time of a component to the deadline of the problem it solves is what makes real time systems tractable at all.

Answer: A

The BDI model separates the agent's picture of the world, its objectives and the plans it has committed to, so that commitment persists rather than being reconsidered at every step. That persistence is the model's central contribution, since an agent that reconsiders constantly achieves nothing. Hardware stages, architectural layers and decision quantities are different decompositions. Distinguishing a desire from an intention captures something real about acting under limited time.

Answer: B

An agent that reconsidered every option at every step would never accomplish anything, so adopting an intention closes off deliberation and directs effort until the intention is achieved or explicitly dropped. Knowing when to reconsider is the model's hard problem. Learning, achievability and ownership are not the distinction. Committing to a course of action, at the cost of ignoring alternatives, is a genuine feature of bounded rationality rather than a defect of it.

Answer: C

Simon's insight was that real agents cannot compute the optimal action in the time available, so rationality must be defined relative to their resources rather than as perfect optimisation. Satisficing, choosing a good enough option, follows from it. Spatial bounds, utility caps and goal limits are unrelated. Redefining a normative standard so that it can actually be met is what made decision theory applicable to agents that must act in real time.

Answer: D

Rather than searching for the best, a satisficing agent stops as soon as it finds something good enough, which is often the rational thing to do when search itself is costly. It follows directly from bounded rationality. Exhaustive maximisation is the idealisation it replaces, constraint satisfaction is a different problem class despite the similar word, and minimal action count is one possible objective. Knowing when to stop searching is itself a decision, and treating it as one is the insight.