Intelligent Agents Questions and Answers

Answer: C

With infinitely many options, the agent must optimise over the space rather than compare a list, which brings in gradients, sampling and parametrised policies. Robot control and vehicle steering are the usual examples. Execution time, state count and measurement are separate concerns. Whether a decision can be made by enumeration or requires optimisation is one of the sharpest practical divisions in agent design and it determines the whole class of algorithm used.

Answer: D

A policy specifies an action for every state, so an agent following one need not plan afresh at each step. It is the standard object of reinforcement learning, where the goal is to find a policy maximising long term reward. Constraints, plans and performance measures are distinct concepts, though a plan is a sequence rather than a mapping. Distinguishing a policy from a plan matters because one is robust to unexpected states and the other is not.

Answer: A

A plan is a sequence that assumes everything goes as expected, whereas a policy covers every state and so responds correctly when the world departs from the plan. That robustness is why stochastic domains are usually solved for policies. Policies typically need more memory and more computation, and neither guarantees optimality of length. Choosing between a sequence and a mapping is choosing between efficiency and robustness, and the environment decides which matters.

Answer: B

In a very large space a policy cannot be represented explicitly, so the agent plans from its current state and replans when reality departs from the plan. That is the standard approach in robotics and logistics. Determinism removes the need for either, unlimited computation would favour a policy, and frequent goal change favours planning too though for a different reason. Choosing between precomputation and on demand computation is a trade of memory against time that appears throughout computing.

Answer: C

A situated agent is coupled to its environment through a continuous loop of perception and action rather than reasoning about an abstract description of it, which is the emphasis behaviour based robotics brought to the field. Fixed position, knowledge population and role assignment are unrelated senses. The contrast with a system that reasons about a symbolic description supplied to it is what gives the term its point.

Answer: D

Embodiment holds that the shape, dynamics and sensors of a body are part of the solution rather than an inconvenient wrapper, so that a well designed morphology can make control trivial. Passive dynamic walkers, which walk downhill with no computation at all, are the standard demonstration. Implementation medium, self representation and visibility are different notions. Letting physics perform part of the computation is a genuinely different design stance and not merely a philosophical one.

Answer: A

A walker with no motors, sensors or processor descends a shallow slope with a natural gait, showing that the dynamics of a well shaped mechanism can do work that a controller would otherwise have to do. It is the cleanest argument for taking morphology seriously. It does not show that controllers are unnecessary in general, nor make claims about simulation or about wheels. A demonstration that removes the computation entirely is a strong form of the embodiment argument.

Answer: B

Percepts, actions, environments and performance measures describe a chess program, a warehouse robot and a spam filter equally well, which lets techniques and vocabulary transfer between them. It guarantees nothing about behaviour, removes no need for domain knowledge and asserts no equivalence between problems. A common abstraction that makes different problems comparable is valuable for exactly that reason, and it is why the framework opens most treatments of the subject.

Answer: C

The PEAS description fixes what success means and what the agent can perceive and do, and every subsequent choice follows from it. Selecting an architecture or an algorithm first is the commonest way of solving the wrong problem well. Languages, platforms and data are implementation decisions that come later. Insisting that the task be specified before the solution is chosen is unglamorous discipline that prevents a large proportion of wasted effort.

Answer: D

Strange but persistent behaviour is usually the agent optimising exactly what it was told, so re-reading the objective is more productive than suspecting a fault. Hardware, data and architecture are worth checking afterwards. The habit of suspecting the specification before the implementation is the practical form of everything the alignment and specification gaming literature has to say, and it costs nothing to adopt.