COMP9414 Chap.2 Rational Agents and Task Models
Rational Agents and Task Models
An agent perceives, decides and acts in an environment, but those verbs become computational only after the model names percepts, legal actions, relevant world features and the preferred outcome. This chapter turns that loop into a task specification.
It connects environment properties to memory and monitoring needs, distinguishes goal satisfaction from utility, and shows why representation errors cannot be rescued by a more sophisticated search algorithm.
What this chapter covers
- 01
Percepts entering an agent from its environment
- 02
Actions changing the world and future observations
- 03
Fully and partially observable task models
- 04
Deterministic and stochastic action outcomes
- 05
Static and dynamic environments
- 06
Goals, path costs and utility as different roles
Worked example · free
Formulate a library-return agent
- +1Record cart location, remaining boxes, loaded status, battery band and latest corridor status in the changing state.
- +1Use location, box, load and corridor observations as percepts; define move, collect, unload and recheck actions.
- +1Require both boxes cleared, the cart at sorting and no crate still loaded for the goal.
- +1Accumulate travel time with a penalty for missing closing, and re-sense because the corridor is dynamic and partially observed.
Key terms
- Agent
- A decision-making system that receives percepts and can affect its surroundings through actions.
- Percept
- Information supplied to an agent about the environment at a point in its interaction.
- Rational action
- An action selected to maximise expected utility under the model and information available.
- Partial observability
- A condition where current percepts do not reveal the complete task-relevant state.
- Utility
- A preference measure used to compare outcomes or trade-offs among possible actions.
Rational Agents and Task Models FAQ
What makes an agent rational?
Rationality is relative to the stated objective, model and information. An agent can make the best available decision under uncertainty and still obtain a poor result. Evaluation should distinguish decision quality from luck and check whether the supplied utility matches the intended outcome.
Why is the environment more than a map?
Relevant objects, other agents, changing conditions and hidden information can alter action legality or outcome. A spatial graph captures location, but a delivery task may also require parcel status, capacity, door state and observations that update while execution proceeds.
How do environment properties change a design?
Partial observability creates a need for history, belief or sensing; stochastic effects require contingencies or expected utility; dynamic change requires monitoring; continuous variables require an abstraction or motion representation. A useful classification changes the proposed action or stored information.
Exam move
Take a familiar autonomous system and draw the full agent-environment loop. Label the incoming percepts, outgoing actions and world changes that can occur without the agent. Classify the task along observability, determinism, change and discreteness, but attach one design consequence to every classification.
Next write a state proposal and test it with paired histories: if two histories map to the same state, should every future action have the same meaning? Add any missing variable that changes legality, goal satisfaction or cost. Separate the goal test from utility by constructing two goal-reaching plans with different costs.
Finish by naming one stale or missing observation that could make a rational offline choice fail during execution and specify the monitoring action that would detect it. On a second worksheet, distinguish facts the agent observes from facts the modeller knows. This catches answers that quietly grant the controller complete access to the environment.
For each percept, note its possible error or delay and decide whether the state should include a history, confidence or last-observed time. Then perform an objective audit: invent behaviour that scores well under the written utility but violates the intended purpose. Repair the metric or add a constraint and explain why the change alters rational action.
Practise transferring the same physical setting between tasks, such as navigation, delivery and emergency response, to see how goals and relevant state variables change. End by writing one sentence that distinguishes decision quality from outcome quality under uncertainty. These exercises turn environment labels into design consequences and prepare you to formulate search problems without hidden information.
Add a counterfactual pass in which one environment property changes while the physical setting stays fixed. Explain which sensor, state variable or policy assumption must change and which parts of the original design remain sound. Use a two-column error log for information failures and objective failures, because they require different repairs.
Finish by checking that every claimed percept is genuinely available to the agent at decision time rather than supplied only by the problem author.
Working through Rational Agents and Task Models in COMP9414? Sia is AskSia’s AI Computer Science tutor — ask any COMP9414 Rational Agents and Task Models question and get a clear, step-by-step explanation grounded in how COMP9414 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
