Source

Adapted from my own handwritten and hand-typed notes from Year 1 to Year 3 of my EEE degree

An agent is an entity that:

  • perceives its environment through sensors (eyes, ears, cameras, infrared range sensors)
  • acts on the environment through effectors (hands, legs, motors)

The stream of inputs is the percept sequence, and the agent maps it to actions.

Rational Agent

A rational agent does the right thing. A rational action maximises the expected value of an objective performance measure, given the percept sequence to date.

Rationality depends on:

  • the performance measure (e.g. floor area cleaned for a cleaning robot)
  • everything the agent has perceived so far
  • built-in knowledge about the environment
  • the actions that can be performed

Rational does not mean all knowing: the agent maximises expected performance with what it knows.

Autonomous Agent

An autonomous agent does not rely entirely on built-in knowledge, it adapts to the environment through experience. A purely pre-programmed agent only works when all its built-in knowledge is correct. Example: a driverless car that learns in a driving centre, then NTU, public roads, highways and City Hall.

Describing the Task

Driverless taxiMedical diagnosis system
Perceptsvideo, speed, acceleration, engine status, GPS, radarsymptoms, findings, patient's answers
Actionssteer, accelerate, brake, horn, displayquestions, medical tests, treatments
Goalssafety, reach destination, profit, obey laws, comforthealthy patient, faster recovery, low cost
EnvironmentSingapore streets, highways, traffic, pedestrians, weatherpatient, hospital, clinic

Agents differ in how much they store and plan, see Agent Types.