Source
Adapted from my own handwritten and hand-typed notes from Year 1 to Year 3 of my EEE degree
Four basic agent designs, each adding something to the one before.
Simple Reflex Agent
Works by condition action rules:
- find the rule whose condition matches the current percept
- perform the action linked to that rule
- example: if car-in-front-is-braking then initiate-braking
It has no memory, so it only works when the current percept is enough to decide.
Reflex Agent with State
Same as above, but it also keeps an internal state (memory of past percepts and of how the world evolves and what its actions do). The rule is matched against the percept and the stored state.
Example: if yesterday-at-NTU and no-traffic-jam-now then go-Orchard.
Goal Based Agent
Has explicit goal information and predicts "what will the world be like if I do action A". It chooses actions that lead towards the goal.
Example: a driverless taxi at a junction decides left, right or straight based on whether it gets closer to the destination (Orchard).
Utility Based Agent
Many action sequences may reach the same goal. A utility function measures how happy the agent is in a state, and the agent picks the actions that maximise utility.
Example: going to Orchard via PIE or AYE, which one gives the lower fare.
Comparison
| Agent | Uses current percept | Internal state | Goal | Utility |
|---|---|---|---|---|
| Simple reflex | yes | no | no | no |
| Reflex with state | yes | yes | no | no |
| Goal based | yes | yes | yes | no |
| Utility based | yes | yes | yes | yes |
Which design is needed depends on the Task Environment Properties.