Intelligent Agents (COMP30024 W1 L2)

Updated: March 03, 2022

Intelligent Agents

Chess: logical, well-organized problem with finite life time and objective.
Medical diagnosis: ill-defined problem, where environment is changing. (multiple diseases, patient may not reveal everything, etc.)
Walking on two legs: real-time control, getting through obstacles, maintaining stability, etc.
Taxi driver: complex, quickly evolving environment.

How do we characterize These different types of intelligent agents?

(Characterise requirements for an agent in terms of its percepts, actions, environment, and performance measure.)

**Percepts/observations** of the environment
- Made by sensors (How does the model sense the environment?)
- Input
- How does the model sense the enviroment?
**Actions** which may affect the environment
- Made by actuators (What action does the agent perform?)
- Output
- i.e. Ability to exelerate, move, accept money, change the state of a chess board, etc.
**Environment** in which the agent exists
- What’s being sensed and modified?
- More complex, need to be modeled.
**Performance measure** of the desirability of environment states
- What’s the object applied to the agent?
Example: Automated taxi
- Percepts: video, accelerometers, temperature gauges, engine sensors, keyboard, GPS, …
- Actions: steer, accelerate, brake, horn, speak/display, …
- Environment: city streets, freeways, traffic (other vehicles, traffic lights), pedestrians, weather, customer, …
- Performance measure: safety, reach destination, maximize profits, obey laws, passenger comfort, …
Agents can be evaluated empirically, sometimes analysed mathematically.
Agent is a function from percept sequences to actions.
Ideal rational agent picks actions that is expected to maximise its performance measure (based on the percept sequence and its built-in knowledge)
- Rational != omniscent (know everything)
- Rational != clairvoyant (automatically predict the future state)
- Rational != successful (need to experiment, then do the best it can)

What is the functionality that lies inside the agent? (Choose and justify choice of agent type for a given problem.)

**Simple reflex agents**
- Assumes that the world is fully observable (no missing information)
- No long term reasoning (may result to infinite loops)
- Tasks where this type can be sufficient: automatic door (detects person in front of the sensor)
**Model-based reflex agents**
- Assumes some of the world is not observable, and uses a model to infer the state of the environment
- Long-term decision, but still needs improvement
**Goal-based agents**
- Looks ahead to achieve desirable future state
- Searches for the best action to get to the focused, specific, single goal
**Utility-based agents**
- Dealing with different goals (short & long)
- Utility: contains a measure of how desirable a future state is
  - Consider which one would maximize my happiness
  - Mapping state to a value (optimize the expected outcome)
  - More general, thinks about the trade-off between achieving different goals
- Multiple goals when trying to maximize the profit in a taxi
  - Keeping the customer happy
  - Getting the customer to the destination
  - Obeying the traffic rules (avoiding fines)
  - Not taking a route that is unnecessarily long

You should choose the agent type depending on the environment. (characterise the environment for a given problem.)

Environments may or may not be:

**Observable**
- Percept contains all relevant information about the world; Everything is in view.
- ↔ partially observable: Environments cannot be directly sensed via percept.
**Deterministic**
- Current state of the world uniquely determines the next.
- Having high certainty about the consequences (i.e. I will get to the beach by 3:30pm if I take this train.)
- ↔ stochastic: The next state is randomly determined.
**Episodic**
- Only the current (or recent) percept is relevant. (Only consider short-term actions)
- ↔ sequential: The consequences are long-term.
**Static**
- Environment doesn’t change while the agent is deliberating.
- (i.e. chess: board does not change while the agent makes a move.)
- ↔ dynamic: environment changes while making the decision.
  - (i.e. driving cars: environment changes constantly.)
**Discrete**
- There are finite number of possible percepts/actions.
- (i.e. chess: number of squares on the board, number/state of pieces are finite.)
- ↔ continuous: infinite number of possible percepts/actions.
  - (i.e. driving a car)

The environment type largely determines the agent design.
The real world is partially observable, stochastic, sequential, dynamic, and continuous.