Intelligent Agents

PEAS Framework

> [!important] PEAS Framework > > - Performance measure > - Environment > - Actuators > - Sensors

Taxi driver

P: Safety, speed, legal regulations, comfort, profit, impact on other road users E: Roads, other traffic, police, pedestrians, customers, weather A: Steering, accelerator, brake, signal, horn, display S: Cameras, radar, speedometer, GPS, engine sensors, accelerometer, microphones, touchscreen

Performance Measure

Performance measures are things to consider for the agent, such as:

• fit (Who is the agent best for?)
• purpose (What is being optimised)?
• information available
• side-effects (What, if there are, are unintended effects)
• costs

Rational agents

Agents that choose actions that maximise performance measure.

Environment

Properties of the task environment

• Fully vs partially observable
• Episodic vs sequential

Observability

A fully observable environment allow its’ agents’ sensors to give it access to the complete state of the environment at each point in time.

A chessbot's environment is fully observable

The agent is able to see the whole state of the environment, which is all the pieces on the chessboard, and their positions at one time.

A smart driving system's environment is partially observable

The agent is only able to see the state of the environment around them, and not the state of the environment further away from them (limited information)

Episodic (vs sequential)

An episodic agent’s experience is divided in different atomic episodes, consisting of a perception and an action, and the choice of action depends only on the episode itself.

If this is not followed, the agent is thus sequential.

A chessbot's environment is sequential

The choice of move made depends on the state and the moves made on the board before.

A TicTacToe bot's environment is episodic

The choice of move only depends on the current state of the game.

Static (vs dynamic)

A static environment is unchanged while deliberating.

If it changes, then it is considered dynamic. If the environment itself does not chnage with the passage of time, but the agent’s performance score does, then it is semi-dynamic.

A chessbot's environment can be static, or semi-dynamic

The chessboard does not change while the agent deliberates over the action. However, it can be considered semi-dynamic, in cases where there is a time element to the game, causing there to be a drop in performance score.

A self-driving system 's environment is dynamic

The self-driving system’s environment will change as it moves across the roads.

Discrete (vs continuous)

A discrete environment has a limited number of distinct, clearly defined percepts and actions, while a continuous environment does not.

A chessbot's environment is discrete

There are a limited number of moves that a piece can take and there is a limited number of states of the board.

A self-driving system's environment is continuous

There is an infinite amount of actions that can be taken by the system, and an infinite amount of states at any point of time.

Single-agent (vs multi-agent)

A single-agent environment dictates that the agent is operating by itself, while a multi-agent environment dictates otherwise.

A Solitaire agent is single-agent

There is only one decision maker choosing the cards.

A chessbot is multi-agent

There are two agents - as there are two sides to the game.

Agent Function

Agent function

The agent function maps from percept histories to actions :

Agent program f, given an input.

The agent program is then an implementation on the physical architecture of the agent to produce the function

Thus, the agent itself is completely specified by the agent function. These agent function can follow (a combination of) common agent structures.

Agent Structures

Simple reflex agents

The simple-reflex agent structure is a simple structure that takes the precepts, passes it through some condition-action rules to decide the action taken.

Example

A condition-action rule that can be seen is:

• if the front is movable, move
• else, move downwards

Goal-based agent structure

The goal-based agent structure evaluates the agent based on what happens if the action is done, and if it matches the goal taken.

Utility-based agent

The utility-based agent structure evaluates the agent based on how "happy" the agent is if the action is done. This utility usually has a measure of its own, which also has to be determined.

Learning agent

Exploration vs Exploitation

Agents can choose to explore or exploit based on their current knowledge.

Exploration

Learning more about the environment around it

Exploitation

Maximising gain based on the current knowledge of the environment around it.

Example

An agent can choose to go through a route it has already known (exploitation) or navigate through unknown routes to find a better route (exploration)