Intelligent Agents and Architectures

• The lecture introduces Intelligent Agents and their architectures
• Luís Paulo Reis is the director of LIACC, Associate Professor at DEI/FEUP, and President of APPIA, at the University of Porto

Agent Concept

• Agents perceive their environment through sensors
• Agents act upon their environment through actuators
• A Human agent's sensors include eyes, ears, nose, touch, and taste
• Actuators for humans include legs, arms, hands, and other body parts
• A Robotic agent's sensor includes cameras, infrared sensors, range finders, and microphones
• Actuators for robots include motors, wheels, and speakers
• Agent actions at any given time depend on all past percepts

Rational Agents

• A Rational Agent does the right or correct thing
• Correct actions are those that ensure an agent is more successful
• Success can be measured by evaluating actions (e.g., a Robot Driver, Chess Playing Program, Spam email classifier)
• An Ideal Rational Agent acts to maximize performance based on knowledge and perceptions
• Rationality differs from omniscience, which is "all-knowing"

Intelligent Agent Functions

• The agent function maps the history of perceptions into actions with the formula [f: P* → A]
• The agent program runs on physical architecture to produce the agent function
• Agent = Architecture + Program
• The agent function is an abstract mathematical description
• The agent program is a specific implementation within a physical system

Structure of Intelligent Agents

• Agents exhibit actions based on sequences of perceptions
• Al Task involves designing Agent programs and architecture
• An agent is comprised of an Architecture and a Program
• Software Agents differ from Physical Agents

Agent Requisites

• Agents must perceive environments via sensors
• Agents are expected to decide on actions to execute
• Agents execute actions in environments via their actuators
• Agents react to sensors and control actuators
• Unlike Objects, Requisites decide what to do

Agent Program Types

• Agents interact with environments using sensors and actuators
• Internal Data Structures are updated using perceptions
• Updated internal structures are used to determine actions

Types of Agents

• Simple reflex agents
• Agents representing the world
• Objective-based agents
• Utility-based agents
• Learning Agents

Vacuum cleaner example

• A Vacuum cleaner World provides a simple illustration of agent functions
• An agent perceives Place and content, e.g., [A, Dirty]
• Possible agent actions are Left, Right, Suck, or NoOp
• A simple agent function dictates that if the current state is Dirty, then suck; otherwise, move to another square

Nature of Environments

• When designing an agent, the initial step is characterizing its task and the task environment
• PEAS stands for Performance measure, Environment, Actuators, Sensors
• Setting must be specified before an intelligent agent is designed
• An agent design includes specification of Performance Measure, Environment, Actuators, and Sensors

The Nature of Environments: PEAS Examples

• For a medical diagnosis system
  • Performance is a healthy patient and reduced costs
  • The environment is a patient, hospital, and staff
  • Actuators include the display of questions, tests, diagnoses, and treatments
  • Sensors include touchscreen/voice entry of symptoms and findings
• For a taxi driver agent
  • Performance includes safety, speed, legality, comfortable trip, and maximized profits
  • The environment includes roads, traffic, pedestrians, and customers
  • Actuators include a steering wheel, accelerator, brake, signal, and horn
  • Sensors include cameras, sonar, speedometer, GPS, odometer, engine sensors, and a keyboard
• For a part-picking robot
  • Performance is measured by the percentage of parts in correct bins
  • The environment includes a conveyor belt, parts, and bins
  • A jointed arm and hand act as actuators
  • Cameras, tactile sensors, and joint angle sensors all serve as sensors

Properties of Environments

• Environments can be Accessible or Inaccessible
  • Accessible environments have sensors that detect everything relevant
• Environments can be Deterministic or Nondeterministic
  • Deterministic environments have next states determined by the previous state and agent actions
• Environments can be Episodic or Non-Episodic
  • Episodic environments can be divided into episodes, where a future action does not rely on past actions
• Environments can be Static or Dynamic
  • Dynamic environments change while an agent is thinking
• Environments can be Discrete or Continuous
  • Discrete environments have a finite number of perceptions and actions
• Environments can have a Single Agent or Multi-Agent
  • In a environment with a single agent, there is only one agent operating
  • In a environment with multiple agents, there are several of agents acting cooperatively or competitively

Environment Types

• The environment type affects agent design
• The simplest environment has the following parameters
  • Fully observable, deterministic, episodic, static, discrete, and single-agent
• Most real situations are
  • Partially observable, non-deterministic (stochastic), sequential, dynamic, continuous, and multi-agent

Example Agent Types

• Simple reactive/reflex agents
• Agents with world representation
• Objective-based agents
• Utility-based agents
• Learning Agents
• BDI Agents

Simple Reflex Agents

• Simple Reflex Agents use condition-action rules tables

Condition-Action Rules

• Makes a direct link between current perception and action

Agents with World Representation

• Agents maintain an internal state comprised of the agent's representation of the world

Objective Based Agents

• Goal definition to the state of the world

Utility Based Agents

• Maps current state to a value

Learning Agents

• Improves over time by using machine learning

BDI Agents

• Incorporates Beliefs, Desires, and Intentions

Multi-Agent Systems

• Composed of multiple agents exhibiting autonomous behavior and interoperability
• MAS are useful as a natural solution to distributed problems

Multi-Agent Systems Motivation

• Distributed knowledge/information
• Addressing problem dimensions
• Human-machine interfaces
• Project clarity/simplicity
• Legacy systems
• Efficiency
• Robustness/scalability
• Problem division
• Information privacy

CONTROLEX Agent

• Designed to control room temperature using T1 and T2 as room/outside temperature perceptions
• Actions include AQ to turn on the heater, NAQ for turning off the heater, AC to turn on the cold air, NAC for turning off cold air, AJ for opening windows, and NAJ for closing windows
• It is intended to keep the room temperature between 22 and 24 degrees
• When possible, an agent should use the windows to control temperature to avoid wasting energy
• When temperature is more than 2 degrees away from the desired band (below 20 or above 26 degrees), windows must be closed and the air or heater used

POOLEX Agent

• It is designed to control water level and temperature of a swimming pool -
• It is intended the pool's temperature is between 25 and 27 degrees and the water level is between 1.3 and 1.5 meters
• TEMP perceptions corresponding to the pool temperature and ALT corresponding to the pool's water height
• Actions include: AS (Open water outlet), FS (close water outlet), AEQ (Open hot water inlet), FEQ (close hot water inlet), AEF (Open cold-water inlet and FEF - close cold-water inlet)
• It is possible to connect one of the inlets and the water outlet simultaneously, but there is no guarantee that the level will remain the same
• The water inlet cannot be connected when the water level is above 1.45 meters

RATEX Agent

• The agent is a Robot that tries to solve a simple maze
• Agent PEAS and its environment are classified
• A simple algorithm to solve mazes with the agent is used
• Agent only wants to move in the maze without hitting any other robot

RATEX Sensors

• 2 drive wheels (MLef, MRig)
• 3 proximity sensors (SLef, Scent and SRig)
• 1 floor sensor (SGround)
• 1 headlight sensor (SLight)