Modeling AI systems with Agents

Modelling AI Systems

• One way to model an AI System is by using agents.
• An agent makes decisions or arrives at conclusions; for example, a person, machine or software.

Agents: A Formal Definition

• An agent perceives its environment through sensors and acts upon it through actuators.
• Sensors of a human agent include: eyes, ears, nose, hands, and skin.
• Actuators of a human agent include: mouth, hands, feet, and other body parts.
• Sensors of a robotic agent include: cameras and infrared range finders.
• Actuators of a robotic agent include: robotic arms and various other motors.

Agents and Environments

• The term percept refers to an agent's perceptual inputs at any given instant
• Percept sequence is the complete history of everything the agent has ever perceived.
• The agent function maps from percept histories to actions.
• f: P* => A
• The agent program runs on the physical architecture to produce f.
• Agent = Architecture + Program

Vacuum-cleaner World Example

• The agent example is a Vacuum cleaner.
• Percepts are the location and contents ie. [A, Dirty].
• Actions are Left, Right, Suck, NoOP.

Vacuum-cleaner Agent

• The agent has a list of actions based on the rooms being clean or dirty:
• [A, Clean] Action = Right
• [A, Dirty] Action = Suck
• [B, Clean] Action = Left
• [B, Dirty] Action = Suck
• [A, Clean], [A, Clean] Action = Right
• [A, Clean], [A, Dirty] Action = Suck
• [A, Clean], [A, Clean], [A, Clean] Action = Right
• [A, Clean], [A, Clean], [A, Dirty] Action = Suck

Rational Agents

• Building any agent/AI system should aim to be rational.
• A rational agent strives to "do the right thing", based on perception and performable actions, choosing the most successful action.

The Right Action and Performance Measures

• Performance measures should align with desired environmental outcomes, not preconceived agent behaviours.
• Deciding if an agent is doing the right thing involves measuring the outcome and cost of actions.
• The vacuum-cleaner agent performances can be measured by the amount of dirt cleaned up, the amount of time taken, electricity consumption, and noise generated.

Actions of Rational Agents

• A rational agent selects actions expected to maximize its performance measure. This decision is based on evidence and built-in knowledge.
• An agent is autonomous if its behavior stems from its own experience, including learning and adaptation abilities.

PEAS: Performance, Environment, Actuators, Sensors

• An agent's PEAS collectively refers to its task environment.
• When designing an agent, defining the task environment as fully as possible is the first step.

PEAS Examples

Automated Taxi

• Performance: Safe, legal, fast, comfortable, maximizes profits
• Environment: Roads, other traffic, pedestrians, customers
• Actuators: Steering wheel, accelerator, brake, signal, horn, display, screen
• Sensors: Cameras, speedometer, GPS, odometer, engine sensors, microphone, touch screen

Medical Diagnosis System

• Performance: Healthy, minimal costs, no lawsuits
• Environment: Patient, hospital, staff
• Actuators: Screen display (questions, tests, diagnoses, treatments, referrals)
• Sensors: Touchscreen/voice for entry of symptoms and findings

Part-Picking Robot

• Performance: Percentage of parts in correct bins
• Environment: Conveyor belt with parts, bins
• Actuators: Jointed arm and hand
• Sensors: Camera, joint angle sensors

Interactive Math Tutor

• Performance: Maximize students’ scores on tests
• Environment: Students
• Actuators: Screen display (exercises, suggestions, corrections)
• Sensors: Keyboard, touch screen, other input devices

Properties of Task Environments

• Fully Observable (vs. Partially Observable): The agent's sensors have complete access to the environment’s state at any given time (Chess vs Poker).
• Deterministic (vs. Stochastic): The environment's next state is fully determined by the current state and agent's action (Chess vs throwing dice).
• Episodic (vs. Sequential): The agent's experience is divided into atomic "episodes" (perceiving and performing a single action); action choice depends only on the episode itself (defective parts robot vs. chess).

Environment Types

• Static (vs. Dynamic): The environment remains unchanged while the agent is deliberating (Chess vs. Taxi driving). Semidynamic environments do not change with time but do affect the agent’s performance (Chess with a clock).
• Discrete (vs. Continuous): A limited or finite number of distinct, clearly defined percepts and actions.
• Single agent (vs. Multiagent): An single agent is operating in the environment on its own.

Environment type examples

Chess with a clock

• Fully observable: Yes
• Deterministic: Yes
• Episodic: No
• Static: No
• Discrete: Yes
• Single agent: No

Chess without a clock

• Fully observable: Yes
• Deterministic: Yes
• Episodic: No
• Static: Yes
• Discrete: Yes
• Single agent: No

Taxi Driving

• Fully observable: No
• Deterministic: No
• Episodic: No
• Static: No
• Discrete: No
• Single Agent: No

More Task Environment Examples and Properties

• Crossword puzzle is observable, deterministic, episodic, static, and discrete.
• Chess with a clock is observable, deterministic, episodic, static, and discrete.
• Poker and Backgammon is unobservable, deterministic, episodic, static, and discrete.
• Taxi driving and Medical diagnosis is unobservable, deterministic, episodic, static, and discrete.
• Image Analysis and Part-picking robot is unobservable, deterministic, episodic, static, and discrete.
• Refinery controller is partially observable, single agent, stochastic, sequential, dynamic and continuous.
• English tutor is partially observable, multi agent, stochastic, sequential, dynamic and discrete.

Agent Types

• The four basic types of agents from least to most complex include:
• Simple reflex agents
• Model-based reflex agents
• Goal-based agents
• Utility-based agents
• Learning agents

Simple Reflex Agents

• The simplest agent type selects actions based on the current percept, ignoring the past.
• Reactive agents have no memory.
• An intelligent car wiper or vacuum agent are examples, as decisions are based on the current location and the presence of dirt.

function REFLEX-VACUUM-AGENT((location,status)) returns an action
   if status = Dirty then return Suck
   else if location = A then return Right
   else if location = B then return Left

Model-Based Reflex Agents

• Reflex agents do not work well in partially observable environments.
• The most effective way to handle partial observability is for the agent to keep track of the part of the world it can't see now and to maintain it's internal state.
• The agent needs to have a model of the world.

Goal-Based Agents

• Knowing the current state is not always enough to decide what to do.
• Goal information describes desirable situations, like reaching a passenger's destination.
• Search and planning are often employed to fulfill a goal.

Utility-Based Agents

• Goals alone are not enough to generate high-quality behavior in most environments.
• Goals provide a binary distinction between “happy” and “unhappy” states.
• A more general performance measure allows a comparison of different world states, according to exactly how happy they would make the agent.

Learning Agents

• Learning allows an agent to operate in initially unknown environments.
• Learning makes agents more competent than what its initial knowledge alone might allow.