Intelligent Agents and AI Methodology

• The main goal of AI is to create a rational system that acts like a human when faced with a task.
• Due to the lack of defined methodology, researchers used different criteria to achieve this goal.

Agents

• Julián Vicente and Vincent Botti, two prestigious researchers, attempted to define a universal methodology, calling it "Intelligent Agents or Paradigm of Agents" in 1999.
• These are autonomous systems that provide solutions to problems by dividing the problem into several sub-problems and coordinating with other agents.
• It combines three areas of knowledge: Software Engineering, Distributed Systems, and, Artificial Intelligence

Software Engineering

• A software approach similar to object-oriented engineering, where a problem is broken down into independent software entities
• This involves encapsulation, independence, message passing between objects, classes, and inheritance.

Distributed Systems

• It has physical components that are located across multiple locations.
• A communication network connects the components via message passing.
• It involves sharing data, connectivity, networks, protocols, and the Internet to communicate

Object vs Agent:

• Objects have the ability to send messages to other objects
• Example: Java's private and public methods
• Objects are aware of their state but do not have control over their behavior
• Objects cannot prevent others from using their methods

Agents

• They communicate with other agents and request actions on their behalf
• Objects simply do what is asked of them, while agents do not
• Agents can be reactive, proactive, or social
• In the case of multi-agents, each agent has its own execution thread

Agent Definition

• Russell & Norvig: An agent perceives and acts on its environment via sensors and actuators.
• Maes: It refers to an autonomous system within a dynamic, complex environment that acts to fulfill goals.
• Wooldridge & Jennings: An agent is a computational system located in an environment that acts autonomously to achieve its goals.

Agents and the environment

• Agents gather information via sensors and act via actuators
• They execute complex tasks that benefit the user
• There are software programs involved in the agent, though not all computing software are agents.
• Key differences are the: Autonomy, and The ability to face different situations with reactive, proactive, and communicative responses.

Reactivity (Reaction)

• Reactive agents lack a symbolic model of the environment but act on their current information
• Their goal is to maintain a warm home without wasting energy.

Proactivity

• Agents anticipate events and take initiative, in order to achieve final goal
• Their goal is to maintain a warm home without wasting energy.

Communication

• Agents interact with other agents to complete their tasks or assist others.
• They use Agent Communication Language (ACL) to communicate
• Their goal is to maintain a warm home without wasting energy.

Simple Reactive Agent Structure

• These agents possess limited intelligence, making decisions solely based on their current perception
• The process only considers the current state
• Knowledge is stored in condition-action rules

Simple Reactive Agent Design

• The Interpret-input function translates perception into an abstract description of the current state
• The Rule-match function applies the first rule from the set that matches the current situation.
• The problem is that infinite loops may occur in partially observable environments due to lack of memory
• Choosing a random action is not rational.

Simple Reactive Agent Example

• Obstacle-avoiding mobile robot w/ proximity sensors: front, right, left.
• Possible actions: forward, turn right, turn left
• The robot's behavior is defined in a table, in modules.

Reactive Agent: Summary

• It is based on the current state to choose an action
• Does not consider consequences
• Reactivity is not necessarily arational.

Planning Agents

• "What happens if..."
• Make "decisions" on implications or hypothesis
• Have model of how the environment responds

Features

• They take into account how the environment will be in the future.
• Optimal versus Complete planning

Planning Agents

• Memory-based -(models) store past states
• Goal-oriented - possess information regarding a goal state needed to achieve
• Utility-oriented - have decision-making resources used with all of the prior information stated

Architecture of an agent based on models

• Takes into account the partial visibility of environments by storing previous states to internal memory
• This requires coding two types of knowledge: evolution, and action

Design Example

• The agent has Reflex-Agent-With-State
• It stores actions, condition-action rules, a description of the current world
• Knowledge of internal state is not always sufficient
• For example, knowing which of two possible actions to take is a good step
• A goal needs to be set

Agents Example Design

• An obstacle-avoiding mobile robot
• It uses its range sensors, to move forward or turn right and left
• It will contain a table to store this, or it may generate a "map" via movement based on previous actions and environments for this robot

An Architecture with a Goal Structure

• Utilizes goals and models to know what to do, and how to achieve them

Agent Design Goals

• It takes an action sequence to implement the process, which may include searching the area, and planning
• This leads to lower agent efficient, but greater maneuverability
• When achieving a number of "goals", are all solutions equal?

Examples

• This is seen in the robot moving through an environment to map its path
• The action, needs to be driven toward a "fixed" goal

Utility-Based Model Architecture

• In addition to the model and goal, it utilizes a "utility" function that indicates a reward for each state
• It chooses and action that maximizes return

Utility:

• A robot avoids a collision during a utility function.

Other Features

• Franklin and Nwana added extra descriptors not required to the agent.
• They stated the agent should have : Persistence, Rationality, Adaptability, Mobility, Truthfulness (veracity), Good Will

More Efficient Agents: Learning

• Automatic Learning
• Programs enhance behavior based on experiences (software)
• Learning from experience is a continuous mode
• It can come from labeled examples, for supervised learning
• Or though trial an error known as reinforcement learning
• Ultimately, any kind of learning enhances the way that the agent responds

Examples Include:

• Autonomous Car
• Traffic controls
• Stock market
• Video game bot

Environments

• Types of Agents is conditional to the Agents Design
• Fully/Partial-Visibility
• Requires memory
• Discreet/Continuous
• The limited or endless amount of combinations it may produce
• Stochastic/Determinist
• Occurs through desynchronized events
• Episodic/Sequential
• With a need to plan longterm
• Static/Dynamic
• Solo or Multi agent
• Is it self-contained or requiring help?

Visibility

• The agent is complete if it is able to use enough information

Visibility

• Agents require information to proceed
• Real-world interactions makes this increasingly unique

Discreet/Continuous

• Finite action
• An environment can be limited, making it a smaller role
• However more complex agents may not use this
• Example is taxi driving, driving angels are endless
• Discreet environments allows more user creation

Determinist VS Stokastik

• Deterministic environments provides a controlled situation
• Deterministic situations provide more organized and repeatable data
• Stokastik environments have more user creation, and less reliability

Episodic Sequences

• Occurrences that take place in a moment to define what an Agent does
• The end is predetermined once the first "episode" begins
• Sail classification is included, and is episodic by nature
• In reality, other sequences may not be able to be separated, so what is done now, may affect decisions in the future
• Example: video games

Dynamic Vs. Static

• Static environments help maintain steady thought processing
• Board games are static
• Dynamic environment have a constant changing environment that needs monitoring and a quick witted response
• Dynamic agents take more time
• Having a static environment help improve the creation

Multi-Agent

• It requires many agents connected together for function and design.
• Here are several multi-agent categories that allow all of this to coordinate smoothly:
• Cooperation
  • Smooth functions require these agents to communicate
• Resolving Conflict
  • In the event there is conflict, negotiation resolves most "attacks"
  • There are a number of negotiation systems that resolve most problems through coordinated creation in different roles in coordination
• Knowledge Sharing
  • This allows proper access to knowledge and coordination

Agent Environments

• Agents can be fully or partially observed, discrete or continuous, stochastic or deterministic, episodic or sequential, static or dynamic, and single or multi agent.

Architectures of Agents

• There are a number of AI Architectures which include these:

Deliberative

• AI that have been here awhile

• The idea to manage a model with abstraction

• Here, this can be sequential and consistent

• Some major models here that calculate function:

  • The model of senses
  • To " think, Plan
  • Finally have actuator

• Issues:

  • The sequence relies on all "modules" so if one fails, the entire model is inconsistent

• This method cannot overcome those, so it fails to evolve properly

Reactive

• Reacting by generating input with a certain set of guidelines

• This is quicker and more responsive than AI architecture

• However, generating responses that don't generate the previous, so it builds its model with:

• sensory Inputs

• Sense to Act

• *Generates Actuators - Lacks the ability to build from it

• Some issues are: It not able to generate anything outside of that frame

Hibrido - Fusion Reactionary (Evolution AI)

• Includes everything necessary to manage the data
• The models it builds are limited in data however as the framework that develops is unable to provide data.

Operation-Based Model

• Inspired by biological systems
• Runs simultaneous, and are in constant communication with each other
• But may require additional computing power during critical situations