Artificial Intelligence: Types & Capabilities

• Course explores into Artificial Intelligence & Expert Systems (ITE 153) during the 2nd Semester of AY 2024-2025, taught by Dr. Lumer Jude Doce from the IT & Physics Department

Types of Artificial Intelligence

• Artificial Intelligence is categorized based on capabilities and functionality.

AI Based on Capabilities (Type-1)

Weak AI or Narrow AI

• Narrow AI performs dedicated tasks with intelligence and is the most common AI.
• It cannot operate beyond its limitations and is trained for specific tasks
• Also termed weak AI, failure is possible if it exceeds its limits.
• Apple Siri exemplifies Narrow AI with a limited pre-defined range of functions.
• IBM's Watson supercomputer falls under Narrow AI with expert systems, machine learning, and natural language processing.
• Narrow AI is used in playing chess, e-commerce suggestions, self-driving cars, and image recognition.

General AI

• General AI can perform any intellectual task with human-level efficiency, being able to think like a human by itself
• The idea is to create systems that are smarter than humans.
• Currently, there is no system that can perform tasks as perfectly as a human.
• Researchers are focused on developing General AI, however, it requires time and effort to develop these systems.

Super AI

• Super AI is a level of intelligence where machines surpass human intelligence including cognitive properties and it is an outcome of General AI.
• Super AI's key characteristics are the capability to think, reason, solve puzzles, make judgments, plan, learn, and communicate.
• Still a hypothetical concept with development a world-changing task.

AI Based on Functionality (Type-2)

Reactive Machines

• Reactive machines are basic AI types.
• These AI systems do not store memories or past experiences for future actions, they focus on current scenarios and react accordingly.
• IBM's Deep Blue and Google's AlphaGo are examples of reactive machines.

Limited Memory

• Limited memory machines store past experiences or short-term data.
• They can only use stored data for a limited time
• Self-driving cars exemplify limited memory systems, they store recent speed, distance of nearby cars, the speed limit and other info.

Theory of Mind

• Theory of Mind AI should understand human emotions, beliefs, and interact socially.
• This type of AI is still under development.

Self-awareness

• Self-awareness AI is the future of Artificial Intelligence.
• These machines will be highly intelligent with their own consciousness, sentiments, and self-awareness while being smarter than human minds.
• Self-Awareness AI is a hypothetical concept.

Agents in AI

• An AI system studies rational agents and their environments.
• Agents perceive their environment through sensors and act through actuators.
• AI agents can possess mental properties like knowledge, belief, and intention.
• An agent runs in a cycle of perceiving, thinking, and action
• Human-Agents have eyes, ears, and limbs.
• Robotic Agents use cameras, infrared range finders, NLP, and motors.
• Software Agents use keystrokes and file contents as sensory input and display output.
• The world has many agents, these include thermostats, cellphones, cameras, and even humans.
• Sensors detect environment changes by sending data to electronic devices allowing Agents to observe the environment.
• Actuators convert energy into motion, for movement and control.
• Effectors affect the environment, such as legs, wheels, arms, fingers, etc.

Intelligent Agents

• Intelligent agent is an autonomous entity, it acts upon an environment using sensors and actuators to achieve goals and it learns from the environment.
• Thermostats are examples of intelligent agents.
• The four rules for an AI agent, are the ability to perceive, its observation must be used to make decisions, decisions must result in actions, and the actions must be rational.

Rational Agents

• Rational agents have clear preferences, model uncertainty, and maximize performance.
• They perform the right things, creating tools for game and decision theory.
• Rational actions are necessary in reinforcement learning, where correct actions yield rewards and incorrect actions yield negative rewards.
• Rational agents in AI are similar to intelligent agents.
• The rationality is measured by performance, prior knowledge, actions, and percept sequences.

Structure of AI Agents

• Task of AI is to design an agent program that implements the agent function
• An intelligent program is a combination of architecture and a program. Agent = Architecture + Agent program
• Architecture is the machinery that AI executes on.
• Agent Functions map a percept to an action, and they are implemented by an agent program. An program on physical architecture creates function f: (f:P* → A)

PEAS Representation

• PEAS models define AI agent properties, and is made up of four words.
• P: Performance measure, the objective for success.
• E: Environment.
• A: Actuators.
• S: Sensors.
• For self-driving cars, its Performance is its safety, time, legal compliancy and comfort, it operates on roads with other vehicles etc, with actuators being steering, accelerator, brake, horn et and it uses camera, GPS, speedometer and other sensors.

Types of AI Agents

• Agents are in five classes based on degrees of perceived intelligence and all can improve over time.
• Simple Reflex Agent
• Model-based reflex agent
• Goal-based agents
• Utility-based agent
• Learning agent

Simple Reflex Agent

• Simple reflex agents take actions based on current percepts and ignore the history and are only effective in a fully observable environment.
• The agent works on condition-action rules, mapping states to actions.
• A Room Cleaner cleans if dirt is in the room
• However, simple design has limited intelligence, no knowledge of unobservable states, is large to store, and they are not adaptive.

Model-Based Reflex Agent

• Model-based agents function in partially observable environments by tracking situations with a model with factors on how the world operates and an internal state that represents current conditions based on past percepts.
• These agents utilize the model, which is world knowledge, to perform actions.
• Agents need to state updates and information on the world's evolution and the agent's effects on the world,

Goal-Based Agent

• Knowing the current environment isn't sufficient, therefore agents need to know its goal.
• Goal-based agents have capabilities of model-based agents while possessing goal information.
• They choose actions to achieve the goal.
• Different scenarios are considered when to decide if the end goal is achieved or not and can cause proactive action.

Utility-Based Agent

• Utility agents have measurements of success at a given state, also acting on how to achieve it.
• Utility-based agents are useful for choosing the best action among multiple options.
• Utility functions map states to real numbers to assess action efficiency.

Learning Agent

• Learning agents can use past experiences or data to automatically adapt
• They start with basic knowledge and evolve through learning automatically.
• Key components are a learning element for improvements, the critic to provide feedback, a performance element, and a problem generator for new experimental actions.
• Learning agents can learn, analyze, and seek new improvements.

Agents Environment

• The environment is the surroundings excluding the agent itself with it being described by the agent's present situation.
• It provides the agent to sense and act upon with it mostly being said to be non-feministic.
• Based on Russell and Norvig, an environment can be fully or partially observable, static or dynamic, discrete or continuous, deterministic or stochastic, single-agent or multi-agent, episodic or sequential, known or unknown, and accessible or inaccessible.

Fully Observable vs Partially Observable

• A fully observable environment is when agents can access the environment's complete state at all times, which is easy as there is no need to keep track of past history.
• A partially observable environment are environments that can not be completely sensed by all agents.
• An unobservable environment are agent with no sensors.

Deterministic vs Stochastic

• The environment is deterministic if an agent's current state and action determine the next state.
• A stochastic environment is random and unpredictable by the agent.
• Agents don't need to worry of uncertantiy in a determined fully observable environment.

Episodic vs Sequential

• Episodic environments involve one-shot actions with the current percept.
• Sequential environments require a memory of past moves to determine the next action..

Single-Agent vs Multi-Agent

• Single Agents is when only one agent is involved.
• Multi-agents is when multiple are involved.
• Problems are different in multi-agent environments.

Static vs Dynamic

• A static environment remains same while agents deliberate.
• A dynamic environment can change even as the agent is acting.
• Taxi driving exemplifies a dynamic environment, while a crossword puzzle is static.

Discrete vs Continuous

• Descrete is when there is a finite number of percepts and actions.
• Continuous is whent here is no finite number of percepts and actions.
• Chess is a descrete environment, while self-driving would be continous.

Known vs Unknown

• Known/Unknown describe an agent's knowledge to act in an environment, not the environment itself.
• In a known environment, the results of the agent's acttions are known, while in an unknown, it needs to learn.
• It's possible to that a known environment to be partially observable and an Unknown environment to be fully observable.

Accessible vs Inaccessible

• Accessible environments happen when an agent can get complete and accurate information.
• Inaccessible environments involve something such as the state of the environment cannot be obtained.
• An empty room with its termperature and state easily defined is an example of an accesbile environment.
• Information on an event happening on earh is inaccessible.