Intelligent Agents and Percept-Action Sequences

Intelligent Agents

• Agents and their environments are studied to handle interactions between AI-based software and its environment.
• An agent is anything that can perceive its environment through sensors and act upon it through actuators.
• Robotic agents have sensors like RGB/infrared cameras and actuators like motors or arms.
• Software agents have sensory inputs such as file contents, network packets, and human input (keyboard/mouse) and outputs like displays, sound.
• An agent's action choice depends on its built-in knowledge, where input leads to an agent function that calculates the output.
• Example: A vacuum cleaner agent in a world has sensors to detect locations (A or B) and statuses (dirty or clean), chooses actions like going right, left, or do nothing, and starts in location A.

Percept-Action Sequences

• Agents can use tables of possible percept-action sequences for all possible decisions. This is critical for decision-making.
• The agent function returns an action based on the current location and status, such as sucking if something is dirty.

Rational Agents

• Performance measures evaluate the environment states.
• These measures represent desirable actions (human perspective) or designer goals in software agents.
• Designing a measure according to desired behavior, not how the agent thinks, should be achieved.
• Rational agents choose actions that maximize expected utility, taking into account both percepts and built-in knowledge.

Simple Reflex Agents

• These agents make decisions based only on the current percept, ignoring past history.
• Example: A reflex-vacuum-agent acts only if its location is dirty; otherwise, it does nothing.

Model Based Reflex Agents

• They handle partial observability by keeping track of parts of the world not directly visible.
• They use an internal state to record what they can't currently see, improving decision-making.

Goal-Based Agents

• These agents have goals to guide their actions.
• They choose actions that lead to achieving those goals while keeping track of the current status of the world and desires.

Utility-Based Agents

• They handle conflicting goals by assigning a utility function to each outcome. This function maximizes the expected utility of options.
• In scenarios with partial observability, this helps in decision-making.

Learning Agents

• They learn from experiences to improve performance. Using the external performance standard, it evaluates how the agent does.

Knowledge-Based Agents

• These agents have a knowledge base with previous knowledge.
• Based on this knowledge and the current state of the world, the agent can determine the next appropriate action to take.

Procedural Approach

• Desired behaviors are directly encoded by the agent designer.
• This is a direct approach to the design of an agent which is done through programming (if-else statements)

Declarative Approach

• The agent designer encodes specifications of what is needed to be done, while a reasoning algorithm manages the output in detail. (for instance, through reasoning rules)

Ontologies and Knowledge Graphs

• Ontology is to the design of theories of being (existence)
• Taxonomy is a categorization that shows relationships between existing things, such as child-parent relationships.
• T-Box contains concepts and relations used to explain real-world things
• A-Box shows the facts, statements, or individuals based on T-Box's concepts and relations.

Monotonic Inference

• Classical logic (where adding new information in a set of sentences does not invalidate previously derived conclusions).

Non-Monotonic Reasoning

• It is when adding new information can revise previous conclusions because knowledge is incomplete.
• Some example scenarios: doctors making initial diagnoses.

Common Sense Reasoning

• It's about creating plausible conclusions through assumptions of the world being normal and as expected.

Answer Set Programming

• A paradigm of logic programming used to solve complex problems.

User Data Modeling

• Discusses the Dunning-Kruger effect, design methods (engineering, user-centered, participatory), and technological determinism.

User Profile and User Profiling

• Summarizes user interest and characteristics for better experience.

Supervised Machine Learning

• Methods to find the relationship between the input and output from data.

Loss Functions

• Metrics to measure the closeness between predicted values and actual ones. Used to calculate error and optimization of model parameters.

Regression

• Learning a model of predicting numerical output from the input.

Linear Regression

• A model where the output variable is an affine combination of input variables.

Classification

• Learning a model to categorize the input into different classes.

Logistic Regression

• A model that uses a sigmoid function for separating classes.

Support Vector Machine

• Finding the best border (hyperplane) to distinguish classes with the greatest margin.
• Used for separating data points into different classes.

k-Nearest Neighbors

• A classification method to predict the category of a new data point by finding the k nearest neighbors in the training set.

Unsupervised Machine Learning

• Methods to discover hidden patterns or structures in data without labeled outputs.

Clustering, k-Means, and Gaussian Mixture Models

• Statistical methods to categorize data into groups and find their distribution.