Intelligent Agents in AI

Intelligent Agents

• Agents are entities that perceive their environment through sensors and act upon it through actuators.
• Examples include humans, robots, and thermostats.
• The environment is the part of the universe whose state influences the agent's actions and perception.

Simple Terms

• Percept: An agent's sensory input at any given moment.
• Percept sequence: A complete history of everything perceived by the agent. Actions depend on built-in knowledge and the entire percept sequence, but not on unperceived information.

Agent Function & Program

• Agent function: A mathematical description of an agent's behavior, mapping any percept sequence to an action.
• Agent program: The actual implementation of the agent function, running within a physical system.

Example: Vacuum-cleaner world

• Perception: Location (A or B) and cleanliness (clean or dirty).
• Actions: Move left, move right, suck up dirt, do nothing.
• Example agent function: If a square is dirty, suck; otherwise, move to the other square.

Good Behavior: Rationality

• Rational agent: An agent whose actions maximize its success according to a performance measure.
• Correct action: The action maximizing the agent's success is considered correct (rational).
• Performance measure: Evaluates any given sequence of environmental states, establishing how successful an agent's actions are. It must define desirable actions.

Performance Measure

• The performance measure determines how successful an agent is.
• Examples: Percentage of correct actions, minimizing time or cost, etc.
• Performance measures should reflect actual desires; e.g., cleaning a floor, regardless of chosen actions.

Rationality

• Rationality depends on the performance measure, the agent's environmental knowledge, possible actions, and the percept sequence.
• A rational agent selects the action expected to maximize its performance, given the evidence from the percept sequence and built-in knowledge.

Omniscience

• An omniscient agent knows the actual outcome of its actions in advance.
• However, omniscience is unrealistic; a truly rational agent does not need to be capable of foreseeing all possible future outcomes.

Learning

• A rational agent is not limited to its current percept alone but should also consider past percept sequences (learning behavior).
• It learns by adapting its actions based on experience to improve its performance in similar situations next time.

Autonomy

• Autonomous agents rely primarily on their own perceptions and experience, rather than solely on pre-programmed knowledge.
• Rational agents should learn to compensate for partial or incorrect initial knowledge. This learning allows the agent's behavior to become independent of pre-programmed parameters.

The Nature of Environments

• Environments can be artificial (e.g., video games, flight simulators) or real-world simulations.
• Software agents (softbots) operate within these artificial, yet complex environments.

Task Environments

• Task environments are the problems rational agents solve.
• PEAS description specifies the task environment adequately:
• Performance measure
• Environment
• Actuators
• Sensors

Properties of Task Environments

• Fully observable vs. partially observable environments
• Deterministic vs. Stochastic environments
• Episodic vs. sequential environments
• Static vs. dynamic environments
• Discrete vs. continuous environments
• Single agent vs. multiagent environments
• Known vs. unknown environments (unobservable aspects or unpredicted events introduce uncertainty in the environment)

The Structure of Agents

• Agents combine architecture and program.
• Architecture includes sensors and actuators (motors).
• The agent program implements the agent function mathematically.

Agent Programs

• Input for an agent program (acting on the percept): only the current percept.
• Input for an agent function: The entire percept sequence.
• Implementation as a look-up table (agent function): Agent actions are pre-programmed for all possible scenarios, storing a mapping of possible states.
  • Agent program example—a table-driven agent program (storing a table of percept-action pairs).

Types of Agent programs

• Simple reflex agents
• Model-based reflex agents
• Goal-based agents
• Utility-based agents

Simple Reflex Agents

• These agents use condition-action rules (“if... then...”) only based on the current state of a perceived environment.
• Efficient but with limited applicability.
• Suitable only for fully observable environments, because hidden aspects of the environment are not accounted for.

Model-Based Reflex Agents

• These agents use internal states to predict future states of their environments.
• They model how the world evolves independently, and how their actions affect the environment.
• They are suitable for partially observable environments.
  • Requires a model of how the state evolves based on the current state and taken actions.

Goal-Based Agents

• These agents make decisions based on goals (desired outcomes) rather than simple condition-action rules.
• It uses a goal or set of goals and estimates how likely it is to reach a goal based on different sets of possible actions.
• Agents use search (planning subfields) to find appropriate action sequences.

Utility-Based Agents

• Utility-based agents rank states and actions based on their usefulness (utility), not just whether they achieve a goal.
• These agents rank different potential sequences of states based on utility.
• These agents provide a degree of success (utility) for how successful the agent is, not just a binary or yes/no answer for goal achievement.
• Suitable where multiple goals conflict, and trade-offs must be made.

Learning Agents

• Learning agents adapt to new information. They improve their performance over time and experience.
• These agents employ a feedback mechanism from the performance in an environment to improve the learning element
• Crucial components include a learning element, performance element, critic, and problem generator.