Introduction to Intelligent Agent and its types
- 1. Unit 1- Intelligent Agents Dr.V.Nirmala Assistant Professor, Artificial Intelligence and Data Science Easwari Engineering College
- 2. Introduction to Intelligent Agents Welcome to this presentation on Intelligent Agents! We'll explore the fundamental concepts of AI agents, their interaction with environments, and the properties that define their behavior. An agent is anything that can perceive its environment through sensors and act upon that environment through actuators, such as human agents, robotic agents, or software agents. An intelligent agent is a system that perceives and acts in an environment to achieve specific goals. It can make decisions rationally, adapt over time, and operate autonomously.
- 3. Agents and Environments An agent interacts with its environment by perceiving it through sensors and acting upon it via actuators. A percept is a single input received at one time, while the percept sequence is the complete history of all received percepts. The agent's behavior is fundamentally defined by the agent function, which maps every possible percept sequence to an action. This abstract function is implemented by the agent program, which runs on a physical machine or system. For instance, in a vacuum-cleaner world, the agent uses sensors to detect its location and the presence of dirt. Its actions could be to suck up dirt or move to an adjacent square. The challenge in designing intelligent agents lies in creating an agent function that enables optimal actions, often requiring sophisticated analysis tools for evaluation.
- 4. Concept of Rationality and Rational Agent Defining Rationality A rational agent consistently chooses actions that are expected to maximize its performance measure, based on its current percept sequence, prior knowledge, and available actions. Performance Measurement Performance is evaluated by the desirability of the resulting environment states, not just the agent's internal assessment. It's about outcomes, not intentions. Rationality vs. Omniscience Rationality is about acting optimally given available information, not having perfect knowledge or avoiding all errors. An agent can be rational even if unforeseen events lead to poor results. Information & Autonomy Rational agents actively gather information and learn from experience to improve. Autonomy means they adapt beyond initial programming to handle unknown environments. Consider a vacuum-cleaner agent: it's rational if it cleans dirty squares efficiently based on its sensors and environment knowledge. However, if it moves unnecessarily and the performance measure penalizes movement, it might be considered irrational for wasting energy.
- 5. Task Environment and PEAS Framework The task environment is crucial for defining the problem an agent is designed to solve. It encompasses all relevant aspects that influence the agent's design and behavior. The PEAS description (Performance, Environment, Actuators, Sensors) is a structured framework used to fully specify the task environment before designing an intelligent agent. This systematic approach ensures all critical aspects are considered. Performance Measure Defines the criteria for success, such as safety, speed, or comfort for a taxi, or profit maximization for a business agent. Environment The external world where the agent operates, including elements like roads, traffic, weather, or other interacting agents. Actuators The means by which the agent performs actions, such as steering, accelerating, braking, or displaying information. Sensors The tools through which the agent perceives its environment, like cameras, GPS, sonar, or speedometers.
- 6. PEAS Framework in Action: Automated Taxi Driver To illustrate the PEAS framework, let's consider the example of an automated taxi driver, a complex yet relatable intelligent agent. Performance: Safe operation, minimizing accidents. Adherence to traffic laws and regulations. Comfortable rides for passengers. Maximizing profits by optimizing routes and fuel efficiency. Actuators: Steering control for navigation. Accelerator for speed control. Brakes for stopping and slowing down. Horn for warning other drivers or pedestrians. Display screen for communicating with passengers. Environment: Diverse road types (highways, city streets, unpaved roads). Dynamic traffic conditions and other vehicles. Varying weather (rain, snow, fog) and lighting (day, night). Pedestrians, cyclists, and animals. Obstacles, road signs, and traffic signals. Sensors: Cameras for visual perception of surroundings. Sonar and radar for proximity detection. Speedometer for current vehicle speed. GPS for precise location and navigation. Accelerometer for detecting vehicle movement and tilt.
- 7. Properties of Task Environments Understanding the characteristics of a task environment is fundamental to designing effective intelligent agents. These properties dictate the complexity and challenges an agent will face. 1. Fully vs. Partially Observable The observability of an environment refers to how completely an agent's sensors can capture its state. Fully Observable The agent's sensors provide access to the entire state of the environment necessary for decision-making. Example: In a chess game, all pieces and their positions are always visible to both players. Partially Observable Sensors do not provide complete information, either due to noise, occlusion, or hidden aspects. Example: A taxi driver cannot directly perceive the intentions of other drivers or pedestrians. Unobservable The agent receives no information from the environment and must rely on its initial programming or random strategies.
- 8. Properties of Task Environments (Continued) 2. Single-Agent vs. Multi-Agent This property defines the number of decision-making entities interacting within the environment. Single-Agent Only one agent is making decisions, and its actions do not depend on or affect other agents. Example: A lone player solving a crossword puzzle. Multi-Agent Multiple agents exist, and their actions can significantly impact each other, requiring coordination or competition. Competitive: Agents have opposing goals (e.g., chess, poker). Cooperative: Agents share common goals (e.g., multiple robots collaborating on a task). Mixed: Environments with both competitive and cooperative aspects (e.g., taxi driving, where drivers compete for fares but cooperate to follow traffic laws).
- 9. Properties of Task Environments (Continued) 3. Deterministic vs. Stochastic This characteristic refers to the predictability of action outcomes within the environment. Deterministic The outcome of each action is entirely predictable based on the current state. Example: In a simple logic-based board game, a move always results in the same new board configuration. Stochastic Outcomes involve an element of randomness or uncertainty, meaning an action may not always produce the same result. Example: A taxi driver might face unpredictable traffic jams, sudden weather changes, or unexpected pedestrian behavior. 4. Episodic vs. Sequential This property differentiates environments based on the dependency of decisions over time. Episodic Each decision or "episode" is independent of previous actions; the agent's past actions do not affect the current state. Example: An automated inspection system checking parts for defects, where each part is a new, unrelated instance. Sequential Decisions are interdependent; earlier actions influence later ones and the overall outcome. Example: In chess or driving, current choices significantly impact future possibilities and results.
- 10. Properties of Task Environments (Continued) 5. Static vs. Dynamic This property distinguishes environments based on whether they change while the agent is deliberating. Static The environment remains unchanged while the agent is making a decision. Example: Solving a crossword puzzle, where the clues and grid do not change once presented. Dynamic The environment can change autonomously while the agent is thinking or acting. Example: In driving, traffic patterns and pedestrian movements evolve continuously, even as the agent decides to brake or accelerate. 6. Discrete vs. Continuous This property describes the nature of states, actions, and time within the environment. Discrete The environment has a finite set of distinct states, actions, or time steps. Example: Chess has discrete board positions and a defined set of legal moves. Continuous States, actions, or time change smoothly and can take on any value within a range. Example: In taxi driving, the vehicle's position, speed, and the flow of time are all continuous variables. 7. Known vs. Unknown This property relates to the agent's initial knowledge of the environment's rules and effects. Known The agent has complete knowledge of the rules governing the environment and the consequences of its actions. Example: Playing a game with clearly defined rules, like checkers. Unknown The agent must explore and learn how its actions affect the environment to build an internal model. Example: A new video game with unfamiliar controls and mechanics that the player must discover.
- 11. Structure of Agents An intelligent agent is a combination of its physical or virtual architecture and the agent program that dictates its behavior. The architecture provides the sensory input and carries out the physical actions, while the program processes the input to determine the appropriate output. Sensors Mechanisms for perceiving the environment, gathering data. Agent Function The abstract mathematical mapping of percept sequences to actions. Agent Program The concrete implementation of the agent function, runnable on a machine. Actuators Components for taking actions and influencing the environment. In essence, the sensors feed information to the agent program, which then decides on an action, executed by the actuators. This continuous loop allows intelligent agents to interact dynamically with their surroundings.
- 12. Four Basic Types of Agent Programs Intelligent agents can be categorized into four fundamental types based on their decision-making mechanisms and internal complexity. Each type represents a different approach to processing environmental information and selecting actions. Simple Reflex Agents These agents choose actions based solely on the current percept, without considering historical information or future consequences. They follow condition-action rules directly. Model-Based Reflex Agents These agents maintain an internal state to track aspects of the world that are not evident in the current percept. They use this internal model to make more informed decisions. Goal-Based Agents These agents act to achieve specific goals, considering future actions and their potential outcomes. They can plan and reason about sequences of actions to reach desired states. Utility-Based Agents These agents choose actions to maximize a utility function that measures the desirability of different outcomes. They can handle conflicting goals and make trade-offs between competing objectives.
- 13. Simple Reflex Agents : ï Simple reflex agents select actions based solely on the current percept, ignoring the percept history. They operate using condition-action rules (if- then rules) that directly map perceived conditions to actions. ï Example: The classic vacuum-cleaner agent that cleans if the current square is dirty and moves otherwise is a simple reflex agent. Another example is a driver braking immediately when the brake lights of the car in front turn on. ï Working Mechanism: ï The agent perceives the current state via sensors. ï The percept is interpreted into an abstract state description. ï The agent matches this state against a set of condition-action rules. ï The first matching rule determines the action to execute. ï This process is simple and fast, often implementable by Boolean logic circuits.
- 14. Simple Reflex Agents : ï Advantages: ï Very simple and efficient to implement. ï Works well in fully observable environments where the current percept contains all necessary information. ï Limitations: ï Cannot handle partially observable environments because it ignores history or internal state. ï May get stuck in infinite loops or make poor decisions if the percept is ambiguous or incomplete. ï Lacks learning and adaptation capabilities. ï Example of Problem: A vacuum agent with only a dirt sensor but no location sensor might endlessly move left or right without cleaning effectively.
- 15. Model Based Reflex Agents ï Model-based reflex agents maintain an internal state that represents unobserved aspects of the environment, updated based on the percept history and a model of how the world evolves and how the agent9s actions affect it. ï Internal State: Keeps track of information not directly observable in the current percept. For example, a driving agent remembers positions of other cars not currently visible or the agent9s destination. ï World Model: Encodes knowledge about: ï How the environment changes independently of the agent (e.g., a car behind getting closer). ï How the agent9s actions change the environment (e.g., turning the steering wheel turns the car). ï Decision Process: ï Update State: Combine previous internal state, recent action, and current percept to update the internal model. ï Rule Matching: Use condition-action rules on the updated state to select the next action.
- 16. Model Based Reflex Agents ï Advantages: ï Can operate in partially observable environments by maintaining and updating internal state. ï More flexible and intelligent than simple reflex agents. ï Can handle uncertainty by making best guesses about the current state. ï Limitations: ï Building and maintaining accurate models can be computationally expensive. ï Models may not capture all real-world complexities and require frequent updates. ï Still limited by the predefined set of condition-action rules. ï Example: An automated taxi uses sensors to perceive the environment, updates its internal model of traffic and road conditions, and decides actions like braking or lane changes based on this model.
- 17. Goal-based agents ï Goal-based agents select actions based on both the current state and a goal that describes desirable situations (e.g., reaching a destination). They combine a model of the world with goal information to choose actions that achieve the goal. ï Key Features: ï Use search and planning to find sequences of actions leading to the goal. ï Consider future consequences of actions (<What will happen if I do this?=). ï More flexible than reflex agents because goals and models are explicit and modifiable. ï Can adapt behavior easily by changing the goal without rewriting rules.
- 18. Goal-based agents ï Example: A taxi deciding whether to turn left, right, or go straight based on the destination goal rather than just reacting to immediate percepts. ï Applications: Robotics, autonomous vehicles, game AI, and any domain requiring planning and goal achievement. ï Advantages: ï Clear direction for decision-making. ï Adaptable to different goals and environments. ï Can handle complex tasks requiring multi-step planning. ï Limitations: ï Requires knowledge of goals and environment model. ï Planning can be computationally expensive.
- 19. Utility-based agents ï Utility-based agents extend goal-based agents by using a utility function that assigns a numerical value to each possible state, representing the agent9s preferences. They choose actions to maximize expected utility, balancing trade-offs among conflicting goals. ï Key Features: ï Handle conflicting goals (e.g., speed vs. safety) by weighing utilities. ï Make decisions under uncertainty and partial observability by maximizing expected utility. ï Internalize the performance measure as a utility function guiding rational behavior. ï Expected Utility: Agents compute the expected utility of possible outcomes, averaging over uncertainties weighted by their probabilities, and select the action with the highest expected utility. ï Example: A taxi choosing a route that balances shortest time, safety, and fuel efficiency rather than simply reaching the destination. ï Advantages: ï More nuanced decision-making than binary goal satisfaction. ï Can adapt to stochastic and uncertain environments. ï Supports principled trade-offs among multiple objectives. ï Limitations: ï Requires defining a suitable utility function. ï Computationally complex to calculate expected utilities and optimal actions.