Unit 3: Intelligent Agents - Subjective Questions

An intelligent agent is an autonomous entity that perceives its environment through sensors and acts upon that environment through actuators to achieve specified objectives.

• Environment: The external world in which the agent operates.
• Sensors: Components used to obtain percepts from the environment. Examples include cameras, microphones, and temperature sensors.
• Agent program: Processes percepts and determines an appropriate action.
• Actuators: Components used to execute actions. Examples include motors, robotic arms, displays, and speakers.

The interaction can be represented as:

Environment → Sensors → Agent → Actuators → Environment

For example, a robotic vacuum uses infrared or distance sensors to detect obstacles, processes the sensor data, and controls its wheels and cleaning mechanism through actuators.

Agent architecture is the physical or computational platform on which an agent program operates. The agent program is the software that maps percepts or percept sequences to actions.

The relationship is commonly expressed as:

Agent = Architecture + Agent Program

• The architecture provides sensors, processing resources, memory, communication facilities, and actuators.
• The agent program implements the decision-making logic.
• The architecture must be compatible with the requirements of the program.

For example, in an autonomous car:

• Cameras, radar, processors, steering, and braking systems form the architecture.
• Software for object detection, route planning, and vehicle control forms the agent program.

A single agent program may behave differently when deployed on architectures with different sensors, processing power, or actuator capabilities.

The PEAS framework specifies the task environment of an intelligent agent. PEAS stands for Performance Measure, Environment, Actuators, and Sensors.

For an autonomous taxi:

• Performance Measure: Passenger safety, legality, travel time, fuel efficiency, passenger comfort, successful arrival, and operating cost.
• Environment: Roads, traffic signals, pedestrians, passengers, other vehicles, weather conditions, road signs, and maps.
• Actuators: Steering, accelerator, brakes, gear controls, horn, indicators, doors, display, and speech system.
• Sensors: Cameras, radar, LiDAR, GPS, speedometer, odometer, microphones, engine sensors, and proximity sensors.

A PEAS description clearly identifies what counts as success, where the agent operates, how it can act, and what information it can observe. It is therefore useful when designing an agent and evaluating its behavior.

The four PEAS components are:

• Performance Measure: The criteria used to evaluate the success of the agent's behavior.
• Environment: Everything external to the agent that may affect its operation.
• Actuators: Mechanisms through which the agent changes the environment.
• Sensors: Mechanisms through which the agent receives information from the environment.

The performance measure should be defined by the designer because an agent might otherwise optimize an unintended internal measure. For example, a vacuum agent rewarded only for collecting dirt might repeatedly spread and recollect dirt. A well-designed measure should instead evaluate outcomes such as cleanliness, energy consumption, completion time, and avoidance of damage.

Thus, the performance measure must describe the desired external results, not merely the actions or internal states of the agent.

A rational agent selects the action expected to maximize its performance measure, given its percept sequence, prior knowledge, and available actions.

A rational agent is not necessarily:

• Omniscient: It does not know the actual outcome of every action in advance.
• Always successful: Uncertainty, incomplete observations, actuator failures, or unexpected environmental events may produce poor outcomes.
• Perfect: It is limited by the information and computational resources available to it.

Rationality depends on four factors:

• The specified performance measure.
• The agent's prior knowledge of the environment.
• The actions available to the agent.
• The complete percept sequence received so far.

For example, an autonomous vehicle may rationally brake after detecting an obstacle, but an unavoidable collision may still occur because of insufficient stopping distance. The decision can remain rational even when its result is unsuccessful.

Agent environments can be classified along the following dimensions:

• Fully observable vs. partially observable: A chess agent can observe the complete board, whereas a medical diagnosis agent cannot directly observe every condition inside a patient.
• Single-agent vs. multi-agent: Crossword solving is generally single-agent, while chess involves competing agents.
• Deterministic vs. stochastic: A puzzle may be deterministic, while driving is stochastic because outcomes are uncertain.
• Episodic vs. sequential: Image classification is episodic; driving is sequential because current actions affect future states.
• Static vs. dynamic: A crossword remains unchanged while the agent thinks, whereas road traffic changes continuously.
• Discrete vs. continuous: Chess has discrete states and actions; robotic motion involves continuous position, velocity, and time.
• Known vs. unknown: In a known environment, transition and outcome rules are available; in an unknown environment, the agent must learn them.

These dimensions influence the type of agent architecture and decision-making method required.

A simple reflex agent selects an action using only the current percept. It follows condition-action rules of the form:

If condition, then action.

For example, a vacuum agent may use the rule: If the current square is dirty, then suck.

Its working involves:

• Receiving the current percept through sensors.
• Matching the percept against stored condition-action rules.
• Selecting the corresponding action.
• Executing the action through actuators.

Advantages:

• Simple to design and implement.
• Fast decision-making.
• Suitable for fully observable and predictable environments.

Limitations:

• Ignores percept history.
• Cannot reason about unobserved parts of the environment.
• May repeatedly enter loops.
• Performs poorly in partially observable or dynamic environments.
• Cannot easily adapt when its rules become unsuitable.

A model-based reflex agent maintains an internal state representing aspects of the environment that are not directly visible in the current percept. It updates this state using:

• The previous internal state.
• The latest percept.
• Knowledge of how the environment changes independently.
• Knowledge of how the agent's actions affect the environment.

The internal model allows the agent to infer what may be happening in partially observable environments. Condition-action rules are then applied to the updated state rather than only to the current percept.

For example, a robotic vacuum may remember which rooms have already been cleaned even when they are outside its current field of view. It can update its location after moving and avoid cleaning the same area unnecessarily.

Unlike a simple reflex agent, a model-based agent uses percept history indirectly through its internal state. However, it still selects actions through rules and does not necessarily reason explicitly about future goals.

For example, a simple thermostat reacts only to the current temperature. A model-based climate controller may also consider whether heating is already active, how quickly the room changes temperature, and whether a window is open.

Both may use reflex rules, but the model-based agent makes decisions using a richer representation of the current situation.

A goal-based agent selects actions by considering desired future states called goals. It evaluates possible action sequences and chooses one that can transform the current state into a goal state.

Its main components include:

• A representation of the current state.
• A model of possible actions and their results.
• One or more goals.
• A search or planning mechanism.

Search explores alternative states or action sequences. Planning organizes actions into a sequence that is expected to achieve the goal.

A reflex agent maps a situation directly to an action, while a goal-based agent considers the future consequences of alternative actions. For example, a route-planning agent does not merely turn whenever it sees a road; it examines possible routes and selects actions that lead to the destination.

Goal-based behavior is more flexible because goals can be changed without rewriting every condition-action rule. Its disadvantage is greater computational cost.

A utility-based agent uses a utility function to assign a numerical value to a state or outcome. This value represents how desirable that outcome is to the agent.

A goal only distinguishes between acceptable and unacceptable outcomes, while utility provides a ranking among alternatives. It is useful when:

• Several goal states can be reached with different costs or benefits.
• Goals conflict with one another.
• Outcomes are uncertain.
• Trade-offs are required, such as speed versus safety.

Under uncertainty, the agent can choose the action with the highest expected utility:

where is the probability of reaching state after action , and is the utility of that state.

For example, an autonomous taxi may consider safety, travel time, passenger comfort, and fuel consumption. It chooses the action offering the best overall expected outcome rather than merely any action that reaches the destination.

A goal-based agent evaluates whether a state satisfies a goal, whereas a utility-based agent evaluates how desirable each state is.

A utility-based approach is preferable when multiple solutions exist, resources are limited, outcomes are uncertain, or objectives conflict. For instance, a medical decision agent may compare treatments using expected recovery, side effects, cost, and patient preferences rather than merely asking whether each treatment can address the disease.

A learning agent improves its performance through experience. Its four major components are:

• Performance element: Selects external actions based on current knowledge and percepts. It is responsible for the agent's immediate behavior.
• Learning element: Modifies or improves the performance element using feedback and experience.
• Critic: Evaluates the agent's behavior according to the performance measure and provides feedback to the learning element.
• Problem generator: Suggests exploratory actions that may produce useful new experiences, even if they are not currently considered optimal.

For example, in a game-playing agent, the performance element chooses moves, the critic evaluates results, the learning element updates the strategy, and the problem generator encourages testing unfamiliar moves.

Learning allows the agent to operate in initially unknown environments and adapt to changes. A balance between exploration of new actions and exploitation of known successful actions is essential.

These types represent increasing decision-making sophistication. An advanced autonomous system may combine all of them. For example, an autonomous vehicle may use reflex braking, a model of nearby traffic, goal-based route planning, utility-based maneuver selection, and learning-based perception.

An autonomous intelligent system can perceive, decide, and act with limited or no continuous human control. It uses its own observations and experience to perform tasks and respond to environmental changes.

Major characteristics include:

• Perception: Obtaining and interpreting environmental information.
• Decision-making: Selecting actions according to goals or utility.
• Adaptability: Adjusting to new situations or learning from experience.
• Self-governance: Operating without step-by-step human instructions.
• Robustness: Continuing to function under uncertainty or partial failure.

Major challenges include:

• Incomplete or noisy sensor data.
• Unpredictable environments.
• Safety and reliability requirements.
• Ethical and legal accountability.
• Cybersecurity threats.
• Explainability of decisions.
• Maintaining meaningful human oversight.

Examples include autonomous vehicles, delivery robots, spacecraft, industrial robots, and intelligent software agents.

A multi-agent system consists of multiple autonomous agents that interact within a shared environment. Each agent has its own observations, actions, knowledge, and objectives, although agents may also share a common objective.

Main features include:

• Distributed perception and decision-making.
• Communication or indirect interaction among agents.
• Local knowledge and limited control.
• Coordination of actions and resources.
• Emergent system-level behavior.

Types of interaction include:

• Cooperative: Agents work toward a shared objective, such as warehouse robots moving goods efficiently.
• Competitive: One agent's success may reduce another's success, as in chess or automated auctions.
• Mixed: Agents have both shared and conflicting interests, as in road traffic where drivers share safety rules but prefer different routes and arrival times.

Multi-agent systems can be scalable and robust, but they also introduce challenges involving communication, conflict resolution, trust, and coordination.

Cooperation occurs when agents assist one another in achieving common or compatible objectives. Coordination is the organization of their actions so that dependencies are managed, resources are used effectively, and harmful conflicts are avoided.

Important techniques include:

• Communication: Exchanging goals, observations, intentions, or status information.
• Task allocation: Assigning tasks according to agent capabilities, location, or workload.
• Negotiation: Reaching agreements when agents have different preferences.
• Resource scheduling: Preventing simultaneous incompatible use of shared resources.
• Protocols and social rules: Defining acceptable interaction patterns.
• Centralized coordination: Using a coordinator to assign actions.
• Distributed coordination: Allowing agents to reach decisions using local interactions.

For example, warehouse robots can reserve paths, communicate destinations, and negotiate priority at intersections. Effective coordination improves throughput and safety, while poor coordination may cause collisions, deadlocks, duplicated work, or wasted resources.

Human-AI collaboration involves humans and intelligent agents working together by combining complementary strengths. AI systems offer speed, consistency, large-scale data processing, and pattern recognition, while humans contribute contextual understanding, creativity, ethical judgment, and responsibility.

Effective collaboration should follow these principles:

• Clear role allocation: Specify which decisions belong to the human and which can be automated.
• Transparency: Communicate the agent's recommendations, confidence, and limitations.
• Human control: Provide mechanisms to review, correct, override, or stop the agent.
• Appropriate trust: Avoid both blind reliance and unnecessary rejection of AI outputs.
• Usability: Present relevant information without overwhelming the user.
• Feedback: Allow human expertise to improve future agent behavior.
• Accountability: Establish responsibility for decisions and outcomes.

For example, a medical AI may identify suspicious images and explain its confidence, while a qualified clinician considers patient history and makes the final diagnosis.

Applications of intelligent agents include:

• Healthcare: Perceives symptoms, medical records, and test results; recommends diagnoses, treatments, or alerts.
• Transportation: Perceives roads, vehicles, pedestrians, and traffic signals; steers, brakes, accelerates, or plans routes.
• Finance: Perceives market data, transactions, and customer behavior; trades assets, detects fraud, or assesses credit risk.
• Manufacturing: Perceives machine status, product quality, and production flow; controls robots, schedules tasks, or requests maintenance.
• Education: Perceives learner answers, progress, and interaction patterns; adapts content, gives feedback, or recommends exercises.
• Customer service: Perceives user messages and account context; answers questions, retrieves information, or escalates cases.
• Cybersecurity: Perceives network traffic, system logs, and user behavior; blocks suspicious activity or generates alerts.

In each case, the agent converts percepts into actions intended to maximize a domain-specific performance measure.

A smart hospital delivery robot can be designed as follows.

PEAS description:

• Performance Measure: Correct and timely delivery, patient and staff safety, low energy consumption, hygiene compliance, and minimal disruption.
• Environment: Corridors, rooms, elevators, doors, patients, staff, visitors, medical equipment, and other robots.
• Actuators: Wheels, brakes, robotic compartment locks, display, speaker, lights, and wireless communication module.
• Sensors: Cameras, LiDAR, proximity sensors, RFID reader, GPS or indoor localization, microphones, wheel encoders, and battery sensors.

Environment properties:

• Partially observable because sensors cannot reveal every hidden obstacle.
• Stochastic because people move unpredictably.
• Sequential because each movement affects later decisions.
• Dynamic because the hospital changes while the robot plans.
• Continuous in position and time.
• Multi-agent because it interacts with humans and other robots.

Suitable architecture:

A hybrid model-based, goal-based, utility-based, and learning architecture is appropriate. It must maintain a map, plan deliveries, balance safety against speed, and learn traffic patterns.

Coordination requirements:

Robots should share routes, reserve elevators, negotiate corridor priority, report obstacles, and avoid deadlocks. Human staff must be able to override or redirect them during emergencies.