ARTIFICIAL INTELLIGENCE (BCS515B)

Module1
Topics: Introduction: What Is AI? The State of the Art. Intelligent Agents: Agents
and environment, Concept of Rationality, The nature of environment, The
structure of agents.

1.1 What is AI?

Artificial Intelligence (AI) is a field of computer science dedicated to develop
systems capable of performing tasks that would typically require human
intelligence. These tasks include learning, reasoning, problem-solving,
perception, understanding natural language, and even interacting with the
environment. AI aims to create machines or software that can mimic or simulate
human cognitive functions.
According to Russell and Norvig, AI can be defined as follows:
" AI (Artificial Intelligence) is the study of agents that perceive their environment,
reason about it, and take actions to achieve goals. Each such agent implements a
function that maps percept sequences to actions, and the study of these functions
is the subject of AI."
Agents: In AI, an agent is any entity that can perceive its environment and take
actions to achieve its goals. These agents can be physical robots, software
programs, or any system capable of interacting with the world. The agents sense
the environment through sensors and act on their environment through actuators.
An AI agent can have mental properties such as knowledge, belief, intention, etc.

Source: The Intelligent Agent depicted on books

Percepts: Percepts are the inputs or information an agent receives from the
environment. This could include data from sensors, cameras, microphones, or any
other sources of information.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Actions: Actions are the responses or behaviours the agent can exhibit to achieve
its goals or objectives. These can be physical movements, data processing,
decision-making, or any other form of output.
Function: The central idea of AI is to design a function (or algorithm) that maps
percept sequences to actions, allowing the agent to make intelligent decisions
based on its observations.
Intelligent Agent: This term refers to the function that enables the agent to act in
a way that is considered intelligent, i.e., making decisions that are adaptive and
can lead to goal achievement.
Russell and Norvig's definition underscore the goal of AI, which is to create
systems that can perceive their environment, reason about it, and take appropriate
actions to achieve specific objectives. The field of AI encompasses a wide range
of techniques and approaches, from symbolic reasoning to machine learning and
deep learning, all aimed at building intelligent agents that can perform tasks
typically associated with human intelligence
In Figure, there are eight explanations of AI shown in two groups. The top
ones talk about thinking and reasoning, while the bottom ones focus on behavior.
The left-side definitions judge based on how close to human-like performance
they are, while the right-side definitions assess based on an ideal measure called
rationality. A system is considered rational if it makes the best decisions based
on the information it has.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Let us look at the four approaches in more detail.
1. Acting humanly: The Turing Test approach
Turing Test : The Turing test is a method proposed by Alan Turing to evaluate a
machine's capability to exhibit intelligent behavior that is indistinguishable from
that of a human. In this test, a human evaluator engages in a conversation with
both a human and a machine through text, without knowing which is which. If
the evaluator cannot reliably differentiate between the machine's responses and
the human's, the machine is said to have passed the Turing test, demonstrating
human-like intelligence.

Image Source: [Link]

In order to pass Total Turing Test the computer must have the following features:
1. Proficiency in natural language processing for effective communication
in English.
2. Capability for knowledge representation to store acquired information.
3. Automated reasoning to utilize stored data for answering questions and
deriving new conclusions.
4. Machine learning to adapt to novel situations, detect patterns, and
extrapolate information.
5. Computer vision capabilities for object perception.
6. Robotics proficiency for manipulating objects and navigating its
surroundings.
These six disciplines encompass the majority of artificial intelligence (AI), and
Turing deserves recognition for designing a test that maintains its relevance six
decades later.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
2. Thinking humanly: The cognitive modeling approach
To ascertain whether a program exhibits human-like thinking, understanding
human thought processes is essential. This involves exploring the inner
workings of the human mind through introspection, psychological
experiments, and brain imaging. Once a precise theory of the mind is
established, it can be translated into a computer program. If the program's input-
output behavior aligns with human behavior, it suggests shared mechanisms.
3. Thinking rationally: The “laws of thought” approach

Aristotle, an early philosopher, sought to formalize “right thinking” through

syllogisms, offering patterns for sound reasoning. These principles, foundational
to logic, inspired logicians in the 19th century to develop precise notation for
various objects and their relations. By 1965, programs capable of theoretically
solving any problem in logical notation emerged, paving the way for the logicist
tradition in artificial intelligence.

4. Acting rationally: The rational agent approach

Computer agents are expected to autonomously operate, perceive their
environment, persist over time, adapt to change, and pursue goals. A rational
agent seeks the best outcome by emphasizing correct inferences, aligning with
the Turing Test skills. While correct inference is a facet of rationality, it's not
exhaustive, as some situations lack provably correct actions. The rational-agent
approach, incorporating knowledge representation, reasoning, and learning,
offers generality and scientific advantages. Despite challenges acknowledged in
the book, it adopts the hypothesis that perfect rationality is a useful starting
point, simplifying complex environments for foundational discussions in the
field.
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
1.2 The State of the Art: What can AI do today?

It's hard to give a short answer because AI is used in so many different areas.
Here are a few examples:

 Robotic Cars: In 2005, a driverless car named STANLEY drove through the
Mojave Desert at 22 mph and won a race by completing a 132-mile course.
It used cameras, radar, and lasers to see the environment and had
software to control the steering, braking, and acceleration. In 2007,
another car from CMU, called BOSS, safely drove through city streets in
traffic while following traffic rules and avoiding other cars and pedestrians.
 Speech Recognition: You can book a flight with United Airlines by talking
to an automated system that understands speech and responds
accordingly.
 Autonomous Planning and Scheduling: NASA used AI to manage the tasks
of a spacecraft far away from Earth. The AI planned and monitored its
actions, making sure everything was running smoothly. Newer versions of
this AI helped manage operations for the Mars rovers.
 Game Playing: IBM’s DEEP BLUE was the first computer to beat the world
chess champion, Garry Kasparov, in 1997. This was a huge moment for AI,
as computers began to show a new kind of intelligence in games.
 Spam Filtering: Every day, AI filters out spam emails from your inbox. It
learns and adapts to new tactics used by spammers, making it more
effective than traditional methods.
 Logistics Planning: During the 1991 Persian Gulf crisis, AI was used to plan
and schedule transportation for the U.S. military. It managed the
movement of thousands of vehicles, cargo, and people in a way that would
have taken much longer with older methods.
 Robotics: iRobot has sold over two million Roomba robotic vacuum
cleaners for home use. Their PackBot robots are used in Iraq and
Afghanistan to handle dangerous tasks like clearing explosives.
 Machine Translation: AI can translate languages, like turning Arabic into
English. For example, AI helped translate an Arabic headline about Turkey
and Cyprus without the programmers knowing Arabic—they just used
statistics and machine learning.

These are just some examples of what AI can do today. It’s not magic or science
fiction—just a mix of science, engineering, and math.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
 1.4 Intelligent Agents
Topics: Agents and environment, Intelligent Agents, Concept of Rationality, The
nature of environment, The structure of agents.
1.4.1 Agents and Environments:

An agent is defined as anything capable of perceiving its environment through

sensors and acting upon that environment through actuators. This basic concept
is depicted in Figure 2.1.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
For instance, a human agent employs eyes, ears, and other sensory organs while
utilizing hands, legs, vocal tract, etc., as actuators.
In contrast, a robotic agent might use cameras and infrared range finders as
sensors and various motors as actuators.
A software agent takes in sensory inputs such as keystrokes, file contents, and
network packets and responds by displaying on the screen, writing files, or
sending network packets.
Percept: The term "percept" refers to an agent's perceptual inputs at any given
moment, and a percept sequence encompasses the complete history of
everything the agent has perceived. Mathematically, an agent's behavior is
described by the agent function, which maps a given percept sequence to an
action.
Table: A table serves as an external representation of an agent's behaviour,
specifically presented in tabular form that outlines the agent's actions for every
possible percept sequences. The table contains entries corresponding to
different percept sequences and the corresponding actions the agent would take
in response to those sequences

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Agent Program: An agent program represents the internal implementation of an
agent's behavior. It is a concrete realization of the abstract agent function,
running within a physical system. The agent program executes the specified
actions determined by the agent function in response to the sensory inputs
received by the agent. It involves the use of computational processes,
algorithms, or code that dictates how the agent translates perceptual inputs into
actions within its environment.

Example: Consider a simple example—the vacuum-cleaner world as depicted in

Figure 2.2.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
This world consists of two locations, squares A and B, where the vacuum agent
perceives its location and the presence of dirt. The agent can choose to move
left, move right, suck up dirt, or do nothing. An example of a simple agent
function is presented: if the current square is dirty, then suck; otherwise, move
to the other square. Figure 2.3 provides a partial tabulation of this agent function
and Figure 2.8 provides agent program.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

1.4. 2: Intelligent Agents

In artificial intelligence (AI), intelligent agents are entities that perceive their
environment and take actions to achieve specific goals. These agents are
designed to operate autonomously, making decisions based on their perception
of the world and their programming or learning mechanisms. The concept of
intelligent agents is fundamental in AI as it provides a framework for building
systems that can exhibit intelligent behaviour.

Key characteristics of intelligent agents include:

Perception: Agents have the ability to perceive or sense their environment. This
involves gathering information from the surroundings through sensors or other
means.
Action: Intelligent agents can take actions in response to their perceptions.
These actions are chosen to influence the state of the environment or achieve
specific goals.
Autonomy: Agents operate autonomously, making decisions and taking actions
without direct human intervention. Autonomy allows them to adapt to changing
conditions in their environment.
Goal-Directed Behaviour: Intelligent agents are typically designed to achieve
specific goals. These goals can be predefined by a programmer or learned by the
agent through experience.
Learning and Adaptation: Many intelligent agents have the capability to learn
from experience and adapt their behaviour over time. This learning process may
involve acquiring new knowledge or improving decision-making strategies.
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Communication: In some cases, intelligent agents can communicate with other
agents or systems. Communication enables collaboration and coordination
between multiple agents to achieve common goals.

1.4.3 The Concept of Rationality (Good behaviour)

A rational agent is defined as one that consistently does right thing and makes
decisions leading to favourable outcomes. This is reflected in the correct
completion of every entry in the agent function table. However, determining
what constitutes the "right thing" involves a detailed consideration of
consequences.
 Assessing Consequences: To evaluate the rationality of an agent's actions,
one must examine the consequences of its behaviour. The agent's sequence
of actions influences the environment, causing it to transition through
different states. Desirability of this sequence is gauged by a performance
measure that evaluates resultant environment states.
 Performance Measure: Performance measures play a pivotal role in
evaluating an agent's behavior and defining success. These measures are
instrumental in shaping the assessment of successful behavior, focusing
specifically on the resulting environment states rather than the agent's
subjective perception of its performance.

 Challenges and Considerations in Designing Performance

Measures: Designing a suitable performance measure is a
multifaceted task that involves careful consideration of desired
outcomes in the environment. This complex process requires
avoiding potential pitfalls, such as the introduction of biased
measures influenced by the agent's personal opinion.

o Example: Taking the vacuum-cleaner agent as an example,

the design of a performance measure becomes critical.
Simply measuring dirt cleaned up in an eight-hour shift may
lead to undesired behaviours, such as repetitive cleaning
and dumping. A more effective measure would reward the
agent for maintaining a clean floor.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
  Alignment with Environmental Goals: Emphasizing the
importance of aligning performance measures with actual
environmental goals is crucial. This underscores the need to base
measures on tangible objectives within the environment, steering
clear of preconceived notions about how the agent should behave.
 Philosophical Considerations: Even with careful performance measure
design, philosophical questions arise. Comparisons between agents with
different approaches (e.g., a consistent but mediocre performer vs. an
energetic but intermittently idle cleaner) prompt deeper reflections on
preferences in various contexts.

1.4.4 What is Rationality?

Rationality encompasses the state of being reasonable, sensible, and possessing

a sound judgment. Rational agent is the one, which is capable of doing expected
actions to maximize its performance measure, on the basis of −
 Its percept sequence.
 Its built-in knowledge base.

A rational agent always performs right action, where the right action means the
action that causes the agent to be most successful in the given percept
sequence. The problem the agent solves is characterized by Performance
Measure, Environment, Actuators, and Sensors (PEAS).

Rationality at Any Given Time depends on the following four things:

1. Performance Measure: The criterion defining success.
2. Agent’s Prior Knowledge: Understanding of the environment.
3. Available Actions: The actions the agent can perform.
4. Percept Sequence: The agent’s historical sensory input.

1.4.5 Omniscience, Learning, and Autonomy in Rational Agents

Omniscience and Rationality: Distinguishing between rationality and

omniscience is crucial. An omniscient agent knows the actual outcome of its
actions and can act accordingly; but omniscience is impossible in reality. While
rationality maximizes expected performance.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Information Gathering : Rational agents should perform actions to modify future
percepts, known as information gathering. This concept is vital for maximizing
expected performance.
Learning in Rational Agents: Rational agents are not only expected to gather
information but also to learn from their perceptions. An agent's initial
configuration may reflect prior knowledge, but as it gains experience, learning
becomes imperative.
Think about the dung beetle, a small insect that carefully constructs its nest, lays
eggs, and uses a ball of dung to seal the entrance. If the dung ball is taken away
while the beetle is carrying it, the beetle continues its job as if the dung ball is
still there, not realizing its missing. Evolution has shaped the beetle's behaviour
to expect the dung ball, and any change from this expectation results in
unsuccessful actions.
The sphex wasp is a bit smarter. It digs a hole, stings a caterpillar, drags it into
the hole, checks everything is fine, and then lays eggs. The caterpillar becomes
food for the hatching eggs. However, if someone moves the caterpillar a little
while the wasp is checking, it goes back to dragging the caterpillar, not realizing
the change. Even after many attempts to move the caterpillar, the wasp doesn't
learn that its plan isn't working and keeps doing it the same way.

Autonomy in Rational Agents: Autonomy is emphasized as a crucial aspect of

rationality. An autonomous agent learns to compensate for partial or incorrect
prior knowledge, ensuring effective adaptation to the environment. While
acknowledging the practical need for some initial knowledge, over time, a
rational agent's behavior becomes effectively independent of its initial
knowledge through the incorporation of learning.
In essence, the combination of rationality, information gathering, and learning
enables the design of autonomous agents capable of success across diverse
environments.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
1.4.4 The Nature of Environments
The task environment, in the context of rational agents, refers to the external
system or surroundings within which an agent operates and performs tasks. It
encompasses the conditions, challenges, and dynamics that influence the agent's
behavior and decision-making processes. The task environment plays a pivotal
role in shaping the rationality, learning, and autonomy of an agent.

Specifying the Task Environment

In the design of any agent, the initial crucial step is to comprehensively specify
the "task environment," referred to as PEAS (Performance, Environment,
Actuators, Sensors).
To illustrate the concept, consider the example of a more intricate problem: an
automated taxi driver.
The PEAS description for the taxi's task environment is summarized in Figure 2.4,
encompassing aspects such as performance measures, driving environment,
actuators, and sensors.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
Performance Measure: Determining the desired qualities for the automated
driver, such as reaching the correct destination, minimizing fuel consumption,
trip time, or cost, adhering to traffic laws, and maximizing safety and passenger
comfort, involves addressing conflicting goals and necessitates trade-offs.

Driving Environment: The taxi must navigate a diverse range of roads, contend
with various traffic conditions, interact with passengers, and make optional
choices like operating in different regions with distinct weather conditions or
driving on different sides of the road. The complexity of the environment directly
impacts the design challenge.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Actuators and Sensors: Actuators for the automated taxi include control over
the engine, steering, and braking, along with interfaces for communication with
passengers and other vehicles. Sensors comprise controllable video cameras,
infrared or sonar sensors, speedometer, accelerometer, and various vehicle state
sensors.

Figure 2.5 outlines the basic PEAS elements for different agent types,
showcasing the variety of considerations in specifying task environments.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

1.4.5 Properties of Task Environments

In AI, there are various task environments, and we can classify them based on a
few key dimensions as follows:

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
1. Fully Observable vs. Partially Observable:
Fully Observable: The agent can see the complete environment all the time.
Partially Observable: The agent's sensors may not capture all relevant
aspects due to noise or missing information.

2. Single Agent vs. Multiagent:

Single Agent: An agent operates independently, like solving a crossword
puzzle.
Multiagent: Agents interact with each other, introducing complexities in
decision-making.

3. Deterministic vs. Stochastic:

Deterministic: The environment's next state is entirely determined by the
current state and the agent's action.
Stochastic: Some uncertainty exists, making predictions challenging. For
example, traffic behavior in taxi driving is stochastic.

4. Episodic vs. Sequential:

Episodic: Each decision is independent, not influenced by past decisions.
Sequential: Current decisions affect future ones, as seen in chess or taxi
driving.

5. Static vs. Dynamic:

Static: The environment remains unchanged during decision-making.
Dynamic: The environment evolves continuously, demanding constant
decision-making.

6. Discrete vs. Continuous:

Discrete: Finite states, actions, or percepts.
Continuous: Values or actions vary smoothly.

7. Known vs. Unknown:

Known: The agent knows outcomes or probabilities.
Unknown: The agent needs to learn about the environment.
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
Understanding these dimensions helps in determining the complexity of an
environment. For instance, taxi driving is challenging due to being partially
observable, multiagent, stochastic, sequential, dynamic, continuous, and
unknown in many cases.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

Figure 2.6 lists the properties of a number of familiar environments.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
1.4.6 The structure of agents
The structure of the agent comprises two essential components: the
architecture and the program. The architecture refers to the combination of a
computing device equipped with physical sensors and actuators. The program,
created by AI, embodies the agent function, determining how the agent
translates percepts into actions. The overall structure is denoted by the
equation:
agent = architecture + program.

Agent Programs: All agent programs share a common structure: they take the
current percept as input from the sensors and produce an action for the
actuators.

It's crucial to note the distinction between the agent program, which operates
based on the current percept, and the agent function, which considers the entire
percept history. The agent programs are presented in a simple pseudocode
language.

An example in Fig 2.7 illustrates a basic agent program that tracks the percept
sequence and uses it to determine the appropriate action from a table.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

The table, exemplified for the vacuum world in Figure 2.3, explicitly represents
the agent function embodied by the agent program.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
Constructing a rational agent using this approach requires designers to create a
table containing actions for every possible percept sequence. However, the
table-driven approach faces insurmountable challenges. The size of the lookup
table grows exponentially with the set of possible percepts and the agent's
lifetime, making it impractical for complex environments. For instance, a lookup
table for an automated taxi's visual input would be astronomically large. Despite
these limitations, TABLE-DRIVEN-AGENT fulfils the objective of implementing the
desired agent function.

Types of Agent Programs: There are, five basic types of agent programs as given
below:
1. Simple reflex agents
2. Model-based reflex agents
3. Goal-based agents
4. Utility-based agents
5. Learning Agents
Each kind of agent program combines particular components in particular ways
to generate actions.

1. Simple Reflex Agents:

The simplest kind of agent is the simple reflex agent. These agents select actions
on the basis of the current percept, ignoring the rest of the percept history.

Example: For example, the vacuum agent whose agent function is tabulated in
Figure 2.3 is a simple reflex agent, because its decision is based only on the
current location and on whether that location contains dirt. An agent program
for this agent is shown in Figure 2.8.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
The program illustrated in Figure 2.8 is designed for a specific vacuum
environment.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015
Figure 2.9 provides a schematic representation of this generalized program,
demonstrating how condition-action rules enable the agent to establish
connections from perception to action. Rectangles represent the current
internal state of the agent's decision process, while ovals depict background
information used in the process.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
The agent program, depicted in Figure 2.10, is also straightforward. The
INTERPRET-INPUT function generates an abstracted description of the current
state from the percept, and the RULE-MATCH function identifies the first
matching rule in the rule set based on the given state description. Note that the
conceptual terms "rules" and "matching" are used, but actual implementations
can be as simple as a collection of logic gates constituting a Boolean circuit.

2. Model based Reflex Agents:

In order to address the challenges posed by partial observability, it is most
effective for an agent to maintain an internal state that reflects the unobserved
aspects of the current state. Updating this internal state requires encoding two
types of knowledge in the agent program. First, knowledge about how the world
evolves independently of the agent's actions, and second, information about
how the agent's actions affect the world. This knowledge is referred to as a
"model of the world," and an agent using such a model is termed a model-based
agent.

Figure 2.11 illustrates the structure of a model-based reflex agent with an

internal state, where the current percept combines with the old internal state
to generate an updated description of the current state based on the agent's
model of the world.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
The agent program is shown in Figure 2.12. The UPDATE-STATE function is crucial
in this process, responsible for creating the new internal state description.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

3. Goal-based agents :
Deciding on actions based solely on the current state of the environment may
not always be sufficient. The agent's program combines the goal information
with its model, similar to the model-based reflex agent, to select actions that
align with the goal. Figure 2.13 illustrates the structure of a goal-based agent.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
4. Utility based Agents:
Achieving high-quality behaviour in various environments requires more than
just setting goals. While goals indicate success or failure, they don't distinguish
between different ways of achieving them in terms of efficiency, safety,
reliability, or cost. A more detailed evaluation is needed, and this is where the
concept of utility comes into play.

Utility, a term used by economists and computer scientists, represents a more

refined measure of desirability than the binary distinction provided by goals. A
utility function internalizes the performance measure, assigning a score to
different sequences of environmental states. If the internal utility function
aligns with the external performance measure, an agent maximizing utility is
considered rational. The utility-based agent structure is presented in Figure 2.14.

Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

5. Learning Agents

Learning offers the advantage of enabling agents to operate effectively in

unfamiliar environments, surpassing their initial knowledge. A learning agent
comprises four conceptual components, as illustrated in Figure 2.15. The crucial
distinction lies between the learning element, responsible for improvements,
and the performance element, tasked with selecting external actions. The

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
learning element utilizes feedback from the critic to enhance the performance
element for future improvement.
The critic informs the learning element of the agent's performance relative to a
fixed standard. This standard is crucial, as percepts alone do not indicate the
agent's success.
The problem generator, the last component, suggests actions for new and
informative experiences. While the performance element tends to favor optimal
actions based on current knowledge, the problem generator encourages
exploration for potentially superior long-term actions.

The automated taxi example: The performance element involves the knowledge
and procedures guiding the taxi's driving actions, while the critic observes the
world and provides feedback to the learning element. The learning element, in
turn, formulates rules based on feedback, modifying the performance element
accordingly. The problem generator suggests experiments to improve certain
behaviors.

How the components of agent programs work

The representations of components of agents can be categorized along an axis

of increasing complexity and expressive power: atomic, factored, and
structured.

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|
Atomic representations view each state of the world as indivisible, treating them
as black boxes with no internal structure. This simplification is exemplified in
algorithms like search, game-playing, Hidden Markov models, and Markov
decision processes.

Factored representations, on the other hand, break down each state into a fixed
set of variables or attributes, allowing shared attributes between different
states. This facilitates dealing with uncertainty, as attributes can be left blank.
Factored representations underlie various AI areas, including constraint
satisfaction algorithms, propositional logic, planning, Bayesian networks, and
machine learning algorithms.

Structured representations go beyond variables and values, explicitly describing

objects and their relationships. This level of representation is essential for
capturing complex scenarios, such as the interaction between a truck and a loose
cow blocking its path. Structured representations are foundational to relational
databases, first-order logic, first-order probability models, knowledge-based
learning, and natural language understanding.

Expressiveness is the key axis along which these representations lie, with
increasing expressiveness allowing more concise representation of information.
However, as expressiveness grows, reasoning and learning become more
complex. Balancing the benefits and drawbacks of expressive representations,
real-world intelligent systems may need to operate at various points along this
axis simultaneously to effectively handle diverse scenarios.
Source Book: Stuart J. Russell and Peter Norvig, Artificial Intelligence, 3rd Edition, Pearson,2015

[Link] G S, Professor and HoD, Dept of CSE, SDM Institute of Technology, Ujire-574240
|