Introduction to Intelligent Agents

Maria Gini
CSci 8551, Fall 2017

Maria GiniCSci 8551, Fall 2017 1

 Outline
♦ What is an agent? (slides from Russell and Norvig)
♦ Types of agents (slides from Russell and Norvig)
♦ Why should we use agents?
♦ What’s special about agent-based computing?

Thanks to Russell and Norvig for slides 3 to 12

Maria GiniCSci 8551, Fall 2017 2

 What is an agent? (Part I)
“An agent is anything that can be viewed as perceiving its environment
through sensors and acting upon that environment through effectors,” (from
Russell and Norvig, Artificial Intelligence: a Modern Approach)
If we define the environment as whatever provides input and receives output
(input = sensing and output = acting) then every program is an agent.
sensors

percepts
?
environment
agent
actions

actuators

The agent function maps from percept histories to actions: f : P ∗ → A

The agent program runs on the physical architecture to produce f

Maria GiniCSci 8551, Fall 2017 3

 Rationality
Fixed performance measure evaluates the environment sequence
– one point per square cleaned up in time T ?
– one point per clean square per time step, minus one per move?
– penalize for > k dirty squares?
A rational agent chooses whichever action maximizes the expected value of
the performance measure given the percept sequence to date
Rational =
6 omniscient
Rational =6 clairvoyant
Rational = 6 successful
Rational ⇒ exploration, learning, autonomy

Maria GiniCSci 8551, Fall 2017 4

 Vacuum-cleaner world

A B

Percepts: location and contents, e.g., [A, Dirty]

Actions: Lef t, Right, Suck, N oOp
Percept sequence Action
[A, Clean] Right
[A, Dirty] Suck
[A, Clean], [A, Clean] Right
[A, Clean], [A, Dirty] Suck
.. ..

Maria GiniCSci 8551, Fall 2017 5

 Simple reflex agents
Agent Sensors

What the world

is like now

Environment
Condition−action rules What action I
should do now

Actuators

The agent works by finding a rule whose condition matches the current
situation, as defined by perception, and then doing the action associated
with the rule. The agent has no memory.

Maria GiniCSci 8551, Fall 2017 6

 Vacuum-cleaner world

A B

Percepts: location and contents, e.g., [A, Dirty]

Actions: Lef t, Right, Suck, N oOp

function Reflex-Vacuum-Agent( [location,status]) returns an action

if status = Dirty then return Suck
else if location = A then return Right
else if location = B then return Left

What is the right function? Can it be implemented in a small agent program?

Maria GiniCSci 8551, Fall 2017 7

 Reflex agents with state

Sensors
State
What the world
How the world evolves is like now

Environment
What my actions do

Condition−action rules What action I

should do now

Agent Actuators

The agent works by finding a rule whose condition matches the current
situation, as defined by perception and by the stored internal state, and
then doing the action associated with the rule. The internal state acts as a
memory and allows for a better selection of the rule to apply.
Maria GiniCSci 8551, Fall 2017 8
 Goal-based agents

Sensors
State
What the world
How the world evolves is like now

Environment
What it will be like
What my actions do if I do action A

Goals What action I

should do now

Agent Actuators

The agent has an explicit goal, and when choosing an action it will select an
action that achieves the goal. The decision process requires some plannig
(”what will happen if I do action A”). This agent is more flexible than the
simple reflex agent, and is capable of achieving goals.
Maria GiniCSci 8551, Fall 2017 9
 Utility-based agents

Sensors
State
What the world
How the world evolves is like now

Environment
What it will be like
What my actions do if I do action A

How happy I will be

Utility in such a state

What action I
should do now

Agent Actuators

The agent has a utility function that maps a state onto real number that
describes how well the agent is performing. This agent not only achieves
goals, but it maximizes some measure of performance.

Maria GiniCSci 8551, Fall 2017 10

 Learning agents
Performance standard

Critic Sensors

feedback

Environment
changes
Learning Performance
element element
knowledge
learning
goals

Problem
generator

Agent Actuators

Maria GiniCSci 8551, Fall 2017 11

 PEAS
To design a rational agent, we must specify the task environment
Consider, e.g., the task of designing an automated taxi:
Performance measure?? safety, destination, profits, legality, comfort, . . .
Environment?? US streets/freeways, traffic, pedestrians, weather, . . .
Actuators?? steering, accelerator, brake, horn, speaker/display, . . .
Sensors?? video, accelerometers, gauges, engine sensors, keyboard, GPS, . . .
The environment type largely determines the agent design
The real world is (of course) partially observable, stochastic, sequential,
dynamic, continuous, multi-agent.

Maria GiniCSci 8551, Fall 2017 12

 What is an agent? (Part II)

A weak notion (from Woodridge and Jennings):

• autonomy: agents operate without direct intervention of humans, and

have control over their actions and internal state;
• social ability: agents interact with other agents (and possibly humans)
via an agent communication language;
• reactivity: agents perceive their environment and respond in a timely and
rational fashion to changes that occur in it;
• pro-activeness: agents do not simply act in response to their environment,
they are capable of taking the initiative (generate their own goals and act
to achieve them).

for details see M. Wooldridge and N. R. Jennings. Intelligent Agents: Theory

and Practice. In Knowledge Engineering Review 10(2), 1995.

Maria GiniCSci 8551, Fall 2017 13

 What is an agent? (Part III)

A stronger notion (from Woodridge and Jennings):

An agent has mental properties, such as knowledge, belief, intention, obli-
gation. In addition, and agent has other properties such as:

• mobility: agents can move around from one machine to another and
across different system architectures and platforms;
• veracity: agents do not knowingly communicate false information;
• benevolence: agents always try to do what they are asked of;
• rationality: agents will try to achieve their goals and not act in such a
way to prevent their goals from being achieved.

Maria GiniCSci 8551, Fall 2017 14

 Discussion
Are these agents?

• a thermostat
• a payroll program
• a spell checker

Can you suggest domains and applications where agents could/should be

used?

Maria GiniCSci 8551, Fall 2017 15

 Why should we use agents?
Agents provide three fundamental abstractions:

1. the autonomy abstraction;

2. the cognitive abstraction;
3. the society of agents abstraction.

Computer scientists use abstraction mechanisms (procedural abstractions,

abstract data types, objects, etc.) all the time.

Maria GiniCSci 8551, Fall 2017 16

 The autonomy abstraction

• Autonomy implies ability to make decisions and to initiate actions without

direct human supervision. This allows us to view agents as entities that
have goals to achieve and are capable of initiating actions to achieve their
goals.
• Autonomy often implies the notion of persistence and identity.
• Autonomy also often implies the existence of other autonomous enties.
This allows us to see agents as members of a society of agents (see last
point).

Maria GiniCSci 8551, Fall 2017 17

 The cognitive abstraction

• In everyday life, we use folk psychology to explain and predict the behavior
of complex intelligent systems, such as people.
Example: Maria intended to prepare her slides.
Example: Jane believed it was raining.
• The term intentional system was conied by Daniel Dennett to describe
entities “whose behavior can be predicted by the method of attributing
belief, desires and rational acumen.”
• the term BDI (belief, desires, and intentions) is used to describe agent
architecture that have those properties.
• The intentional stance is a convenient way of talking about complex sys-
tems, which allows us to predict and explain their behavior without having
to understand how their mechanism actually works.

Maria GiniCSci 8551, Fall 2017 18

 Example of cognitive abstraction
Consider a light switch: “It is perfectly coherent to treat a light switch as a
(very cooperative) agent with the capability of transmitting current at will,
who invariably transmits current when it believes that we want it transmitted
and not otherwise; flicking the switch is simply our way of communicating
our desires.” (Yoav Shoham)
The intentional stance description is consistent, but “.. it does not buy us
anything, since we essentially understand the mechanism sufficiently to have
a simpler, mechanistic description of its behaviour.” (Yoav Shoham)

Maria GiniCSci 8551, Fall 2017 19

 The society of agents abstraction
Agents are not entities living alone, they are part of a society of agents.
This means agents can share knowledge (ontologies), communicate with each
other (Agent communication languages, speech acts), cooperate, coordinate,
and negotiate ... but members of a society can also lie, take advantage of
others, masquerade themselves as someone else, etc.
Social laws need to be developed to ensure that agents deliver what they
promise, do not harm other agents, and act rationally.
Cooperation = working together as a team to achieve a shared goal
Coordination = managing the interdependencies between activities.
Negotiation = reaching agreements on matters of common interest.

Maria GiniCSci 8551, Fall 2017 20

 What’s special about agent-based computing?

• Openess. When components of the system are not known in advance,

change over time, and are highly heterogeneous (e.g. programming the
Internet) an agent-based approach allows to create systems that are flex-
ible, robust, and can adapt to the environment by using their social skills,
ability to negotiate with other agents, and ability to take advantage of
opportunities.
• Handle complexity. With large and complex problems, agents offer a
natural way to partition the problem into smaller and simpler components,
that are easier to develop and maintain, and are specialized.
• Natural Metaphor. Agents provide an easy way to conceptualize metaphors.
For instance, an e-mail filtering program can be presented using the
metaphor of a personal digital assistant.

Maria GiniCSci 8551, Fall 2017 21

 What’s special about agent-based computing

• Non determinism. It is almost impossible to predict in advance how agents

will interact. Agents can learn and change their behavior, it is not know
what other agents will exist and what they will do.
• Global behavior emerges from local behavior. Individual behaviors are de-
signed for individual agents, the global behavior of the system will emerge
at run time. Minsky’s “Society of Mind” treats agents as individually very
simple, but giving rise to “intelligence” when acting together, “in certain
very special ways” in societies.
• New social laws. Social laws need to be developed to ensure that agents
deliver what they promise, do not harm other agents, and act rationally.
Agents are to carry out tasks on behalf of the user in an environment
where the user has something significant on the line, such as money,
reputation, privacy, security, etc.

Maria GiniCSci 8551, Fall 2017 22

 Disciplines used in agents research

• Artificial Intelligence provides methods for representing knowledge and

reasoning about it
• Cognitive Science provides tools for modeling cognitive processes, such
as desires, beliefs, intentions, and commitments.
• Game Theory provides mathematical foundations to decision making in
the context of interactions with other players.
• Algorithms and Distributed Systems provide theory and practical meth-
ods for solving didtributed problems, such as constraint satisfaction and
optimization.

Maria GiniCSci 8551, Fall 2017 23

 Where to look for papers on agents

• Journals
– Journal of Autonomous Agents and Multi-Agent Systems (JAAMAS)
– Journal of AI Research (JAIR)
– Artificial Intelligence Journal (AIJ)
– specialized journals
• Conferences
– AAMAS (Autonomous Agents and Multi-Agent Systems)
– AAAI annual conference
– IJCAI (Int’l Joint Conference on AI)
– specialized conferences for robotics (ICRA, DARS, IAS, etc)

Maria GiniCSci 8551, Fall 2017 24