Rational agents: robots & artificial intelligence
Lecture 1 – AI as a discipline / Agent architectures
Alan Turing: “Can machines think?”
John Searle: “The answer is, obviously, yes. We are exactly such machines.”
Irvin John Good (1965): “The first ultra-intelligent machine is the last invention that man
need ever make, provide that the machine is docile enough to tell us how to keep it under
control.”
Can machines think? – Alan Turing (1950)
What do you mean by “machines”?
o Computation => Turing machine
Turing machine: a conceptual machine with an infinite tape. If
something cannot be solved by a Turing machine, it cannot be solved.
What do you mean by “thinking” ?
o Behavior => Turing test
Turing test: If a machine is mistakenly perceived as a human, it
passes the Turing test.
The discipline of Artificial Intelligence (AI) is shaped by multiple scientific disciplines:
Philosophy
Mathematics
Economics (game-theory and decision-theory)
Neuroscience
Psychology
Computer engineering
Control theory and cybernetics
Linguistics (How does language relate to thought? Understanding language requires
aaaaaaaa an understanding of the subject matter and context, not just an
aaaaaaaaa understanding of the structure of sentences.)
The discipline Artificial Intelligence (AI) tries to understand how (human) intelligence works,
and how intelligent agents can be constructed. Within AI
different perspectives are employed (human vs. rational /
thinking vs. acting).
There are five different agent architectures:
Simple Reflex agent
,A simple reflex agent will only work if the correct decision can be made on the basis of the
current percept alone (fully observable). It uses a list of if-statements which can be triggered
by percepts.
It used condition-action rules to know when to act on a certain percept.
Example of condition-action rule : if the car in front is braking, then initiate braking.
Model-based reflex agent
A model-based reflex agent also works with condition-
action rules in order to know when to act on a situation.
However, it also builds a model of the environment.
Example of a model: predicting where the cars are
that I cannot see at the moment, on the basis of
information obtained earlier.
Model-based goal agent
A model-based goal agent does not work with condition-
action rules, but has one or more goals to satisfy. It also
builds a model of the environment and it uses this to know
the consequences of its actions on the environment.
Example:
o Goal: reaching the customer’s destination
o Decision: go left, right or straight?
o Choice: (sequence of) action(s) that realizes
(approaches) the goal.
Model-based utility agent
In addition to goals, a model-based utility agent also have
been given a certain value of utility. This value gets altered
every action the agent makes. Depending on whether this
action is an action which we want the agent to make the
utility increases or decreases. Due to the agent knowing
what its utility will be in a certain state, it can choose to
make the right sequence of actions to get the highest value
of utility.
Example:
o Goal: reaching the customer’s destination
(this can be done in many ways)
o Utility: built-in preference for a certain way
As safe as possible
As fast as possible
Learning agent
The learning agent consists of a learning element and a
performance element.
, The learning element is responsible for making changes.
The performance element is responsible for selecting external actions.
The learning element uses feedback from the critic on how the agent is doing with respect to
a fixed performance standard and determines how the performance element should be
modified to perform better in future. (Percepts do not provide an indication of success.)
The problem generator is responsible for suggesting actions that will lead to new and
informative experiences (the performance element would keep doing the actions that are
best, given what it knowns).
Lecture 3 – Perceptions and actions / Rationality of agents
A simple reflex agent vacuum cleaner.
Agent function: percept sequence -> actions
The agent function characterizes the agent’s behaviour
from the outside.
Percept sequence Action
[A, Clean] Right
[A, Clean][A, Clean][A, Dirty] Suck
Agent program: prior knowledge (condition-action rules) + percept sequence -> actions
The agent program is a concrete/specific implementation of the behaviour in a
physical system, it describes what goes on inside the agent.
The vacuum cleaner agent program:
There are many possible agent functions, but when is an agent’s behaviour good? (rational)
This depends on the consequences of the behaviour of the agent: a sequence of actions
causes a sequence of states of the environment, if this sequence is desirable (from the
human point of view!) the agent has performed well.
Performance measure captures what is desirable, e.g. points awarded for: amount
of dirt sucked in a given period of time; number of clean locations per time step over
that period.
Rationality
What is rational at any given time depends on four things:
The performance measure, defines the criterion of success;
The agent’s prior knowledge of the environment;
The actions that the agent can perform;
The agent’s percept sequence to date.
Lecture 1 – AI as a discipline / Agent architectures
Alan Turing: “Can machines think?”
John Searle: “The answer is, obviously, yes. We are exactly such machines.”
Irvin John Good (1965): “The first ultra-intelligent machine is the last invention that man
need ever make, provide that the machine is docile enough to tell us how to keep it under
control.”
Can machines think? – Alan Turing (1950)
What do you mean by “machines”?
o Computation => Turing machine
Turing machine: a conceptual machine with an infinite tape. If
something cannot be solved by a Turing machine, it cannot be solved.
What do you mean by “thinking” ?
o Behavior => Turing test
Turing test: If a machine is mistakenly perceived as a human, it
passes the Turing test.
The discipline of Artificial Intelligence (AI) is shaped by multiple scientific disciplines:
Philosophy
Mathematics
Economics (game-theory and decision-theory)
Neuroscience
Psychology
Computer engineering
Control theory and cybernetics
Linguistics (How does language relate to thought? Understanding language requires
aaaaaaaa an understanding of the subject matter and context, not just an
aaaaaaaaa understanding of the structure of sentences.)
The discipline Artificial Intelligence (AI) tries to understand how (human) intelligence works,
and how intelligent agents can be constructed. Within AI
different perspectives are employed (human vs. rational /
thinking vs. acting).
There are five different agent architectures:
Simple Reflex agent
,A simple reflex agent will only work if the correct decision can be made on the basis of the
current percept alone (fully observable). It uses a list of if-statements which can be triggered
by percepts.
It used condition-action rules to know when to act on a certain percept.
Example of condition-action rule : if the car in front is braking, then initiate braking.
Model-based reflex agent
A model-based reflex agent also works with condition-
action rules in order to know when to act on a situation.
However, it also builds a model of the environment.
Example of a model: predicting where the cars are
that I cannot see at the moment, on the basis of
information obtained earlier.
Model-based goal agent
A model-based goal agent does not work with condition-
action rules, but has one or more goals to satisfy. It also
builds a model of the environment and it uses this to know
the consequences of its actions on the environment.
Example:
o Goal: reaching the customer’s destination
o Decision: go left, right or straight?
o Choice: (sequence of) action(s) that realizes
(approaches) the goal.
Model-based utility agent
In addition to goals, a model-based utility agent also have
been given a certain value of utility. This value gets altered
every action the agent makes. Depending on whether this
action is an action which we want the agent to make the
utility increases or decreases. Due to the agent knowing
what its utility will be in a certain state, it can choose to
make the right sequence of actions to get the highest value
of utility.
Example:
o Goal: reaching the customer’s destination
(this can be done in many ways)
o Utility: built-in preference for a certain way
As safe as possible
As fast as possible
Learning agent
The learning agent consists of a learning element and a
performance element.
, The learning element is responsible for making changes.
The performance element is responsible for selecting external actions.
The learning element uses feedback from the critic on how the agent is doing with respect to
a fixed performance standard and determines how the performance element should be
modified to perform better in future. (Percepts do not provide an indication of success.)
The problem generator is responsible for suggesting actions that will lead to new and
informative experiences (the performance element would keep doing the actions that are
best, given what it knowns).
Lecture 3 – Perceptions and actions / Rationality of agents
A simple reflex agent vacuum cleaner.
Agent function: percept sequence -> actions
The agent function characterizes the agent’s behaviour
from the outside.
Percept sequence Action
[A, Clean] Right
[A, Clean][A, Clean][A, Dirty] Suck
Agent program: prior knowledge (condition-action rules) + percept sequence -> actions
The agent program is a concrete/specific implementation of the behaviour in a
physical system, it describes what goes on inside the agent.
The vacuum cleaner agent program:
There are many possible agent functions, but when is an agent’s behaviour good? (rational)
This depends on the consequences of the behaviour of the agent: a sequence of actions
causes a sequence of states of the environment, if this sequence is desirable (from the
human point of view!) the agent has performed well.
Performance measure captures what is desirable, e.g. points awarded for: amount
of dirt sucked in a given period of time; number of clean locations per time step over
that period.
Rationality
What is rational at any given time depends on four things:
The performance measure, defines the criterion of success;
The agent’s prior knowledge of the environment;
The actions that the agent can perform;
The agent’s percept sequence to date.
