10 knowledge representation (compatible) -part 3

CP468, Dr. Reem K. Al-Halimi 1

Study Material: •Part III up to end of Chapter 7 in Luger, Artificial Intelligence: Structures and Strategies.

Other References:•James Odell, “Objects and Agents Compared” in Journal of Object Technology, Vol 1, Issue 1, May-June 2002. http://www.jot.fm/issues/issue_2002_05/column4/column4.pdf

Knowledge RepresentationProblem: represent human knowledge into

computationally acceptable languageDesired Features

Exhaustiveness All needed information is in KB.

modifiability new information can be added without sacrificing consistency.

homomorphic mapping of objects information organized in a natural and intuitive fashion

Computational Efficiency

2CP468, Dr. Reem K. Al-Halimi

Approaches to Problem Solving in AIDifferent views

1. Weak Problem Solvers: To create intelligent systems, we simply need to transform the syntactic form of the start state to match that of the desired goal state. Example: General Problem Solver by Newell and Simon

2. Strong Problem Solvers: To create a system that acts intelligently we must represent world knowledge in a form accessible to the system. Example: expert systems such as MYCIN.

3. Subsumption Architecture: “The world is its own model”.

4. Genetic.


Explicit Representation of World KnowledgeLogic as a knowledge representation

languagePropositional LogicPredicate Logic (FOL)

Semantic networksFramesConceptual Dependency


Knowledge Representation Hypothesis1. Knowledge is represented propositionally

(i.e. in a form that explicitly represents the knowledge in question).

2. The behaviour of a system is seen as formally caused by the represented knowledge.


Semantic NetworksDefine objects in terms of their association with

other objectse.g. snow, white, snowman, ice, slippery.

Represent knowledge as a graph:

Concepts at lower levels inherit characteristics from their parent concepts.

Concepts

Relations


Semantic NetworksWell designed semantic networks are a form

of logic.

memberOf(femalePersons, mary)

female Person

s

memberOf

mary


Fig 7.2 Network representation of properties of snow and ice (From: Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009)

8

Semantic NetworksExample

CP468, Dr. Reem K. Al-Halimi


female Person

s

memberOf

mary

male Person

s

memberOf

john

mammals

subsetOf

Persons

subsetOfsubsetOf

sisterOf

legs 2hasMoth

er

legs 1


Semantic NetworksInference MechanismInheritance

e.g. Persons by default have 2 legs. How many legs does Mary have? John?

Use of Inverse Links (through reification)e.g. hasSister(p, s) and sisterOf(s, p)

hasSister

inverseOf

sisterOf



female Person

s

memberOf

mary

male Person

s

memberOf

john

mammals

subsetOf

Persons

subsetOfsubsetOf

sisterOf

legs 2hasMoth

er

legs 1

hasSister


Semantic NetworksAdvantagesSimple and transparent inference processes.Ability to assign default values for categories.Ability to include procedural attachment.


Semantic NetworksDisadvantagesSimple query language may be too limiting to

express complex queries.Does not represent full FOL since it does not

provide means to use negation, disjunction, and existential quantification.

n-ary functions must be mapped onto binary functions.


Semantic NetworksMuch of this work has been done in the arena

of natural language.First implementation in machine translation in

the early 60s.Quillian’s dictionary (late 1960s):

Planes contain single word definitions.Words are defined in terms of other words in a

semantic network format.Program used definitions to find relationships

between pairs of words. (e.g. comfort and cry produce sad)


Fig 7.3 three planes representing three definitions of the word “plant” (Quillian, 1967). (Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009) 15

Semantic Networks


Semantic NetworksWhat Relationships do We Need?Conceptual Dependency theory:

primitives of meaning1. Actions2. Objects3. modifiers of actions4. modifiers of objects


Semantic NetworksWhat Relationships do We Need?Conceptual Dependency Theory:

1. Actions1. transfer a relationship (give)2. transfer physical location of an object (go)3. apply physical force to an object (push)4. move body part by owner (kick)5. grab an object by an actor (grasp)6. ingest an object by an animal (eat)7. expel from an animal’s body (tell)8. transfer mental information (decide)9. conceptualize or think about an idea (think)10. produce sound (say)11. focus sense organ (listen)



primitives of meaning1. Actions2. Objects3. modifiers of actions4. modifiers of objects

conceptual syntax rules built using these primitives constitute a grammar of meaningful semantic

relationships.conceptual dependency relationships

are defined using the conceptual syntax rules can be used to construct an internal representation

of an English sentence.18CP468, Dr. Reem K. Al-Halimi


conceptual dependency relationships are defined using the conceptual syntax rules can be used to construct an internal

representation of an English sentence. Tense and mode are added.

Example: past future transition etc.


Fig 7.6 Conceptual dependencies (Schank and Rieger, 1974). (From: Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009)

20

Semantic Networks


Semantic NetworksConceptual Dependency Example

Example:“John throws the ball”

“John threw the ball”

John *PROPEL*

BallO

BallOJohn *PROPEL*

P


Semantic NetworksConceptual Dependency theory

Advantages Provides a formal theory of natural language semantics reduces problems of ambiguity. representation directly captures much of the natural

language semantics sentences with similar meaning will have similar

representations (canonical form).Disadvantages:

No program exists that can reliably reduce sentences to canonical form.

Primitives not sufficient to represent more subtle concepts.


Framessupport the organization of knowledge into

more complex units reflecting the organization of objects in the domain.

Can be viewed as a static data structure with values attached.


Fig 7.12 Part of a frame description of a hotel room. “Specialization” indicates a pointer to a superclass.(Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009)

24

Frames Example


Frames AdvantagesFrames add power and clarity to semantic nets by

allowing complex objects to be represented as a single frame.

Frames provide an easier framework to organize information hierarchically than semantic nets.

Frames allow for procedural attachment which runs a demon (piece of code) as a result of another action in the KB (this has also been done to some semantic nets).

Both frames and semantic nets support class inheritance.


Assumptions in Knowledge RepresentationKnowledge must be represented internally.Knowledge representation should be done in

a central location (Knowledge Base).Human need to select precisely what

knowledge is to be represented.


Subsumption ArchitectureKnowledge must be represented

internally.Knowledge representation should be done in

a central location (Knowledge Base).Human need to select precisely what

knowledge is to be represented.

need not


Subsumption Architecture

“The world is its own Model”Rodney Brooks


Brooks on the Subsumption ArchitectureScene 13 from E. Morris’ “Fast Cheap and

Out of Control”


Subsumption Architecture“The world is its own Model”Intelligence is the product of the interaction

between an appropriately designed system and its environment.Examples of environments:

The real world for a robot, the internet for a web agent, the set of documents for a text understanding system, a game for a game playing system.

Intelligent behaviour emerges from the interactions of architectures that have organized simpler behaviour


Brooks on the Importance of Intelligent Systems’ Interaction with their environmentScene 15 from E. Morris’ “Fast Cheap and

Out of Control”



Each task handler is a finite state machine.A task handler uses a set of condition action

production rulestask handlers are data driven.

perception-based input Action

task handler



The architecture is a layered collection of task handlers.

Each layer subsumes lower ones.


From fig 7.26 The functions of the three-layered subsumption architecture from Brooks (1991a). The layers are described by the AVOID, WANDER, and EXPLORE behaviours. Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009

34

Subsumption ArchitectureFirst Layer: Avoid


Subsumption ArchitectureFirst Layer: AvoidWatch the avoid behaviour at

http://www.youtube.com/watch?v=ohykDN6-aY4


From fig 7.26 The functions of the three-layered subsumption architecture from Brooks (1991a). The layers are described by the AVOID, WANDER, and EXPLORE behaviours. Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009

36

Subsumption ArchitectureSecond Layer: Wander


Fig 7.26 The functions of the three-layered subsumption architecture from Brooks (1991a). The layers are described by the AVOID, WANDER, and EXPLORE behaviours. Luger: Artificial Intelligence, 6th edition. © Pearson Education Limited, 2009

37

Subsumption ArchitectureThird Layer: Explore


Distributed Problem SolvingMain idea:

No need for one central store of knowledge and general-purpose inferencing scheme

Divide a problem into several smaller problems.These smaller problems interact to solve the

bigger problems. Earlier history example: blackboard systems.

Intelligent behaviour is the result of the interaction between the appropriately designed problem-solving agent and the environment.


Agent-Oriented Problem SolvingAn agent is a problem solver that is:

Situated (interacts with its environment)Autonomous (makes its own decisions without

external intervention)Flexible (responds to stimuli from the

environment, and initiates actions based on situation).

Social (can interact appropriately with other agents or with humans).


Multi-Agent Problem SolversAgents interact to

cooperate towards achieving a common goal.coordinate in organizing the problem-solving

activity.negotiate sub-problem constraints to improve

performance.


Multi-Agent Problem SolversMulti-agent systems form a “loosely coupled

network of agents that work together” to achieve solutions to problems beyond the capabilities of any individual agent.


Agent-Oriented Problem SolvingExample: ROBOCUP“An international research and education

initiative. “Provides “a standard problem where wide

range of technologies can be integrated and examined.” (robocup.org)

Main domain: Soccer.Format:

Two teams of robots.Robots compete in a soccer match on a

standard platform.


Agent-Oriented Problem SolvingExample: ROBOCUPTeam members must be:

situatedautonomousflexiblesocial


Agent-Oriented Problem SolvingExample: ROBOCUPWatch this soccer game from the Humanoid

2008 World Cup at http://www.youtube.com/watch?v=iMM_XQXJUUc

An example of cooperation between robots during a soccer match

http://www.youtube.com/user/DarmstadtDribblers


Are agents Simply Objects with Fancy Stuff?Agents and objects (an instantiation of a class

in OOP) share some similarities but are quite different.


Objects in OOP vs. Agents Similarities: Objects (like agents) have

1. systems with encapsulated states.2. Certain methods are associated with the

object’s state.3. Methods support interaction with the

environment.4. Different objects communicate by message

passing.


Objects in OOP vs. Agents Differences:

1. Objects do not usually control their own behaviour.2. Agents can initiate their own actions. Object

generally do not.3. Objects do not have a social behaviour.4. Agents do not invoke methods in one another.5. Interacting agents usually have their own individual

thread of control.6. Agents can use more than just simple messages to

communicate. 7. Objects are associated with their class. Agents can

have multiple associations which may also change at any time.

8. Emergence can occur from groups of agents but not from objects.


Are agents Simply Objects with Fancy Stuff?Agents and objects (an instantiation of a class

in OOP) share some similarities but are quite different.

However, we can use objects to create agents.


SummaryDiscussed two models of AI problem solving

central store of knowledge: Weak problem solvers Strong problem solvers

No central KB Subsumption architecture . Multi-agent systems.

A combination of those, and other, models exist.

The applicability of one model versus others is affected by the problem at hand, resources, and problem constraints.


10 knowledge representation (compatible) -part 3

Documents

alhalimisemantic networkswell

represented knowledge

human knowledge

form of logic

form accessible

syntactic form

simonstrong problem

general problem solver