280 likes | 536 Views
Knowledge Representation. Peggy Israel Doerschuk. Requirements. adequately reflect the types of knowledge needed allow new knowledge to be added and existing knowledge to be updated permit the derivation of new knowledge promote efficient processing of the information.
E N D
Knowledge Representation Peggy Israel Doerschuk Knowledge Rep
Knowledge Rep Requirements • adequately reflect the types of knowledge needed • allow new knowledge to be added and existing knowledge to be updated • permit the derivation of new knowledge • promote efficient processing of the information
Knowledge Rep Common representation schemes • Logical representation • predicate logic, propositional logic • Procedural representation • hard-coded sequential programs • production systems • network representation • graph representation - semantic nets, conceptual dependencies, conceptual graphs
Knowledge Rep Common representation schemescont. • relational representation • relational databases • knowledge represented by tuples or records • languages like Structured Query Language (SQL) used to manipulate data • hierarchical databases • allow links between related groups of data
Knowledge Rep Common schemes cont. • structured representation • frames, scripts, object databases, object-oriented programming languages • knowledge is inheritable • groups similar objects together • compact representation • allows reasoning at different levels of abstraction
Knowledge Rep Semantic Networks (Quillian) • Models human information storage and retrieval • association of concepts • hierarchical organization - info is stored at its most abstract level • canary is a type of bird; canary is yellow and can fly • flying is stored with bird • traits specific to canary (yellow) are stored with canary
Knowledge Rep Semantic Networks cont. • consists of nodes that represent an object, concept or event and arcs that represent a relationship between two nodes • nodes are represented as rectangles or circles • arcs are represented as directed arrows • Examples: p. 202 of Luger, p. 65 of Bigus, other examples in Rich • strength: inferencing via links, inheritance, flexibility • weakness: too unconstrained
Knowledge Rep Conceptual Dependency • Roger Schank (1974) • models the deep semantic structure of natural language • uses primitive conceptualizations to represent meaning • primitives define conceptual dependency relationships • conceptual dependency relationships are conceptual syntax rules • used to construct internal representation of English sentence • p. 206-210 of Luger
Knowledge Rep Scripts (Schank and Abelson) • used to represent common sequences of events • contains background information and a collection of slots used to describe the scenes • scenes are grouped into different tracks, depending on the particular situation • scripts are limited to common scenes and can't be used for novel situations
Knowledge Rep Components of a script • Entry conditions - must be true for script to be entered • results - true when script is exited • props • roles • scenes • ex: Fig 6.11
Knowledge Rep Frames • consists of a collection of slots (attributes) and fillers (values) associated with the object of the frame • slots can contain descriptive information (data), procedural information (functions), and pointer information (references to other frames) • supports inheritance and inferencing • frames are often linked to show has-a and is-a relationships • example p. 63 of Bigus, Fig 6.12 of Luger, other examples in Rich • frames can be represented as objects in OOP
Knowledge Rep Frames cont. • Let complex object be represented by a single frame • good for representing classes, inheritance, default values
Knowledge Rep Conceptual graphs John Sowa (1984) • two types of nodes in the graph • concepts (concrete or abstract)- boxes • relations - ellipses • arcs connect concepts to relations • each concept box has the name of the type and the individual, separated by : • markers are used to identify individuals • # followed by number • generic marker * marks unspecified individual • Ex: Fig 6.15-6.20
Knowledge Rep Operations on conceptual graphs • create a new graph by either specializing or generalizing an existing graph • copy • restrict - replace concept node with specialization • generic marker replaced by individual marker • type label replaced by subtype • join • simplify • Fig 6.22
Knowledge Rep Propositional nodes in conceptual graphs • Propositional concepts are indicated as a box that contains another conceptual graph • represent modal logics (various ways propositions are entertained - believed, asserted as true, false, possible, probable, etc.) • ex: Tom believes that Jane lines pizza. Fig 6.24, 6.25
Knowledge Rep Subsumpition Architecture • Rodney Brooks (1991) - intelligent behavior emerges from the interactions of architectures of organized simpler behaviors • subsumption architecture used for robot control • collection of task-handling behaviors • each behavior accomplished by a finite state machine that maps perceptions to actions
Knowledge Rep Three-layered subsumption architecture • Each layer has a network of FSMs • FSMs run asynchronously, sending and receiving messages • no central control; each FSM is driven by the messages it receives • Fig 6.26
Knowledge Rep Limitations of subsumption architecture • Myopic - each level sees only local info • no model of the complete environment means no ability to determine globally acceptable actions • no learning • can it scale to very large, complex systems?
Knowledge Rep Agent-Based and Distributed Problem Solving • Characteristics of intelligent agent system: • Situated - interacts with its environment • autonomous - acts independently • flexible - both responsive and proactive (goal directed) • social - interacts with other agents • communicate • bid for subtasks • cooperate, coordinate
Knowledge Rep Multi-agent problem solving • Problems are solved by multiple agents cooperating together, dividing and sharing knowledge of the problem • each agent has incomplete info • no global controller • knowledge is decentralized • reasoning processes are often asynchronous
Knowledge Rep Applications for agent-based problem solving • Manufacturing - modeled as hierarchy of work areas • automated control - transportation systems, air traffic control, etc. • telecommunications - network control, transmission and switching, etc. • transportation systems • information management - info filtering, gathering on the internet, etc. • ecommerce - portfolio management, etc. • interactive games
Knowledge Rep Knowledge Information Interchange (KIF) • Results from efforts of Defense Advanced Research Projects Agency Knowledge Sharing Environment workgroup • Designed to provide a common format for exchanging knowledge between agents • based on predicate logic, syntax similar to LISP • supports definition of objects, functions, relations, rules, and metaknowledge ( knowledge about knowledge)
Knowledge Rep Knowledge Information Interchange cont • a KIF knowledge base is a collection of forms • A form is either a sentence, a rule, or a definition
Knowledge Rep Knowledge Information Interchange cont. • Variables • individual variables begin with ?, sequence variables begin with @ • expressions • terms - objects; sentences - facts; definitions - constants; rules - inferencing steps (=> (EventName “AGENT:STARTING”)(SetIdentifiedIntervalAlarm “NETSCAPE” 20 “minutes”) If we get an AGENT:STARTING event, start an alarm called NETSCAPE to go off every 20 minutes. • operators • term, rule, sentence, definition operators • constants • numbers, characters, strings, objects, functions, relations, logical constants
Knowledge Rep Building a Knowledge Base • The symbolic approach: Knowledge engineer gathers knowledge from domain expert(s) and represents it in a form used by the reasoning system • expert must represent knowledge explicitly • knowledge acquisition bottleneck • the subsymbolic approach: expert networks use neural network to learn to perform classification and prediction tasks • knowledge is encoded in weights between neurons
Knowledge Rep Research areas in intelligent agents • How to decompose problem, synthesize results • interagent communication • how to ensure agents act coherently • coordination • resolving conflicts between agents • how to recognize, avoid chaotic behavior • how to allocate and manage resources • what are the best hardware, software platforms
Knowledge Rep Representing Uncertainty • Use statistical theory • probability of an event ranges from 0 to 1 • unconditional probability P(heads) = 0.5 • conditional probability is expressed as: • P(H|E) probability of hypothesis H given evidence E
Knowledge Rep Representing Uncertainty cont. • Bayes’ theorem: • P(Y|X) = P(X|Y)P(Y)/P(X) • Bayesian network • a directed acyclic graph • each node represents a variable and a conditional probability table defining relationships between parent nodes • uses probability to reason with uncertainty