18 Representing commonsense knowledge

1
18 Representing commonsense knowledge

• The commonsense world
• what is commonsense knowledge
• difficulties in representing
• importance, research areas
• Time
• KR by networks
• different data structures
• non-monotonic reasoning

2
18.1 Commonsense world

• Expert systems toy examples
• limited knowledge
• well-defined, circumscribed domains
• Human-level AI systems?
• Certain topics difficult to conceptualize
  formally
• easy for humans is hard for computers (and vice
  versa)?

3
Commonsense knowledge

• Things a 10-year old knows
• adequate for many things we do
• scientific knowledge required when we need more
  precise descriptions
• Basic functionality of a robot
• rains -gt shelter
• run, stop -gt dont spill coffee
• pay bills in time
• worth formalizing!

4
Naïve knowledge?

• Good for theories of naïve sciences
• basic physics, economics,
• not contradictory to expert knowledge
• regions of a continuum
• different tasks, different expertise
• Formalizing c.s. knowledge
• broad area near the naïve end

5
18.1.2 Difficulties

• Sheer bulk
• ES 100-1000 facts/rules
• CYC around 10.000.000
• No well-defined frontiers
• conceptualizations thorough the whole area
• can not test until its ready

6
Difficulties...

• Some knowledge isnt easily captured by
  declarative sentences
• complex physical objects
• hard to describe in natural language
• mountain view, sunset, ...
• logical description of the concept?

7
Difficulties...

• Many sentences are actually approximations
• Birds fly (? x) Bird(x) -gt Flies(x)
• How to say (compactly formally)
• most birds fly
• unless stated otherwise, if x is a bird then it
  flies

8
Difficulties...

• How to conceptualize
• time discrete or continuous
• past, future(s)
• intention
• Examples
• If I hadnt done x I hadnt done y
• It was not my intention to do x

9
18.1.3 Importance

• Applications
• are most experts systems?
• commonsense applications
• household robot
• clean, laundry, food, household,
• Makes also ESs more useful
• is question in/outside my area?
• Is my expertise relevant to task?

10
Importance...

• Expanding ES capabilities
• analogies, metaphors
• in particular spatial metaphors (above, beyond,
  next)
• metaphors not simply linguistic coincidences?
• Human way of conceptualizing?
• NLP applications

11
18.1.4 Research areas

• Objects materials
• discrete, solid things easy
• hierarchical, fluids, collections
• Space
• relation to other objects, size, shape
• Physical properties
• mass, temperature, volume,...

12
Research areas...

• Physical processes events
• falling, throwing,
• physics differential equations
• AI qualitative physics
• Time
• temporal logics
• AI situations (snapshots) linked by actions
• also an object to be reasoned about

13
18.2 Time

• How to think about it?
• Real line (-infinity...infinity)?
• Countable integers (big bang,1,2, )
• circular?
• Picture used most often in AI
• events occur in time
• event containers intervals
• intervals entities that exist

14
Time intervals

• Description
• name, events Occurs(E,I)
• begin end start(I) end(I)
• Fact start(x) lt end(x)
• Basic relations
• Meets(x,y), Before(x,y), (fig 18.1)
• transitivity, ...
• Use to express c.s. facts

15
18.3 KR by networks

• Taxonomic knowledge
• Semantic networks
• Non-mon reasoning
• Frames

16
18.3.1 Taxonomic knowledge

• Many entities can be arranged in hierarchical
  structures
• organizing, simplifying
• CYC Thing
• concrete/abstract objects, processes,
• X is a P, all P are Q, all Q are R,
• encoded as e.g. semantic networks or frames

17
Example

• Office machines
• LP(S), (?x) (LP(x) -gt P(x))
• (?x) (P(x) -gt OM(x))
• Categories of objects (LP, P, OM)
• Transitivity
• Properties of categories
• expressed with functions equality
• (?x) (OM(x) -gt e_s(x) w_o)
• inheritance!

18
18.3.2 Semantic Networks

• Graph structures for knowledge
• Two kinds of node labels
• relations categories and properties
• objects
• Three kinds of arcs
• subset arcs (isa links)
• membership arcs (instance links)
• function arcs

19
Semantic networks...

• Reasoning about
• (inherited) category properties and
• set memberships
• is more efficient than using (unguided)
  resolution
• just follow instance/isa arcs

20
18.3.3 Non-mon reasoning

• Recall monotonic not
• Default inferences are common
• unless known otherwise, do
• once otherwise is known, reasoning must be
  revised
• Many formalisms/tools studied
• TMSs, default logic, autoepistemic logic,
  nonmonotonic logic, ...

21
Cancellation of inheritance

• Simple non-mon reas. Method
• Example
• energy source is by default wall outlet
• robots use batteries, however
• contradiction?
• Use network from specific to general direction
• Multiple inheritance difficulties
• priorities to default values?

22
18.3.4 Frames

• Record-like structure
• name
• set of attribute-value pairs (slots)
• slot names slot fillers
• metaknowledge (fig. 18.5)
• Variant of sem. Network
• name node
• attributes names of arcs
• values objects

23
Frames...

• Some knowledge hard to express
• disjunctions, negations
• nontaxonomic knowledge in general
• Hybrid systems
• terminological logics
• hierarchical structures
• logical expressions

24
18.4 Discussion

• Property inheritance mechanism has been adopted
  by O-O programming languages
• Closed-world assumption (CWA)
• Knowledge sharing
• Human commonsense reasoning
• analogical reasoning
• case-based reasoning