Knowledge Representation

1
Knowledge Representation
rule-based systems those that use the
knowledge structure if premise(s) then
conclusion(s) forward and backward chaining
are the common inference strategies logical
theorem proving is the general inference
paradigm - when KB is encoded as logical
rules using AND, OR, NOT, implication,
derive the logical truth or falsehood of some
expression using the KB as a theory
- theorem proving is an active area in AI
- theoretical result not all deductions are
decideable thus inference is a
fundamentally complex and undecideable problem
uncertainty also used Prolog form
conclusion - premise, premise, ...
conclusion.
2
Knowledge Representation
Problems with rule-based systems multiple
evaluation of rule (hence saving of user
input) large rule sets inefficient,
maintainability - 50 rule modules maximum
- if problem decomposition is difficult,
perhaps change the knowledge
representation (eg. add frames) uncertainty
choosing a technique, interpreting results
misfit of problems to rule-based paradigm
- not all problems naturally suited
procedural code in rules - forcing
procedural execution (eg. Oops) can be messy
- Prolog hooks are a good solution
3
Other techniques
i) Frames and object-oriented programming ii)
multiple contexts and truth-maintenance
systems iii) model-based representations iv)
blackboards v) Case-based reasoning (CBR)
Rule of thumb When a knowledge base is
difficult to understand and use, consider
changing the representation. Main goal close
the semantic gap between experts knowledge
and computer implementation
4
(i) Frames (review)
knowledge structuring technique "frames"
lmowledge representation exploiting structure
inheritence "object-oriented programming"
associating all data and computation with
objects ( frames) computation done via
message passing benefits - aids
structuring of KB - more generic rules,
therefore reduce size, complexity of KB -
puts data in one area/module (a frame
declaration) - maintainability of KB -
objects, once identified, can contain
characteristics, rather than spread this
knowledge throughout many rules - inference
engine can more effectively use/modify the data
5
Frames (cont)
To use frames during knowledge engineering
i) decide on type of reasoning (forward,
backward, other...) ii) decide on tasks
for rules and processing requirements -iii)
sketch a few rules, design a small hierarchy of
frames, and test AI workstations have
graphical frame interfaces - permits quick
definition, debugging Problems with frames
learning period for proper use and application
efficiency can be lots of execution time
overhead - eg. one object change can
entail 100's of subsidiary commands mis-use
create an inappropriate frame hierarchy/taxonomy
- knowledge becomes difficult to understand
- can also use too much procedural code, when
declarative would be preferred
6
Frames
p.250
7
(ii) Multiple contexts
context single universe/world in which
inference pertains some problems have
different environments which result in different
solutions eg. time, season,
path/direction, geographic location, ....
multiple context expert system performs
separate inferences for different contexts,
using the same ruleset - otherwise, one
inference and rule set to do the same
- complicated KB and rules, that need to
reconcile different contexts -
inefficient too much inferencing and
backtracking one can have multiple "parallel"
inferences for all the contexts then one can
inspect the results for each and determine which
is best alternative - each uses identical
KB rules, but may require their own environments
(variable binding records, run-time
databases, ....) akin to parallel processing
instead of backtracking
8
Multiple contexts
p.256
9
Multiple contexts
p.257
10
Multiple contexts
Applying multiple contexts 1. selecting
context structure for problem - problem
specific - often hierarchical structure
are well-suited 2. Reasoning with contexts
"fit" of context to problem lt---gt
reasoning (i) parallel reasoning mutually
exclusive sub-problems (ii) pruning trim
computation space (like a "cut") -
else resources can be overwhelmed -
dominance once one context is known to be best,
then kill others (iii) merging when 2
contexts have identical facts and assumptions,
incorporate them together
11
Truth maintenance systems
monotonic reasoning set of beliefs grow as
inference proceeds non-monotonic reasoning
adding contexts can reduce belief set
(eg. by retracting facts) Truth maintenance
systems (TMS) (consistency mgmt. systems)
- when one assumption changes, all conclusions
from that assumption must be retracted
(repeated) Three types of TMS i)
Justification TMS simplest - set of
beliefs in assumption set is monotonic, never
shrinks - facts are either believed or not
but not believing a fact is not the same
as believing it is false ii) Logic-based TMS
- each predicate is true, false, unknown
represented in logic - contradictions are
detected efficiently iii) Assumption-based TMS
- conexts used in parallel (as in previous
overheads)
12
TMS
p.266
13
TMS
p.267
14
Multiple contexts
Searching multiple contexts a tree of all
possible inferences (i) breadth-first each
level is generated one-by-one, fairly
- complete, but resource
expensive (ii) depth-first Prolog technique
- efficient, but possibly
non-terminating (iii) controlled more
arbitrary, can move about tree as
required "general theorem proving"
15
(iii) Model-based representation
encode knowledge in a structure which
naturally describes it use realtime monitoring
and diagnosis of machinery - can reason
about common and uncommon failures (i) Static
models graph, node arcs good for
representing connection paths between
components/modules/... supply input
forward chaiing supply output backward
chaining benefits - clarity can
permit visual representation - a
direct simulation of object of interest -
familiarity permits better debugging of KB
- maintenance change components in model,
rather than KB rules - simplification KB is
simpler - efficiency compact representation
of domain, smaller than with rules/facts
16
Static models
p.289
17
Static models
p.291
18
Model-based representation
(ii) dynamic models the state of the
model changes over time, and this state change is
the phenomemon of interest simulation
language/systems supply inputs, system
generates output, which is simulated dynamic
behavior of whole system used when actual
domain unavailable, therefore a formal model is
used as a substitute eg. nuclear
power plant, earth's ozone layer, aircraft
modelling, ... used in many AI systems (eg.
KEE) benefit can ascertain data that would
normally have to be treated as UNKNOWN
19
(iv) Blackboards
blackboard representation of knowledge with
a high-level controlling element first
used in Hearsay II components
knowledge sources (KS) modules with specific
expertise (tests, actions) blackboard
knowledge storage and communication
control overall problem-solving strategy
some control strategies - event-driven
react to events - expectation driven
predict solution based on current BB state, and
act appropriately - request driven
control given to the KS which seems appropriate
to a particular request - goal
directed control given to KS likely to solve
some goal
20
Blackboards
p.304
21
Blackboards

• Some conditions for BB use
• static knowledge sources KS's don't change
  structure (BB does)
• decomposable knowledge get a small number of
  complex KS's
• independence of KS's run independently without
  hooks to one another
• - BB is a medium that buffers them
• hierarchical problems layer problem domain
• - each level taken care of by one or
  more KS's

22
Blackboard characteristics

• Diverse approaches to problem solving
• modular, independent
• Common language for interaction
• formal interfaces
• Flexible representation of information
• Efficient storage and retrieval of information
• permits KSs to inspect BB, and retrieve info as
  necessary
• Organized participation
• a distributed computing problem
• Iterative problem solving
• KSs contribute to solution incrementally

23
Blackboards

• Some conditions for BB use
• static knowledge sources KS's don't change
  structure (BB does)
• decomposable knowledge get a small number of
  complex KS's
• independence of KS's run independently without
  hooks to one another
• BB is a medium that buffers them
• hierarchical problems layer problem domain
• each level taken care of by one or more KS's

24
Blackboards
p.315
25
Blackboards
When to use BB's real-time processing
(speach, signal processing) some scheduling
and planning applications problems which
naturally decompose into separate independent
tasks problems which are best handled
procedurally - this is also a pitfall, as
BB and procedures can be applied to problems
for which declarative rule-based methods
would be perfectly suited
26
(v) Case-based Reasoning (CBR)

• Case operational description of a past problem
  and its solution
• CBR idea experts use vast amount of past
  experiences to derive new solutions
• their compiled knowledge often uses pattern
  matching on long-past experiences
• CBR records and documents a history of past cases
• then attempts to match a given program with
  closest past case
• the more the CBR system is used, the larger its
  case history becomes, and the more effective it
  can be
• ideal when a problem is difficult to formulate in
  terms of explicit rule-based knowledge
• often, the closest case matching a new problem is
  sufficient for correctly diagnosing the problem
• Issue newer cases are usually more pertinent
  than older ones
• CBR system should account for case ages
• Note inductive inference (ID3) is one style of
  CBR
• each example in table is a case