RuleML 08

1
RuleML 08

• Adding Uncertainty to a RETE-OO Rule Engine

Authors D.SottaraM. ProctorP. Mello
Speaker D. Sottara
2
Objectives (long term)

• Intelligent supervision of complex Systems
• Example bio-chemical processes (water treatment)
• Tasks for a Knowledge-Based system
• Monitoring
• Control
• Diagnostics
• Available Resources
• Real-Time Data from probes
• Knowledge (partial) from human experts
• Chosen Architecture
• Rule Based Expert System
• Convenient to encode knowledge
• Models expert behaviour
• Possibly Interactive
• Uncertainty in both Data and Knowledge

3
Rule-Based Expert System

• Standard architecture

• Rules
• Link Premises to Conclusions
• If P then C
• When P then C
• P ? C
• ltHeadgtClt/Headgtlt/BodygtPlt/Bodygt

• Engine
• Preconditions trigger Consequences
• RETE
• RETE-OO (Drools)

Engine
4
Example

• When
• t Tank( temperature 25 level lt 80 )
• s Sample( source t (pH lt 7 orp gt 100))
• Then
• / sub-optimal conditions /

5
Example / with Uncertainty

• When
• t Tank (
• temperature 25
• and level is not full
• )
• and
• s Sample (
• source t and
• (pH lt 7
• or orp gt 100 )
• )
• Then
• / conditions are inadequate in some degree /

Uncertainty should be handled automatically
6
Benefits of Uncertainty

• Robustness
• Input may not be accurate
• What if tank temperature is not available?
• Threshold values are arbitrary
• Is 7.8 different from 7.7?
• and may change
• Convenience
• Expert Knowledge is rarely precise
• Rules may not be always valid
• Unexpected Exceptions

7
Types of Uncertainty

• Uncertainty comes in different forms
• The most commonly used are
• Reliability (Confidence)
• Eg T 25 (70/100)
• Measures the strength of the estimate,
  especially w.r.t. to alternatives
• Aleatority (Probability, prior)
• Both subjective (bayesian) and objective
  (frequentist)
• Eg T 25 (70)
• 30 probability the value is different. May be
  completely different
• Graduality (Fuzzy, posterior)
• Eg T 25 (0.7)
• The value, e.g. 22, is known to be similar to
  the reference 25.

8
Truth Degrees

• The degree of uncertainty can be modelled in
  different ways
• Does not depend (entirely) on the type of
  uncertainty
• More complicated representations may account for
  several types at the same time

• Real Value 0,1
• Interval 0,12
• Distribution 0,1 ? 0,1
• Type II fuzzy set
• Imprecise Distribution 0,1 ? 0,13
• Type III fuzzy set

9
Evaluators

• Given some arguments X and a property P
• P(X) reads X are P?
• E.g. Tank.temperature, 25 are Equal
• P(X) generalizes the concept of characteristic
  function
• P symbolic name
• X arguments
• Needs a definition (semantics)
• Output Truth Degree
• An Evaluator is any implementation of the
  interface P(X)
• Mathematical function
• Map
• Neural network
• Logic program
• Should be invoked at run-time

10
Operators

• Operators are special Evaluators
• Aggregate properties
• E.g. n-ary or checks whether any of its
  argument properties hold
• Truth-functionality
• Output depends only on the output of the
  individual args
• Usually requires independence
• Standard families exist for different logics
• Language support is needed
• Custom operators may be defined and used
• Does not depend (entirely) on the type of truth
  degrees

11

• when Rain (hard)
• then Run (fast)

12
State of the Art

• Flexible languages and tools are needed
• Expressiveness
• Customizability
• Fuzzy-RuleML has been created with similar goals
  at the language level
• Built for narrow fuzzy logic
• Compatible with other standards
• No equivalent mainstream, open-source engine
• Most are RETE-based
• No unified support for uncertainty
• Special purpose engines (e.g. fuzzy logic) are
  common

13
Example

• t Tank( temperature 25 ? level ? is full )?
• s Sample( source t ? (pH lt 7?? orp gt 100
  ))

a
b
a
14
RETEU a-network

• Constraint nodes enclose modular evaluators
• Simple , ! , gt ,
• Complex (custom) contains, high,
• Partial matching type-checking
• Explicit Operator nodes are added
• Nested operators supported
• Uncertainty is optional ? Cast to boolean

?
?
a
?
a
15
RETEU b-network

• Join nodes use Evaluators for constraints
• An operator is applied at the end of the b chain

• Problem
• When to filter an object / tuple ?
• Discard on false no longer applies
• Different Strategies
• Never discard
• False rules still fire
• Performance issues
• Discard on most-likely false
• Also on custom threshold
• Heuristics and many others

Identity?Equality?
16
RETEU Terminal Node

• In classical Modus Ponens
• ltP(x), P(X) ? C(Y)gt / C(y)
• both Premise and Implication are taken into
  account
• Rules have an associated truth degree
• Confidence factors
• Associative rules (support, coverage)
• Gradual rules
• Can be learned
• Using the Implication operator ?
• Terminal nodes yield the truth degree of the
  Premise
• To compute the degree of activation, the
  Implication degree must be considered as well.
• Using the Deduction operator ?
• If a rule is True, it coincides with the Premise

17
Terminal Network

• Implication (rule) truth degree is usually a fact
• May be stored in a-memory
• May be asserted dynamically
• Premise is joined to Implication
• Conclusion is entailed
• Consequence depends on the activation degree
• Uncertain Facts may be asserted

18
RETEU Multiple Evaluation

• Information on a constraint comes from different
  sources
• (possibly at different times due to rule
  chaining)
• Facts a priori information
• Evaluator in the a-node
• Rules other parts of the network
• Equal to the activation degree
• Degrees are merged ?
• Think of intervals
• Propagate again on actual updates
• and cached
• an object may be stored in several memories
• Need conflict resolution strategies
• Backward chaining may help ?

19
RETEU State vs Justify

• TMS distinguish between facts
• Stated provided externally
• Justified entailed logically
• Justifications add support and (usually)
  information
• Two different insertions
• State previous information is overridden
• Evaluators are disabled
• Justifications are ignored
• Justify previous information is combined

20
Conclusions and Future Works

• RETE has to be modified to support Uncertainty
• Few structural modifications
• Many policy modifications
• Node behaviour
• Interactions between nodes
• Goal support different logics
• Customizability many orthogonal degrees of
  freedom
• Evaluators
• Truth degrees
• Operator types
• More inference mechanisms
• Quantifiers
• Learning by Induction
• Drawback Risk of Incoherence
• Use high level packages

21
Prototype applications

• Several experiments are being carried out
• Standard Fuzzy controller
• Real Time Signal Analysis
• Expert Committee using different evaluators
• Statistical
• Neural
• Logical
• Declarative programming
• Emulation of the Self-Organizing Map training
  algorithm