Task Specific Knowledge Based Task Specific Knowledge Based Systems - - PowerPoint PPT Presentation

Task Specific Knowledge Based Task Specific Knowledge Based Systems and Their Application Systems and Their Application

Ahmed Kamel Ahmed Kamel North Dakota State University North Dakota State University

2

Outline Outline

• First Generation Expert Systems

First Generation Expert Systems

• Task Specific Architectures

Task Specific Architectures

• Generic Task Approach:

Generic Task Approach:

– – Structured Matching Structured Matching – – Hierarchical Classification Hierarchical Classification – – Routine Design Routine Design – – Functional Modeling Functional Modeling

• Applications:

Applications:

– – Engineering Design Engineering Design – – Agricultural Farm Management Agricultural Farm Management – – Adaptive Traffic Control Adaptive Traffic Control – – Computer Network Security Management Computer Network Security Management

• Conclusions

Conclusions

3

First Generation Expert Systems First Generation Expert Systems

• Focus is on representation

Focus is on representation

e.g. Rule e.g. Rule-

• based systems

based systems

• All knowledge represented in one form

All knowledge represented in one form

e.g. If e.g. If condition condition then then action action

• Control is separate from knowledge

Control is separate from knowledge

4

Strengths Strengths

• Uniformity of representation

Uniformity of representation

• Modularity of knowledge

Modularity of knowledge

• Change domains by changing rules

Change domains by changing rules

• Rules are easily expressed in English

Rules are easily expressed in English

• Rule trace provides explanation

Rule trace provides explanation facility facility

5

Limitations Limitations

• No guidance in problem analysis

No guidance in problem analysis

• No direct mechanism for expressing control

No direct mechanism for expressing control knowledge knowledge

• Problems with scale

Problems with scale-

• up

up

• Very hard to maintain:

Very hard to maintain:

rules are not really independent rules are not really independent

6

Lesson from First Generation Lesson from First Generation

It is difficult to separate knowledge It is difficult to separate knowledge from its use. from its use.

Knowledge Base Knowledge Base Inference Engine Inference Engine

7

Task Specific Architectures Task Specific Architectures

A task specific architecture A task specific architecture directly directly supports specification of the supports specification of the knowledge needed to carry out a knowledge needed to carry out a specified mapping of input to output. specified mapping of input to output. Examples of tasks include design, Examples of tasks include design, classification, ... classification, ...

8

Level of Generality Level of Generality

too general purpose, lose constraints on knowledge and inference too general purpose, lose constraints on knowledge and inference specialized specialized general purpose general purpose too specific, lose ability for reuse of the approach/shell too specific, lose ability for reuse of the approach/shell

task specific task specific middle ground middle ground

9

Task Specific Architectures Task Specific Architectures

• Several task

Several task-

• specific architectures exist to support

specific architectures exist to support the development of knowledge the development of knowledge-

• based systems

based systems

• Different architectures share a common

Different architectures share a common philosophy, however they differ on the level of philosophy, however they differ on the level of granularity of the tasks and on whether or not granularity of the tasks and on whether or not representation and use of knowledge should be representation and use of knowledge should be separate separate

• Furthermore, different architectures provide

Furthermore, different architectures provide mechanisms for the representation of knowledge mechanisms for the representation of knowledge

• r its use or both
• r its use or both

10

Generic Task Approach Generic Task Approach

Human reasoning is based on a set of generic

problem solving types out of which complex reasoning is composed

An attempt is made to identify these problem

solving types and to build software to support these types of problem solving

For each identified method, a a representation

template and a control regime are provided

The embodiment of which amounts to a language tailored for each specific GT that is identified.

11

Characteristics of a GT Characteristics of a GT

• Information Processing Task (i.e.

Information Processing Task (i.e. input/output relationship) input/output relationship)

• Organization of Knowledge

Organization of Knowledge

• Control principles

Control principles e.g. Hierarchical classification, Routine e.g. Hierarchical classification, Routine Design, Design, … …. .

12

Structured Matching Structured Matching

• A mechanism for imposing a structure on rules

A mechanism for imposing a structure on rules

• Organizes rules into small decision tables

Organizes rules into small decision tables

• Each decision table produces a result using some

Each decision table produces a result using some

• rdered metric:
• rdered metric:

– – 0 0 … … 5 5 – – Excellent, very good, good, neutral, poor Excellent, very good, good, neutral, poor – – strongly match, match, neutral, against, strongly against strongly match, match, neutral, against, strongly against – – etc. etc.

• Decision tables may refer to the results of other

Decision tables may refer to the results of other decision tables decision tables

13

Structured Matching Example Structured Matching Example

14

Hierarchical Classification Hierarchical Classification

• Knowledge Organization

Knowledge Organization

– – hierarchy hierarchy of specialists

• f specialists

– – each specialist is responsible for establishing a each specialist is responsible for establishing a hypothesis hypothesis – – a a subspecialist subspecialist represents a more detailed represents a more detailed hypothesis than its parent hypothesis than its parent

• Processing

Processing

– – establish establish the high level hypothesis (typically using the high level hypothesis (typically using a structured matcher) a structured matcher) – – Ask Ask subspecialists subspecialists to to refine refine to a more detailed to a more detailed hypothesis hypothesis

15

16

17

Classes of Design Classes of Design

• Creative Design

Creative Design

• Innovative Design

Innovative Design

• Routine Design

Routine Design

18

Routine Design Routine Design

A method for performing A method for performing “ “Routine Design Routine Design” ” pioneered by David Brown (WPI) for pioneered by David Brown (WPI) for performing design under repetitive well performing design under repetitive well understood situations understood situations

19

20

Routine Design Specialists Routine Design Specialists

Design Performed by a hierarchy of Design Performed by a hierarchy of cooperating specialists cooperating specialists

Specialist S1

Plan

Planning Decision Constraint Call Specialist S2

Plan

Planning Decision Constraint Call Specialist S2, S3

• • •

Specialist S2

Plan

Planning Decision Constraint

Plan

Planning Decision Constraint

• • •

Specialist S3

Plan

Planning Decision Constraint

Plan

Planning Decision Constraint

• • •

21

Routine Design Plans Routine Design Plans

• Plans represent a step by step method for

Plans represent a step by step method for performing the design performing the design

• A plan may include:

A plan may include:

– – the use of other specialists the use of other specialists – – tasks to be performed tasks to be performed

22

Plan Sponsors Plan Sponsors

• Each plan is associated with a sponsor

Each plan is associated with a sponsor

• A sponsor utilizes a structured matcher to

A sponsor utilizes a structured matcher to assign an applicability rating to its plan assign an applicability rating to its plan

• Applicability ratings may be:

Applicability ratings may be:

– – perfect perfect – – suitable suitable – – neutral neutral – – rule rule-

• out
• ut

23

Plan Selectors Plan Selectors

• Problem

Problem-

• solving agents available to

solving agents available to specialists specialists

• Use domain knowledge in consultation with

Use domain knowledge in consultation with plan sponsors to select a plan for execution plan sponsors to select a plan for execution

24

Design Tasks Design Tasks

• A task is an ordered collection of design

A task is an ordered collection of design steps steps

25

Design Steps Design Steps

• A step is the basic problem solving unit in

A step is the basic problem solving unit in routine design routine design

• A step utilizes a structured matcher to

A step utilizes a structured matcher to assign a value to a design attribute assign a value to a design attribute

26

Design Constraints Design Constraints

• A relation among design attributes that must

A relation among design attributes that must be maintained be maintained

• Constraints can be applied at any point in

Constraints can be applied at any point in the design process the design process

• Failure of a constraint triggers a failure

Failure of a constraint triggers a failure handler handler

27

Failure Handlers Failure Handlers

• Failure handlers may be:

Failure handlers may be:

– – automatic: systematic backing up to a automatic: systematic backing up to a branching point branching point – – knowledge knowledge-

• directed: use domain knowledge to

directed: use domain knowledge to “ “fix fix” ” the design the design

28

Functional Modeling Functional Modeling

• A Method for reasoning about the functional

A Method for reasoning about the functional behavior of systems (engineered or natural) behavior of systems (engineered or natural)

• Relies on hierarchically decomposing the system

Relies on hierarchically decomposing the system along its major functional components along its major functional components

• Functionality of any component is expressed in

Functionality of any component is expressed in terms of the functionality of its subcomponents terms of the functionality of its subcomponents

29

30

31

Device Functions Device Functions

• Functions of clothes Pin:

Functions of clothes Pin:

– – Open, close, hold Open, close, hold

• Functions of Pivot:

Functions of Pivot:

– – transmit_force transmit_force

32

Representing a Function Representing a Function

Function: Open Function: Open

• f device: Clothes_Pin
• f device: Clothes_Pin

toMake toMake: teeth_more_open : teeth_more_open provided: applied_force > restoring_force provided: applied_force > restoring_force by: open behavior by: open behavior

33

Behaviors Behaviors

• A behavior is a procedural implementation

A behavior is a procedural implementation

• f a function
• f a function

34

Applications Applications

36

Design of Composite Materials Design of Composite Materials

Domain of Problem: Domain of Problem: The material design (materials selection The material design (materials selection and processing protocol design) for and processing protocol design) for thermoset polymer composite materials. thermoset polymer composite materials.

Materials Design System Design Requirements Materials List, Processing Protocol

37

Design Architecture Design Architecture

Case Library Design Modifier Routine Designer Design Requirements No cases Design Case, Attributes to alter Design Modification failed Output Design Output Design

38

Problem: Multiple Designs Needed Problem: Multiple Designs Needed

• Market conditions (e.g. cost and

Market conditions (e.g. cost and availability) are very dynamic availability) are very dynamic

• needs vary (lower cost vs. faster turnaround

needs vary (lower cost vs. faster turnaround time) time)

• being a relatively new area, experimentation

being a relatively new area, experimentation with alternatives is desirable with alternatives is desirable A single design is not sufficient A single design is not sufficient

39

Multiple Design Solution Multiple Design Solution

MultipleDesigner Design Selector Family of Designs

Routine Designer

40

Multiple Design Multiple Design

• Allow selection of multiple plans

Allow selection of multiple plans

• Allow assignment of multiple (alternative)

Allow assignment of multiple (alternative) values for design attributes values for design attributes

• Output: a tree of design with each path from

Output: a tree of design with each path from the root to a leaf node representing a the root to a leaf node representing a complete design complete design

• Safeguards must be taken to guard against

Safeguards must be taken to guard against exponential growth exponential growth

41

Extensions to Routine Design Extensions to Routine Design

• During plan selection: rank the plans and select

During plan selection: rank the plans and select the highest ranked the highest ranked ones

• nes
• During step execution: select all possible values,

During step execution: select all possible values, not just the first one not just the first one

• Limiters at step and task levels to constrain the

Limiters at step and task levels to constrain the generated designs generated designs

• Redesign eliminated by explicitly specifying

Redesign eliminated by explicitly specifying possible redesign values in the main design possible redesign values in the main design knowledge knowledge-

• base

base

42

Agricultural Systems Agricultural Systems

Task: Task: The management of all aspects of an The management of all aspects of an irrigated wheat farm in Egypt irrigated wheat farm in Egypt

43

Top Level Architecture Top Level Architecture

Farmer’s Circumstance Farmer’s Circumstance Farmer’s Farmer’s Circumstances Circumstances Farmer’s Preferences Farmer’s Preferences Strategic Planning Module Strategic Planning Module season season Plan Plan Farmer’s Farmer’s Preferences Preferences

44

Strategic Planning Strategic Planning

Planting Date Preplant Tillage Fertilizer/Water Regime Planting Parameters Harvest Strategic Pest Management

Irrigated Wheat Crop Planning Task

Varietal Selection

45

Problem Solving Architecture Problem Solving Architecture

Planting Date Specialist Preplant Tillage Specialist Fertilizer/Water Regime Specialist Fertilizer/Water Regime Specialist Planting Parameters Specialist Harvest Specialist Strategic Pest Management Specialist Varietal Selection Specialist

Strategic Planning Module Strategic Planning Module

Routine Designer

Top Level Controller

Plan: Select Variety, Planting Date, Pest Management, Preplant Tillage, Planting Parameters, Fertilization/Water, Harvest

HC Module Algorithmic Module Routine Design Module

46

Lessons Learned Lessons Learned

• A complex problem solving situation

A complex problem solving situation should be decomposed into a set of should be decomposed into a set of simpler problems simpler problems

• Each of the individual problems can be

Each of the individual problems can be solved separately using an appropriate solved separately using an appropriate generic task method generic task method

• The individual problem solvers should be

The individual problem solvers should be integrated to accomplish the overall goal integrated to accomplish the overall goal

47

Weed Identification Weed Identification

• Weed Identification: looks like a

Weed Identification: looks like a classification problem but: classification problem but:

– – Difficult to describe weeds using text Difficult to describe weeds using text – – Difficult to understand description Difficult to understand description

• Solution: Picture

Solution: Picture-

• Based Hierarchical

Based Hierarchical Classification Tool Classification Tool

48

Picture Based Hierarchical Classification Picture Based Hierarchical Classification

• Hierarchical Classification with each

Hierarchical Classification with each specialist associated with a picture specialist associated with a picture

• Classification accomplished by the user

Classification accomplished by the user using the pictures using the pictures

• Motivation:

Motivation:

– – Some situations are difficult to describe using Some situations are difficult to describe using text text – – Describing pictures makes it hard to understand Describing pictures makes it hard to understand

49

Example Example

• Select either

Select either broad leaf broad leaf weeds or weeds or grasses grasses

50

Example (cont.) Example (cont.)

• For broad leaf

For broad leaf weeds, select weeds, select appropriate appropriate shape for shape for cotyledon cotyledon

51

Example (cont.) Example (cont.)

• Select

Select appropriate appropriate true leaf true leaf

52

Example (cont.) Example (cont.)

• Select the

Select the matching matching seedling seedling

53

Example (cont.) Example (cont.)

• The

The flowering flowering plant plant

54

Applications Under Development Applications Under Development

• Adaptive Traffic Signal Control

Adaptive Traffic Signal Control

• Computer Network Security Management

Computer Network Security Management

55

Traffic Signal Control Traffic Signal Control

• Simulation

Simulation-

• based optimizers are used to optimized

based optimizers are used to optimized the control of traffic signals under the control of traffic signals under “ “normal normal” ” traffic traffic conditions conditions

• Normal traffic conditions may include periodic

Normal traffic conditions may include periodic variations such as weekday morning and afternoon variations such as weekday morning and afternoon rush hours rush hours

• What happens when an exception occurs, such as:

What happens when an exception occurs, such as:

– – An accident An accident – – The end of an even or a popular game at a stadium The end of an even or a popular game at a stadium – – … …etc. etc.

56

Adaptive Traffic Signal Control Adaptive Traffic Signal Control

• Using monitoring devices, traffic volumes at all

Using monitoring devices, traffic volumes at all directions are collected directions are collected

• Traffic volumes are fed into a hierarchical

Traffic volumes are fed into a hierarchical classification system classification system

• The observed pattern is classified as one of several

The observed pattern is classified as one of several pre pre-

• compiled patterns described by various

compiled patterns described by various increments at the different directions (a null increments at the different directions (a null pattern is allowed to allow for no corrective pattern is allowed to allow for no corrective actions to be taken under actions to be taken under “ “normal normal” ” conditions) conditions)

57

Corrective Actions Corrective Actions

• When a deviation from normal is detected, a

When a deviation from normal is detected, a routine planner is invoked to modify the current routine planner is invoked to modify the current signal pattern for the current traffic signal, and signal pattern for the current traffic signal, and those downstream those downstream

• Corrective actions may include modifications to:

Corrective actions may include modifications to:

– – Signal durations Signal durations – – Signal splits Signal splits – – Phase shifts relative to upstream signals Phase shifts relative to upstream signals

• Routine Planners are designed using simulations

Routine Planners are designed using simulations

• n a large number of systematically generated
• n a large number of systematically generated

traffic patterns traffic patterns

58

Computer Network Security Management Computer Network Security Management

• Task: given a new

Task: given a new “ “sensitive sensitive” ” project, project, assign varying degrees of access to the assign varying degrees of access to the project files to the different project project files to the different project personnel personnel

• A hierarchical classification system based

A hierarchical classification system based

• n project sensitivity level, job
• n project sensitivity level, job

classification, and role on project classification, and role on project

59

Conclusions Conclusions

• Extensions are often needed and

Extensions are often needed and implemented as new problems are tackled implemented as new problems are tackled (Will we ever reach a steady state?) (Will we ever reach a steady state?)

• Complex problems requires collaborative

Complex problems requires collaborative effort of more than one problem solver effort of more than one problem solver

• What

What’ ’s next? s next?

– – Educational Tutorials? Educational Tutorials? – – Distributed collaboration? Distributed collaboration? – – Intelligent Agents? Intelligent Agents?