Designing Backward Chaining Systems

Major tasks when developing a backward chaining system:

1. Define the problem

2. Define the goals

3. Define the goal rules

4. Expand the system

5. Refine the system

6. Design the interface

7. Evaluate the system

Example (Financial Planning)

• A popular areas of Expert System applications:

  • the system offers real payoffs

  • the task is well-defined

  • real expert exists

  • many successful expert systems exist in financial counseling.

• Objective: to develop an expert system to assist a financial counselor with providing an investment recommendation for a client.

Task1: Defining the Problem

• source for starting the project: reports, documents and books.

• Identify an expert 0 e.g. an investment advisor at a local firm.

• Listen to the expert how he/she do her job.

• During the problem definition stage of a project, do not interrupt the expert for problem details. Here you are looking for a general understanding of the problem and how it is solved. Details can be obtained during later discussion with the expert.

2.Defining the Goals

• The coding of any backward-chaining system begins with defining the system's goal.

• From the discussion with experts, obtain two principal goal:

  • Determine portfolio mixture

  • Determine investment vehicles within each category.

  Only the first goal will be discussed.

• For simplicity, assume that we consider only 4 kinds of portfolio:

  • Portfolio 1: 100% investment in savings

  • Portfolio 2: 60% stocks, 30% bonds, 10% savings

  • Portfolio 3: 20% stocks, 40% bonds, 40% savings

  • Portfolio 4: 100% investment in stocks

• [Note: should focus the design of the initial system on a small but representative part of the overall problem]

• We have 4 different goal, one for each portfolio, for the backward system. Can write 4 different goal, or 1 variable goal.
  

Task3: Designing the Goal Rules

• Each goal in the system must have a least one rule that can conclude it.

  

• Rule: IF Precondition_1 AND Precondition_2 AND ...
  THEN Portfolio_i

• Talk to domain expert about the process. Find out the factors that the expert use to make decision.

• Useful to construct a decision table (if possible).

• Use this table to write up some rules:
  IF Client's_investment_amount < 10000
  THEN Portfolio advice IS 100% investment in savings

  IF Client's personal state suggests a conservative position
  AND Client's financial state suggests a conservative position
  THEN Portfolio advice IS 100% investment in saving.

  IF Client's personal state suggests a conservative position
  AND Client's financial state suggests an aggressive position
  THEN Portfolio advice IS 60% stocks, 30% bonds and 10% savings

  IF Client's personal state suggests an aggressive position
  AND Client's financial state suggests a conservative position
  THEN Portfolio advice IS 20% stocks, 40% bonds and 40% savings

  IF Client's personal state suggests an aggressive position
  AND Client's financial state suggests an aggressive position
  THEN Portfolio advice IS 100% investment in stocks

• You can encode these new knowledge into an expert system, using a shell.

• Following the encoding of any new knowledge into the epxert system, you should immediately test it, using information specific to the new knowledge. (particularly important during the early stage of the project).

• It is also possible to draw the inference network (or the AND-OR graph) for the rules that has been develop for easy referencing.

• Order of search is also important. Usually, the system will explore the rules in a specific order. We can re-arrange the rules to change the order of search to match with the DE.

• You also need to consider exhaustive search vs. non-exhaustive search. Whether to stop when any one solution found, or continue to find all possible solution.

• The latter is particularly important when the system use numerical certainty factor (or belief factor).

• e.g. Each conservativeness/aggressiveness can have a certainty factor between 0 and 1, and the result also characterized by a certainty factor:
  
  PORTFOLIO RECOMMENDATION BELIEF
  60% stocks, 30% bonds, 10% saving 90%
  100% investment in stocks 85%
  20% stock, 40% bonds, 40% savings 70%
  100% investment in savings 60%

Task4: Expanding the System

• You can broaden or deepen the system's knowledge.

• Broadening 0 teach it about additional issues.

  • E.g. teaching it about additional portfolio mixtures.

  • Fairy easy and usually reserved for later in the project.

• Deepening 0 teach it more about issues it already knows about.

  • E.g. teach the system how to determine the present premises in our goal rules.

  • During the entire development of an expert system, one should be looking for premises that can be expanded deeper.

  • Ask whether the user can effectively answer the question posed by the system. If no, then add knowledge to the system that forces it to seek more primitive and reliable information. (e.g. quantitative answers).

• The system has 3 issues that can be further expanded:

  • Client's investment amount

  • Client's personal state

  • Client's financial state

• Repeat the above procedure, i.e. talk to DE, derive rules, and deepening of premises ....

• Expand one issue at a time. This permits the expert to focus on a single issue and avoid problems commonly encountered when the expert is asked to address multiple issues.

• E.g.

  IF Client IS old
  OR Client's job IS not steady
  THEN Client's personal state suggests a conservative position

  IF Client IS young
  AND Client's job is steady
  AND Client's has children
  THEN Client's personal state suggests a conservative position.

• Avoid using ELSE in rules.
  E.g. IF Client_age < 40
  THEN Client IS young
  ELSE Client IS old

• and later if one decide to add a middle-age concept:
  IF Client_age < 50
  AND Client_age >= 40
  THEN Client IS middle-age

  then problem will arises.

• The resultant personal state and financial state network may look like:

• The use of inference network or AND-OR graph allows easy identification of concepts that can/need to be deepened.

Task5: Refining the System

• we now have a fully functional system that meets our initial objectives.

• Can use variables for numbers to make modification simpler.
  E.g. can use the variable OLD_AGE to store the age which is considered old.
  Initialize OLD_AGE=40

  Rule:
  IF Client_age < OLD_AGE
  THEN Client IS young...

• provide safety net for the system û a course of action that the system takes if it fails to reach a recommendation.

  • Can be as complex as loading in anther knowledge base that will try to correct the situation.

  • Or simply displaying a message that recommendation not found.

  • Or provide an "intelligent safety net", such as why recommendation not found.

• Providing Milestones

  • worthwhile to keep the user informed of the important finding during consultation.

  • E.g. after finding the personal state of the user, display it.

Task6: Interface Design

• Each system should have an introductory display

  • at least tell the user about the overall purpose of the system

  • can also tell the user a little about how the primary task is accomplished.
    E.g.

  
  

• When asking users to answer a question, details should be provided as what is expected from the user, possibly with an example.

• Better to provide multiple choice questions if possible.

• Conclusion display (final recommendation if found).

Task 7: System Evaluation

• ask experts for one of his past cases and compare the result with the recommendation of the system.

• Test with more cases.

Future Direction

• broadening

  • recommending additional portfolios

  • recommending investment vehicles.

Example Session:

SYSTEM: Welcome to ....
Please press RETURN to begin our session

USER: RETURN

SYSTEM: How much money in dollars does the client want to invest? Please enter the dollar amount without commas, e.g. 10000.

USER: 50000

SYSTEM: What is client's age

USER: 30

SYSTEM: How many years has the client been with the present company?

USER: 5

SYSTEM: Layoffs at the organization are:
Low
High

USER: Low

[The user may have trouble with this question. May need deepening]

SYSTEM: The client has
children
No children

USER: children

SYSTEM: After considering the client's personal issues, I would suggest a conservative position. I will next look into the client's financial issues.
Please press RETURN to continue the session.

USER: RETURN

...

SYSTEM: I would recommend the client invest the money in
60% stocks, 30% bonds, 10% savings.
The major reasons why I recommend this investment is that I found the client's personal state suggests a conservative position and financial state suggests an aggressive position.



Designing Forward-Chaining Systems

Major tasks in designing a forward chaining system:

1. Define the problem

2. Define input data

3. Define data-driven structure

4. Write initial code

5. Test the system

6. Design the interface

7. Expand the system

8. Evaluate the system

Example: (Automobile Diagnostics Domain)

• problem is better understood than problems such as design or scheduling.

• Provide tangible benefits.

• Most diagnostic problem use backward chaining approach, as most of the problem addressed had a finite number of possible faults. Each fault treated as a goal.

• However, in many diagnostic problems, the number of possible fault conditions can be very large.

• E.g. in automobile diagnostics, there are hundreds of possible fault condition, impractical to treat each fault as a goal.

• The expert usually begins by gathering some initial information, reasons with it to pursue other important information, and finally, identifies the faulted automobile part.

• Inherently data-driven and is best managed with a forward-chaining expert system.

Task1: Define the problem

• information is obtained from the repair manual.

• The manual usually provide a table or decision-tree to aid in the isolation of the faulty subsystem.

• E.g. the problem "engine does not start"

Section 1

Starting Problem

...

Task2: Define Input Data

• every forward system needs to first obtain some initial data to get started.

• Need to write a rule whose only task is to ask for some information about the problem û startup rules.

• This rule is written to automatically fire when the system starts. When it fires, it asks fro problem information.

• E.g.
  Rule 1 Start Diagnosis
  IF Task IS Begin
  THEN ASK Car question.

• Ask the user to select from a list of car problems, such as
  Car won't start
  Car hesitates at high speeds
  Car idles rough

Task3: Define Data-driven Structure

• Modify the forward structure to include task information

• e.g.
  IF Task is test battery condition
  AND Lights don't brighten
  THEN Task is test battery

• Using a rule of this form, the rule will only fire if the rule is relevant to the current task.

• This is similar to grouping the rules.

Task4: Write Initial Code

• the purpose is to determine if we have effectively captured the problem's knowledge in a good rule structure.

• A good structure is one that not only provides correct results, but also a template to follow in the development of other rules. (e.g. the rules structure described above).

• Example:
  Rule 1 Start Diagnosis
  IF Task IS begin
  THEN ASK Car Problem

  Rule 2 Car won't start
  IF Car problem IS won't start
  THEN Task IS test cranking system

  Rule 3 Car hesitates at high speed
  IF Car problem IS hesitate at high speeds
  THEN Task IS test fuel system.

  Rule 4 Step 1.1 - test the cranking system
  IF Task IS test cranking system
  THEN ASK Engine turns
  ...

Task5: Test the System

• Assume that "Task IS begin" is initialized into the WM by the system.

• A typical dialogue looks like:

SYSTEM: What is the problem with the car?
- won't start
- hesitates at high speed ...

USER: won't start

SYSTEM: Please turn on the ignition.
How does the engine turn?

- Slowly or not at all
- normally

USER: not at all

...

Task6: Design the Interface

• begin the development of the system's interface early in the project and develop it in parallel with the development of the system's knowledge base.

• For example, use of graphical interface familiar to the user, say, one that is similar to the car front panel.

• You should bear in mind that the system will most likely be used by someone not familiar with a computer.

Task7: Expand the System

• include more faults into the system.

Task8: Evaluate the System

• Work directly with the manual for initial testing

• Work with expert

• Finally, real test cases.