Chapter 2 Slides

advertisement
Chapter 2: Case Studies of Several Case-Based Reasoners
•
•
•
This chapter describes some classic CBR systems that illustrate different
ways in which CBR can be used
One of the best known CBR systems is the planning system CHEF
• CHEF generates plans for cooking, that is, recipes
• It was built by Chris Hammond, in the AI Lab, at Yale
We will use CHEF to illustrate the flow of reasoning in CBR
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
1
CS 682, AI:Case-Based Reasoning,
Prof. Cindy Marling
2
CHEF
•
•
•
•
•
In a CBR system, the input describes the problem you want to solve or
the situation you want to interpret
Sample Input for CHEF:
• Request for stir-fry beef and broccoli recipe
Next we need to assign indexes
Indexes are features of the new problem that characterize it and that are
relevant to finding similar problems in the case base
Indexes for CHEF example:
• Stir-fry-dish
• Contains-beef
• Contains-broccoli
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
3
CS 682, AI:Case-Based Reasoning,
Prof. Cindy Marling
4
CHEF, continued
•
•
•
•
The input and indexes are used to retrieve a similar past case
The past case contains the prior solution
Matching rules are used to find the most similar, or most useful, case
CHEF wants a recipe for stir fry beef and broccoli. It retrieves:
Recipe for Stir-Fry
Beef and Green Beans
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
5
CHEF, continued
•
•
•
The next step is to adapt the old solution so that it will meet all of the
goals of the present problem
CHEF has a rule that tells it one vegetable can be substituted for another
in a recipe
It also has rules to account for differences in ingredients. For example,
broccoli needs to be chopped and cooks for less time than green beans
Recipe for Stir-Fry
Beef and Green Beans
Recipe for Stir-Fry
Beef and Broccoli
Cook time X
Cook time Y
Step: Chop broccoli
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
6
CS 682, AI:Case-Based Reasoning,
Prof. Cindy Marling
7
CHEF, continued
•
•
•
•
•
The next step is to test out the new plan or solution to see if it is
satisfactory
CHEF uses a simulator that checks a recipe against known criteria for
good stir-fry cooking
CHEF’s evaluation for this example:
• Failure: The broccoli is soggy
When a failure is detected, it needs to be explained, so that the system
will learn from it and not make the same mistake again
CHEF’s simulator knows that the broccoli is soggy because it sat too
long in the meat juices
• Explanation: meat juices to blame
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
8
CS 682, AI:Case-Based Reasoning,
Prof. Cindy Marling
9
CHEF, continued
•
•
•
•
The causal analysis guides the repair of the plan
Repair strategies are stored and indexed based on the types of problems
known to occur
CHEF’s strategy for soggy broccoli is called SPLIT-AND-REFORM
It repairs the recipe by adding in steps to stir fry the broccoli separately
and then mix it in
Recipe for Stir-Fry Beef and Broccoli
Cook time Y
Step: Chop broccoli
Step: Stir-fry broccoli in separate pan
Step: Mix broccoli in with other ingredients
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
10
CHEF, continued
•
•
•
•
•
Now CHEF has a good recipe to store in the case base
To avoid the soggy vegetable problem in the future, CHEF indexes this
recipe by
• Stir-fry-recipe
• Contains-beef
• Contains-broccoli
• Avoids soggy vegetable problem
CHEF improves its performance by checking for this problem in the
future
A CBR system learns by acquiring new cases
A CBR system learns by anticipating and avoiding failure
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
11
CS 682, AI:Case-Based Reasoning,
Prof. Cindy Marling
12
CASEY
•
•
•
•
CASEY diagnoses cardiac disease based on the diagnoses of earlier
patients
It was built by Phyllis Koton, at MIT
The features, or indexes, for a patient are measures and states
• Measures are clinical data, like heart rate
• States can be qualitative measurements, like high left arterial
pressure, or the presence of a disease, like pericarditis, or therapies
given, like nitroglycerin
To support a diagnosis, CASEY builds a causal explanation, in the form
of a graph, showing links between measures and states to indicate
causality
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
13
CASEY, continued
•
•
•
CASEY compares two cases by determining what the differences are
and whether or not they are significant
• An old case may have a feature that the new case does not have
• If it was not used in the diagnosis, then it is irrelevant whether or
not the new case has it
• If the feature was used in the diagnosis, then CASEY looks to
see if the new case has another feature that would cause the
same states as the missing feature. If so, the match is still good.
If all differences are irrelevant or can be “fixed up,” the cases are
considered to be similar, and the diagnosis is made
The diagnosis is adapted to to the new patient by substituting the new
causal relations and the new patient’s data
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
14
CASEY, continued
•
•
•
•
•
•
CASEY was built on top of the Heart Failure Program, a model based
system that was considered to be accurate, but slow
Whenver CASEY could not find a similar enough case in its case base to
make a diagnosis, it invoked the Heart Failure Program
CASEY obtained initial cases for its case base by running this system
CASEY also took knowledge about which features were and were not
significant from this system
CASEY’s diagnoses were found to match those of the Heart Failure
Program most of the time, but it operated one to two orders of magnitude
faster
Efficiency was CASEY’s original claim to fame, but it has since been
recognized for integrating reasoning modes and for illustrating CBR’s
usefulness in medical domains
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
15
JULIA
•
•
•
•
•
JULIA is a CBR design system
It designs menus for dinner parties, serving in the role of a virtual caterer
It was built by Tom Hinrichs, at Georgia Tech
The inputs to JULIA are the constraints that any satisfactory menu must
meet
Example:
(host tom)
(guests jlks-research-group)
(cost cheap-meal)
(ease-of-prep easy)
(ingredients (tomato cheese))
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
16
JULIA, continued
•
•
•
•
•
•
JULIA checks its case base for matching menus and interactively asks
questions to try to narrow down to a good menu
In our example, JULIA asks “Would you like lasagne, garlic bread and
red wine?
The user may agree, but if the user disagrees, JULIA will look for more
menu choices in the case base
The user may also add new constraints at any time, much as they would
in interacting with a human caterer
JULIA uses both cases (specific dinner party menus) and prototypes
(outlines of menus)
It also integrates CBR with constraint problem solving, another AI
approach applicable to the menu planning process
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
17
HYPO
•
•
•
•
•
•
HYPO is an example of interpretive CBR, rather than problem solving
CBR
It was built by Kevin Ashley, at U Mass Amherst
HYPO’s task is argumentation for legal reasoning in the domain of trade
secrets law
Trade secrets law cases typically involve employees of one company
giving or selling proprietary information to another company, giving that
company a competitive advantage
Trade secrets law cases are decided in courts of law based on precedents
found in previous cases
HYPO works by inputting a current case and finding “most-on-point”
cases to it from a case base of around 30 real trade secrets cases
• Cases are most-on-point if they share the most features in common
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
18
HYPO, continued
•
•
•
•
•
The features used in HYPO were those determined by legal scholars to
be important in deciding trade secrets cases
Example features:
• Had the accused employee signed a nondisclosure agreement
• Had the accused company bribed employees to switch employment
or to give them secrets
• Had the plaintiff company already disclosed the secrets to others
• Had the accused employee given the plaintiff’s notes, diagrams, or
tools to the accused company
HYPO retrieves cases that have both desired and undesired outcomes
It suggests hypothetical ways of strengthening a case by locating
additional information that would be helpful in getting a positive ruling
Lawyers then argue their cases based on the precedents HYPO provides
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
19
PROTOS
•
•
•
•
•
•
•
PROTOS is a CBR classifier
It was built by Ray Bareiss
It uses classification to diagnose an audiological patient as having one of
a number of known audiological disorders
PROTOS contains approximately 200 cases and 15 different disorders
It was built by having an expert audiologist classify real audiological
patients
To classify a new case, PROTOS uses reminding links to suggest a
potential classification
It compares the new case to the cases in that classification first
• If it finds a close match, it extends that classification to the new case
• If not, it uses difference links and censor links to suggest a new
classification to try
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
20
PROTOS, continued
•
•
There are other AI approaches to classification that are easier to use
when they are applicable
• These include neural networks and decision trees
Automated classification techniques did not work well in PROTOS’s
domain because of the nature of the data
• There were 58 different features a patient could have, but on
average, a given patient only had values recorded for 11 of these
• Neural networks and decision tree techniques do not deal well with
this much missing data
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
21
CLAVIER
•
•
•
•
•
•
CLAVIER is a design system that was fielded at Lockheed by D.
Hennessy and D. Hinkle
It was the first CBR system to be used on a factory shop floor (1990)
CLAVIER determines how to best load an autoclave, a large,
pressurized, convection oven
Lockheed manufactures parts for aerospace applications
One critical step is to cure these parts in an autoclave
• Lockheed engineers do not fully understand the properties of the
autoclave and believe them to be too complex to define using
equations or rules
Before CLAVIER, shop floor operators worked from sample oven
configurations
• They tried to fit the parts that were ordered into configurations
known to produce good results
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
22
CLAVIER’s Oven Configurations
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
23
CLAVIER, continued
•
•
•
CLAVIER began by using 20 of these manual configurations as cases
It input a prioritized list of ordered parts and looked for the best
matching configuration
• If multiple configurations could be used, the best match was the
configuration that used the highest number of high priority parts
• If no configuration could be completely filled out, the best match
was the one with the fewest missing parts
Initially, Hennessy and Hinkle tried to adapt incomplete configurations
by including parts similar to the missing ones
• This worked OK from a research perspective, but every mistake was
very expensive
• Lockheed engineers decided to just cure the missing (unordered)
parts and save them against future orders
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
24
CLAVIER, continued
•
•
•
•
CLAVIER became more successful as it acquired cases over time
Each new successful configuration was added to the case base until there
were 150 cases
At this point, they could find an exact match 90% of the time
• It grew close to becoming a table lookup system
Advantages of CLAVIER were:
• It used CBR outside the research lab
• It captured load configuration expertise, which could be used by
non-expert employees
• It provided working solutions even though engineers didn’t
understand the problem well enough to derive solutions from first
principles
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
25
Other CBR Systems
•
•
Not all CBR systems implement all phases of the CBR flow chart or
work in exactly the same way
One important twist is that CBR is useful for advisory and/or
educational systems
• These systems are typically less automated than the ones we’ve
discussed so far
• Often, just retrieving past examples for a user can help a person to
identify and/or solve a problem on their own
• A typical application is the CBR help desk, which works by storing
cases of reported problems that the help desk operators have already
solved. Since different users report the same problems over and
over, good solutions often already exist that can be recalled.
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
26
Download