This is a test - Ohio University

Download Report

Transcript This is a test - Ohio University

Chapter 3: Reasoning Using Cases
•
•
•
In this chapter, we look at how cases are used to reason
We’ve already seen that there are two main types of CBR
• Problem solving: planning (CHEF), design (JULIA), diagnosis
(CASEY)
• Interpretation: understanding, justification and projection (HYPO)
Both types of CBR may be used in the same system
• MEDIATOR was an early CBR system that solved disputes
• It was a problem solving system in that it generated plans to end
disputes
• It was interpretive in that it had to understand the reasons behind a
dispute in order to propose a good compromise
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
1
Problem Solving and Interpretive Tasks Addressed by CBR:
Planning
•
•
Planning is the process of generating a sequence of steps for achieving
some desired state
Planning is a difficult task
• The order of steps is important
• You need to project the consequences of executing each step
• You need to be sure the preconditions for a step to succeed are
met
• Checking for interactions with traditional planning techniques is
exponential in the number of steps in a plan
• Things don’t always go according to plan
• You may need to replan quickly if things go wrong
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
2
Planning, continued
•
•
•
CBR helps by storing complete plans that already have the step
interactions worked out
Plans are normally stored so that parts of plans, as well as whole plans,
can be accessed
• This facilitates quickly changing plans, when needed
Examples of CBR planning systems include:
• CHEF
• A system used by the Italian Forest Service to plan the management
of large forest fires
• A system used by the U.S. Navy to plan the evacuation of civilians
who get caught in the middle of dangerous situations
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
3
Design
•
•
Design is the process of describing some concrete object that satisfies a
set of constraints
• Note that nothing in the design tells you how to GET the desired
object. For that, you would need a plan.
In design, problems may be underconstrained or overconstrained
• An underconstrained problem has few constraints and many
possible solutions
• If JULIA needs to find a menu that’s cheap and easy to prepare,
there are hundreds of possibilities. That’s fine for a human
designer, but doesn’t give a system much to go on
• An overconstrained problem has constraints that conflict with each
other
• There may be no solution at all that meets all of the constraints.
Again, humans may be able to deal with this, but a system can
not find a solution in this situation
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
4
Design, continued
•
•
The underconstrained case is more common, found, for example, in
designing buildings and cars and menus
• Here, CBR helps by providing examples of good solutions
• For example, a car designer considers past models in designing
next year’s models
CBR systems have been used to design menus (JULIA), autoclave oven
layouts (CLAVIER), buildings, landscapes, and mechanical devices
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
5
Explanation and Diagnosis
•
•
•
•
•
In explanation, we find reasons why something happened
• This is sometimes called the credit assignment or the blame assignment
problem
Diagnosis is the most common type of explanation application
The input to a diagnostic system is a list of symptoms or problems, and the
output is an explanation for these problems
CASEY diagnoses heart failures, and PROTOS diagnoses hearing disorders
Troubleshooting is a real world problem in which diagnostic CBR systems
are used
• Deployed systems are typically simpler than CASEY or PROTOS
• In a computer support help desk, for example, adaptation and
justification can be omitted
• If a close match is not found, a technician is called to assist
• Many common problems will have solutions already stored
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
6
Justification and Adversarial Reasoning
•
•
•
Adversarial reasoning means making convincing arguments that our own
position is right and that our opponent’s position is wrong
Justification is the same thing, except that there need not be an opponent
HYPO is an adversarial reasoning system
• When a new case comes in, HYPO finds relevant features and
retrieves similar cases
• Some similar cases will support the lawyer’s position and some will
oppose it
• HYPO makes a 3-ply argument
• It starts with a supporting case to make an initial argument
• It takes an opposing case to make counterarguments
• It finds additional supporting cases to counter the
counterarguments
• This helps lawyers avoid being surprised by adversaries in court
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
7
Classification and Interpretation
•
•
•
Interpretation, or understanding a current situation, is often a process of
determining if the current situation fits a particular type of classification
• If we know what type of problem we have, we’re better able to deal
with it
There is no general purpose methodology for getting a computer to
understand a situation
CBR systems do this is domain dependent ways
• PROTOS and HYPO do this by comparing cases based on features
they determine ahead of time are important
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
8
Projection
•
•
•
Projection is predicting the effects of a decision or a plan
It helps in evaluating proposed solutions
• If we can predict that broccoli will get soggy in advance, we can
avoid making it soggy to begin with
• Projection is especially important in planning, to help ensure that the
steps we take lead toward our goal
Battle Planner is a CBR system used for projection
• This was used at West Point, a military academy, to train cadets
• Student commanders plan battle strategies, and Battle Planner tells
them if they would win or lose
• Battle Planner’s cases are real historical battles
• If it projects a loss, the student commander knows to change the
proposed strategy and learns to be a better military tactician
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
9
Case-Based Reasoning vs. Rule-Based Reasoning
•
•
•
Rule-based reasoning (RBR) is the traditional way in which expert
systems were built
A rule is a knowledge representation expressing a relationship among
objects. It contains a piece of knowledge that can be combined, or
chained together, with other pieces of knowledge to build a solution to a
problem
The major differences between CBR and RBR are:
• Rules are patterns. Cases are constants.
• Rules are fired that match input exactly. Cases are retrieved that
match input partially.
• Rules are applied in an iterative cycle of small events. Cases are
retrieved that approximate an entire solution and are then adapted.
• Rules are small, ideally independent but consistent, pieces of
domain knowledge. Cases are large chunks of domain knowledge,
possibly redundant, in part, with other cases.
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
10
Case-Based Reasoning vs. Model-Based Reasoning
•
•
In Model-Based Reasoning (MBR), inferences are made based on some
physical or mathematical model of the problem domain.
• CASEY relied on an MBR system that used a physiological model
of the human heart
Major differences between CBR and MBR are:
• MBR systems store causal models of devices or domains. CBR
systems store examples of devices or satisfactory solutions for a
domain.
• MBR requires that a formal model exists. CBR can work whether a
domain is formalizable or not.
• MBR is good for evaluating proposed solutions, but doesn’t tell how
to generate a solution to begin with. CBR gives example solutions
that can be adapted or reused.
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
11
Multi-Modal Reasoning
•
•
Multi-Modal Reasoning (MMR) systems solve problems by combining
multiple reasoning modalities or approaches.
• These include, but are not limited to, RBR and MBR
This quarter, we are studying CBR
• In building real systems, it is often advantageous to combine CBR
with other approaches
• Don’t let anyone (even me!) tell you that there is a single AI
approach that is right for all problems
CS 682, AI:Case-Based Reasoning, Prof. Cindy Marling
12