Often the top level compares evidence from different source types, for example,
Chemical Evidence, Physical Evidence (Historical and Exam), Imagery (X-ray)

See figures 1-5 in "Structured Matching: A Task-Specific Technique for Making Decisions"

Also see section 3.1, "HYPER" for some specifics on representation and control

Issues:

• Tool for user to build matchers
• Tool to perform proper control
• Tool for interaction with database (cases)
• Interaction with higher level ps's (storage, names, etc)

Task Formulation:

• Given a hypothesis and a set of features, determine the degree to which the hypothesis matches the situation (specific to classification).
• Given a small number of choices for a decision and a set of parameters relevant to making the decision, where each parameter can take a value from a small number of values, determine a choice.

Form of Knowledge:

Patterns that map the parameters & results of other simple matchers to choices

Organization of Knowledge:

• Hierarchical Organization of evidence abstractions (simple matchers),where each simple matcher makes a subdecision based on a number of inputs. Each simple matcher specifies how selected parameters and the results of its children should be mapped to a choice.
• The top node (root) represents the degree to which the hypothesis matches the situation (i.e., makes the final decision). Subnodes represent evidence components for their parent nodes.

Method:

In HYPER, the evaluation of a structured matcher can be initiated in a top-down fashion, i.e, each simple matcher calls on its children as needed in order to evaluate the conditions of a pattern. Information flows bottom-up.

For each hypothesis, set up a subtask to establish or reject it. If it is established, test its successors. If it is rejected, it and its successors are rejected. The top-down control strategy, called Establish-Refine, can be varied under specific conditions. [T. Bylander and S. Mittal. CSRL: A Language for Classificatory Problem Solving and Uncertainty Handling. AI Magazine 7(3):66-77, 1986] elaborate on this simplified account.

Specialists:

• hierarchy of specialists mirrors diagnostic classification hierarchy
• each node in hierarchy corresponds to state of the situation under consideration (hypothesis)
• domain knowledge distributed across specialists

Advantages:

• divides problem into managable pieces
• organizes knowledge into pieces which become relevant when corresponding concepts become relevant.
  • confirmatory rules
  • exclusionary rules
  • recommendation rules

(* note: these could be rules, could be logic, could be frames, could be pattern recognition, etc.

The organization is important, not the lower level implementation choice)

- nodes higher in the hierarchy represent general hypothesis; lower represent more specific hypothesis (top-down)

Establish: specialist establishes because relevant (high confidence)

Refines by invoking subs

       (which then perform establish/refine)

Reject: confidence low - performs no further actions - prunes hierarchy

Suspend: moderate confidence - may later refine if requested (blackboard - secondary to...)

      * non establish-set lower ; ask for more info

Recommendation: pathognomonic rules

Succinctly:

Knowledge Representative primitives:

1. Classification hierarchy
          subclass of, superclass of
2. Classificatory knowledge implies a specific organization

Organization of Knowledge:

• Classificatory knowledge
• Distributed among classificatory concepts
• Concept can establish or reject itself (degrees allowed)

Control Regime:
Variations on establish-refine on the hierarchy of concepts

Goal Types:
Establish , reject , suspend <>, refine<>, explain

See "CSRL: A Language for Expert Systems for Diagnosis"

Generation of explanation

(if a b)
b .
a

(If ( drunk ?person) (not (walk-straight ?person)
(not (walk-straight Jack) .
(drunk Jack)

Only a guess...Not a valid logical implication

Jack could have just got off a roller coaster.

Abduction is only plausible inference

Careful : Our notion of explanation => causality

causality and logical implication are not the same thing

(if (in ?patient ward5) (have ?patient cancer))
(have Eliza cancer) .
(in Eliza ward5)

(cause  ?x  ?y)        : x causes y
?y                  .  : and y is true 
?x                     : so hypothesize ?x as explanation

Diagnosis:

D is a collection of data (facts, observations)
Hypothesis H is plausible under the circumstance
H explains D (would, if true, explain D)
No other hypothesis explains D as well as H does
Therefore, H is correct

Abductions go from data describing something, to an explanatory hypothesis that best accounts for that data. Other names

inference to the best explanation

the explanatory inference

The mechanism:

a) a classification machine for selecting plausible hypotheses

b) a specialized means-ends machine for assempling plausible hypotheses into a best explanation.

The assembler is used repeatedly by an overview critic which

• produces an initial "best explanation" for the case
• nvestigates the space of alternative explanations to critique the initial one
• produces a final "best explanation"

• A and B are mutually compatible - explanatory capabilities overlap
• Hypothesis A is a subhypothesis of B(a more detailed refinement)
• A and B are mutually incompatible
• A and B cooperate additively where they overlap

• as complete as possible
• maximally plausible
• parsimonious

Each plausible hypothesis delivered by the classifier comes with:

• a description of which findings it explains
• a plausibility value (confidence in this hyp)
• significance
• interactions with other hypotheses

One knows what they want to achieve
and what they have. Means-ends centers around the detection of differences between the current state and the goal state...and what means can be used to get to the goal state

Design by Plan Selection and Refinement

Task:

Design an object that satisfies certain specifications.

Design is separated into a hierarchy of subdesign problems, mirroring the object's component structure. For each node in the hierarchy, there are plans for making commitments for some component parameters. Each component is designed by choosing a plan, based on some specifications, which instantiates some design parts and designs further subcomponents to fill in other parts. [Brown, D.C., Chandrasekaran B. Design Problem Solving: Knowledge Structure and Control Strategies. Morgan Kaufmann, Inc. San Mateo, CA 1989] is the definitive reference on this topic