Class Notes: Chapter 7

Chapter 7: Section 7.1

What is the goal of expert systems research and machine learning research?

Use decision rules to process complex information
Decision rules are IF-THEN rules
One point of interest to notice:
Many systems work the same way to control behavior: IF condition, THEN action

In what way do expert systems research and machine learning research instantiate physical symbol systems?

There are two symbol structures in machine learning systems:
The original data base of observations
The emerging decision rules
The system works to transform the former into the latter, using heuristic search

A Side Trip: Production Systems

A production is an IF-THEN rule: IF condition THEN action
A production system is a set of productions arranged in hierarchical order

Start with the first production
Go down the list until a condition is satisfied
Execute the action
Return to the top

An action can call another production system

Example of a Production System: How to play tic-tac-toe

Assume player is X, opponent is O
Define a stimulus pattern [- - -] as a straight line of three symbols, X, O or ? (blank)
Rule 1: [X?X] or [?XX] or [XX?] ? [XXX]
Rule 2: [O?O] ? [OXO], [?OO] ? [XOO], [OO?] ? [OOX]
Define a pattern [-- - --] of two intersecting lines
Rule 3: [X? X ??] ? [X? X ?X]

Production Systems

Production systems make for a valuable, general-purpose method of representing algorithms
For many psychologists they form the basis of procedural knowledge – knowledge of skills
Procedural knowledge is clearly independent of declarative knowledge: Knowing How versus Knowing That
Why do we believe they are modular?

The Structure ofKnowledge

Data for Production Systems

Where does a PS look to see if a condition is satisfied? Where do the actions take place?
For tic-tac-toe we use an external register
For much of our thinking, the register is internal
We call it working memory

Where does working memory find its data?
Sometimes in an external stimulus, sometimes in long term memory

Representing Information in Long Term Memory

LTM is clearly associational, but what format do the associations use?
For most adherents to the PSS hypothesis, there is a close connection between language and memory
Information is stored as propositions

Associations and Activation

Memory is a network of associations
The network is organized in meaningful patterns
Some of those associations will be active at any one time (this is working memory)
We retrieve information by allowing the activation to spread through the network of associations
This might be automatic, or controlled by a production system

Memory Processes

This framework can be used to understand memory phenomena
Encoding: Establishing certain associations
Storage: The organization of the associations
Interference from other associations
Retrieval: Spreading activation may, or may not, activate the intended memory

Other Relevant Findings

Bransford-Franks Study:
Test sentences:
The red jelly was on the table
The jelly was on the plate on the table
The ants were in the kitchen
The ants ate the jelly

Recognition:
1. The jelly was sweet
2. The ants ate the red jelly on the plate
Greater confidence for #2 than for #1

Reaction Time Studies

Test questions:

1. Does a canary have wings?
2. Does a fish swim?
3. Does a sparrow eat?
4. Is a cardinal red?
5. Does a bird have wings?

Typical Findings: #5 faster than #1, #4 faster than #3

These findings led to the development of numerous association network theories

Tests of the Model: The Fan Effect

Subjects study the following sentences:
The lawyer is in the park
The painter is in the boat
The lawyer is in the chair
The sailor is in the house
The student is in the park
The lawyer is in the house
The sailor is in the park
Later they are tested for recognition of a sentence (Was the lawyer in the park?)

Predictions

Notice the varying size of “fans” of associations
Compare Lawyer-Park with Painter-Boat
The response time is longer if the fan is larger
Activation is divided among more associations

Chapter 7: Section 7.3

Explain the basics of how WHISPER works?

The behavior of blocks in a world where gravity exists

How does WHISPER differ from other physical symbol systems?

The symbols are diagrams, not words
WHISPER deals with an analog system
PSSH can apply to both digital and analog representations
Representation of the world in terms of symbols
Manipulation of symbols using heuristic search

Chapter 7: Section 7.4

What sort of knowledge does SHAKEY have about its environment and itself?

Robot research has been a major area for development of physical symbol systems
SHAKEY: Statements represent rooms and objects in the environment, and the robot itself
The knowledge is the kind of knowledge that Lashley described in accounting for planning behavior

What are the levels of SHAKEY’s software?

1. The physical vehicle and connections
2. Low Level Activities: Basic physical capabilities
3. Intermediate: Operate on specific object
4. Planning mechanisms (reminiscent of intentional level actions)
5. Executive program: Capable of means-end analysis

Intelligence in SHAKEY

At the abstract level, SHAKEY can reason using the predicate calculus
Predicate calculus: an extension of propositional calculus
Propositional calculus deals with simple propositions. Example: if p then q, q is false, therefore p is false
Predicate calculus adds in the notion of quantification: Some, all, none

SHAKEY uses the STRIPS planner: Purely mechanical algorithms (operators) for identifying ILAs that lead to a goal state
Closely related to Simon & Newell’s general problem solver
Heuristic search using means-end analysis
ILAs are the operators that move SHAKEY through the problem solving space

Is it significant that SHAKEY inhabits a real world rather than a virtual one?

Real world, but SHAKEY is still limited in application

After seeing the illustrations of physical symbol systems given in Chapter 7, what do you think are the strengths and weaknesses of the physical symbol system approach?

Let’s summarize the difficulties with physical symbol systems
Later chapters will explore some alternatives

Physical Symbol Systems

The systems are rule based
Problem space needs to be explicitly represented
Processing involves axioms and inference rules operating one at a time (serially)
Rules and axioms operate on the syntactic structure of physical symbols

Plausibility of PSS

The major objections:
Symbol-grounding problem and the Chinese room (Is this a serious problem?)
Frame problems

The frame problem: How can a formal system represent the changes brought about by an action without explicitly representing all the things that the action does not bring about?

Frame Problems and PSS

Some commentators have argued that frame problems present an “in-principle” objection to PSS
(Alleged) impossibility of formalizing commonsense reasoning
Often accompanied by emphasis on “situatedness” and “embodiment” of real cognitive agents (see later chapters)

Practical Problems with PSS: Many cognitive abilities are difficult to model in a rule-based way

Context effects in perception (produces an explosion of rules)
Pattern completion
Language (particularly languages such as English with many irregular forms)

Soft Constraints

Much of cognition is governed by “soft” constraints
Rules that hold “for the most part" and don’t fully determine outcomes
They are often in conflict with each other
Hard to formulate rules governing which constraints should taken precedence in which circumstances

Biological Plausibility

PSS are serial in operation
Far too slow for many cognitive processes
Graceful degradation vs brittleness

Learning:

Typically PSS models consist of completed cognitive abilities
Not easy to see how programs could emerge in the normal course of development
Related to lack of flexibility and lack of a real-time dimension