COGNITIVE MODELLING: LECTURE 1, 2002

How do we use language, problem-solve and operate effectively in a complex, dynamic world?

How do we learn to carry out these functions?

COGNITIVE MODELLING

http://www.informatics.sussex.ac.uk/users/bend/cogmod/outline.html

Course Outline

Aims and Objectives

Reading List -- Green et al. ``Cognitive Science''

Computer Classes -- Week 2 onwards: working in pairs (if possible)

Seminars -- Week 1 onwards: Sharples et al. (1989), Chapter 2 of Green et al.

Assignment -- Log of computer-based experiments

Assessment -- Essay plus Unseen

Types of Models

TYPES OF MODELS

Non runnable: description

diagrammatic -- box/arrow, flowchart (see Green et al.)

descriptive -- ``working memory can hold 7 plus/minus 2 chunks''

mathematical/logical e.g. an equation or formula

Runnable: description plus behaviour

Mechanical/hydraulic -- ``automata'', ``inference engine''

Computer program

ARTIFICIAL INTELLIGENCE

Contributed runnable and non-runnable models to Cognitive Science. Contrast ``engineering'' type AI.

For runnable cognitive models:

Parts can be changed, rules of interaction of parts can be changed: behaviour of model can be compared with behaviour of what is being modelled

Granularity and level of model -- what counts as ``behaviour''?

What aspects are not being modelled?

BUGGY EXAMPLE -- SURFACE BEHAVIOUR

Which child would you like to test: a, b, c, or d? a
Now set some subtraction sums for the child.
You will be asked to type the top row of the sum,
then the bottom row.
Press <RETURN> instead of typing a number
when you want to finish with this child.
Testing child a
    Top number:    ? 23
    Bottom number: ? 4
  2  3
     4
------
  2  1

The program will now set five sums and you should answer
them as if you were child "a". If you want to quit
earlier then press <RETURN> when asked for your answer

Top number:    683
Bottom number:  76
What answer would child "a" give? 613

  6  8  3
     7  6
---------
  6  1  3

Yes - that is the child's answer

Top number:    929
Bottom number: 235
What answer would child "a" give? 714
  9  2  9
  2  3  5
---------
  7  1  4
Yes - that is the child's answer
Congratulations. You appear to understand child "a"

This is my description of the bug:
Child a: the child always subtracts the smaller digit from
         the larger in a column

BUGGY EXAMPLE -- INNARDS

production rules e.g.:

prule fd: [m ?m][s ?s] => assert([nextcolumn]);
                          assert([finddiff]); endprule;
prule b2a: [s greater m] => assert([borrow]); endprule;
prule bs1: [borrow] => addtentom(); endprule;
prule bs2: [borrow] => decrement(); endprule;
prule cm: [m ?m][s ?s] => compare(); endprule;
prule in: [processcolumn] => readmands(); endprule;
prule nxt: [nextcolumn] => assert([processcolumn]);
                           shiftleft(); endprule;

production rule interpreter

subsidiary parts of the program e.g. printing procedures

POSSIBLE EXPERIMENTS

Parts can be changed, rules of interaction of parts can be changed:
select rule mix and order, change rules themselves
change way rules are interpreted
behaviour of model can be compared with behaviour that is being modelled: offer different sums

Granularity and level of model -- what counts as ``behaviour''?
choice of digits in columns in answer, some intermediate digits

What aspects are not being modelled? timing, understanding meaning of subtraction, subtraction in context, motivation ...

DIFFERENT FLAVOURS OF ARTIFICIAL INTELLIGENCE

``Nouvelle AI'' & ``GOFAI -- Good Old-Fashioned AI'':
How far does do the processes being modelled seem to depend on there being some kind of an internal symbolic representation of the ``world''?

Complete simple organism interacting with its environment e.g. a simulated snail on a simulated leaf

A robot fish in a real pond

A (presumed) single (human) mental process -- problem-solving, language understanding, planning

The overall cognitive architecture in which such processes operate and are integrated into a coherent whole.

REPRESENTATION OF WORLD IN BUGGY

For model of child ``a'' -- static set of production rules that mimic his/her behaviour.

A production rule interpreter that uses the rules in a consistent manner.

A dynamic ``working memory'' that changes according to the state of the problem solving.

Models of Symbolic Computation

For seminar in week 1 read Sharples et al. Foreword and Chapter 3 For seminar in week 2 read Green et al. Chapter 2

propositional representations

production systems

connectionist and parallel distributed processing

hybrid models

architectures

CONCLUSIONS

The AI modelling approach to Cognitive Science

The strengths and weaknesses of modelling and of particular models

Kinds of Cognitive Models

Weeks 2-5

Runnable	Symbolic	Type	Domain
Yes	Yes	Production Systems	Arithmetic
Yes	Yes	Propositional	Language
Yes	No	Neural Networks	Language
Yes	No	Behaviour-based Modelling	Robotics