Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watmath!clyde!cbosgd!ucbvax!vax-d.rutherford.ac.UK!mdw
From: mdw@vax-d.rutherford.ac.UK.UUCP
Newsgroups: comp.ai.digest
Subject: (none)
Message-ID: <7960.8708281200@vd.rl.ac.uk>
Date: Fri, 28-Aug-87 08:00:41 EDT
Article-I.D.: vd.7960.8708281200
Posted: Fri Aug 28 08:00:41 1987
Date-Received: Thu, 3-Sep-87 06:13:15 EDT
Sender: daemon@ucbvax.BERKELEY.EDU
Distribution: world
Organization: The ARPA Internet
Lines: 31
Approved: ailist@stripe.sri.com

Subject: Empirical AI ?

In V5 #201 Andrew Jenning suggested that AI is empirical research.
In V5 #205 Spencer Star countered that it is not since programs do not
confirm or disconfirm theories and do not yield replicable quantitative 
results.

Programs are implementations of models. Models are instantiations of 
theories. Theories suggested by a large range of cognitive scientists
can be empirically tested by writing programs (e.g. Newell, Anderson,
Johnson-Laird, the PDP group, Norman and Rumelhart etc...). If the theory
states that certain phenomina can be produced from a set of 
processing assumptions and a set of data, and a program embodying
these assumptions and using such data cannot produce the phenomina
to an level of abstraction acceptable to the theory, then the 
theory is disproved. 

The requirement that enables programs to
act as tests is that the instantiating and implementational
tradeoffs made are not contrary to any part of the theory. The 
instantiation and implementation processes may involve the recruitment
of additional assumptions to those in the theory, which the theorist
may wish to add to the theory, but this is optional; these additional
assumptions specify a program which is one of a set of programs which
could be derived from the theory. The test can be replicated using
other programs derivable from the theory. The use of models (implemented 
in programs or toy construction sets) can act as tests of theories.

Michael Wilson
SERC Rutherford Appleton Laboratory
U.K.