The Delta rule can be implemented using Excel. The User interface
is not, perhaps, quite as nice as it could be made in Java or C++, but
(given a working knowledge of Excel) you can see exactly what is going
on.

I have produced a set of spreadsheets which learn the network which will compute the logical AND of two inputs. I have generated a number of spreadsheets, using differing output functions.

- The delta rule using a linear output function (0.5 * activity) (filename delta2.xls) (corrected 13 Dec 1999)
- the delta rule failing to learn NAND because it diverges
- The delta rule using a logistic output function (1/(1+exp(activity))) (filename delta1.xls)
- The delta rule using a tanh output function (tanh(activity)) (filename delta3.xls)

In the first three of these, 1 is TRUE and 0 is FALSE. In the fourth,
-1 is used for FALSE. (There are sometimes problems making the Netscape
Browser start up Excel from these links. Ideas? Try right-clicking...but
this doesn't always work either). Try ftp, either from the browser (using
ftp://ftp.cs.stir.ac.uk)
or from a terminal window:

ftp ftp.cs.stir.ac.uk

(then login as anonymous)

cd pub/staff/lss

binary

get delta.xls

...and so on for the other files

Once you have the spreadsheet program, try altering the learning rate. What happens when it is very small? or when is is large? Can you suggest a mechanism for determining when it is too large?

It should also be a simple matter to alter it to make it learn different 2-input predicates, such as NOR, NAND and OR.

Back to Biologically-inspired Computing Home
page

email lss
last updated 15 November 1999