-- for Abstraction
-- for Problems
-- Emotion Patterns
Selfridge's Original Pandemonium
The present author confesses that his interest in Pandemonium
lies mostly in a simple generalization of the architecture, and
that he has done little research into the important work by
Selfridge and work that followed.
Anyone interested in correcting, critiquing, contributing to or
editing this part of the website is encouraged to contact
Brian Marshall at
The following is a very brief description. A somewhat more
detailed description is available at
In 1959, O. G. Selfridge proposed an architecture, that he called
Pandemonim, that could be useful for pattern recognition
[SEL1]. He demonstrated that it could
effectively used to distinguish dots and dashes in hand-keyed
morse code and to distinguish hand-printed characters out of a
ten possibilities [SEL2].
Selfridge's pandemonium is composed of four layers:
- a bottom layer of demons that acquire and store input data,
and then pass it on to...
- a layer of computational demons that perform computations on
the data they receive and attempt to identify evidence of
particular patterns, which they pass on to..
- a layer of cognitive demons that consider the evidence
received, and where sufficient evidence exists, shriek
- a top layer composed of a decision-making demon that decides,
based on the shrieking it hears about evidence, what
pattern was actually presented to the bottom layer.
Selfridge's pandemonium used two types of learning:
- initially, in a phase of supervised learning, the weights
between computational demons and cognitive demons are
trained in order to optimize performance, and,
- secondly, useless demons, that contribute little to
decision-making, are eliminated, and useful demons are
used to create new demons, either by mutation or by
combining two useful demons.