Ref.
ftp://ftp.vub.ac.be/pub/projects/Principia_Cybernetica/Texts_General/PCP_Web.txt
Collective Mental Maps (CMM) with algorithms for weighted links for
"Bootstrapping" the Principia Cybernetica Website:
.... "This philosophy can be straightforwardly applied to the
development of knowledge in a system such as Principia Cybernetica Web.
Individual nodes can be seen as pieces of knowledge, and webs of linked
nodes can be seen as knowledge systems. Recombination takes place when
links are changed, so that a node which was connected to one node, is
now connected to another node. Mutation happens when the content of a
node is changed, or a new node is created. Creation of new nodes and
links by different contributors provides a continuous source of
variation. If we want to maintain or improve the quality of knowledge,
we will also have to apply selection. Selection could be done
"manually", by members of the editorial board using their own judgement,
or automatically, by a computer program applying certain formalized
selection rules. The first option is limited by subjectivity and the
bounded cognitive capacities of a human being, the second is limited by
the difficulty to express quality judgements in a formal way, and by the
rigidity of the resulting rules. The most effective approach seems to
consist in a mixed human-computer system, where intuitive judgement is
complemented by computing power (Heylighen, 1991b).
A possible implementation of such an approach could be found in
Thagard's (1992) ECHO program. Here, the judgement of human participants
determines whether two pieces of knowledge (say, propositions or
arguments in a discussion) are either coherent (confirm each other) or
incoherent (contradict each other). The neural network-like computer
program uses these binary coherence relations to decide which of two or
more competing knowledge systems is best, in the sense that its
elements are most coherent with each other and with the whole of the
other knowledge. Although this program was until now only used to
reconstruct some key debates from the history of science, explaining why
one theory eventually replaced its rival, it would seem very promising
to apply such an approach to steer an on-going discussion. Although this
is not really used as yet, PCP Web's annotation function allows
contributors to choose a link type for their annotations, distinguishing
arguments that support a thesis from arguments that refute it. A complex
web of such arguments and counter-arguments could be analysed by
ECHO-like algorithms, attracting the attention towards the more
generally coherent approaches.
Another selection criterion for knowledge, besides coherence, is
simplicity. Paraphrasing Ockham's Razor: all other things being
equal,the simpler a knowledge system is, the better it is (Heylighen,
1994). Key aspects of this criterion can be implemented in a relatively
simple way in a knowledge web. The most straightforward aspect is the
ease with which a piece of knowledge can be located within a web. This
is of particular importance for distributed hypertext systems, which can
contain millions of interlinked nodes, making it virtually impossible to
find a particular node without a priori information. Hierarchical
classification, as discussed earlier, has fundamental shortcomings, and
is poorly suited for a system developed in parallel by different
contributors without co-ordination. Moreover, its model of the
organization of knowledge is inadequate for cognitive systems based on
semantic networks. Free creation of associative links can provide a much
richer model, but is more likely to produce labyrinthine anarchy.
We have recently developed a method that allows an associative hypertext
network to "self-organize" into a simpler, more meaningful, and more
easily usable network. The term "self-organization" is appropriate to
the degree that there is no external control or editor deciding which
node to link to which other node: better linking patterns emerge
spontaneously. The information used to create new links is not internal
to the network, though: it comes from all users collectively. In that
sense one might say that the network "adapts" to its users, or "learns"
from the way it is used.
Algorithms for such an adaptive web can be very simple. Every potential
link is assigned a certain "fitness" or "strength". For a given node A,
only the links with the highest fitness are actualized, i.e. are
accessible to the user. Within the node, these links are ordered by
strength, so that the user will encounter the strongest link first. We
have formulated three separate learning rules for adapting the
strengths:
1) A link, say A -> B, which is directly chosen by the user, increases
its strength. This rather obvious rule can only reorder the links that
are already available within the node. By definition it cannot actualize
new links, since these are not accessible to the user. This necessitates
another rule.
2) A user might follow an indirect connection between two nodes, say A
-> B, B -> C. In that case the potential link A -> C increases its
strength. This is a weak form of transitivity. It opens up an unlimited
realm of new links. Indeed, one or several increases in strength of A ->
C may be sufficient to make the potential link actual. The user can now
directly select A -> C, and from there perhaps C -> D. This increases
the strength of the potential link A -> D, which may in turn become
actual, providing a starting point for an eventual further link A -> E,
and so on. Eventually, an indefinitely extended path may thus be
replaced by a single link A -> Z. Of course, this assumes that a
sufficient number of users effectively follow that path. Otherwise it
will not be able to overcome the competition from paths chosen by other
users, which will also increase their strengths. The underlying
principle is that the paths that are most popular, i.e. followed most
often, will eventually be replaced by direct links, thus minimizing the
average number of links a user must follow in order to reach his or her
preferred destination.
3) A similar rule can be used to implement a weak form of symmetry. When
a user chooses a link A -> B, implying that there exists some
association between the nodes A and B, we may assume that this also
implies some association between B and A. Therefore, the reverse link B
-> A gets a strength increase. This symmetry rule on its own is much
more limited than transitivity, since it can only actualize a single new
link for each existing link.
However, the collective effect of symmetry and transitivity is much more
powerful than that of any single rule. For example, consider two links
A1 -> B, A2 -> B. The fact that A1 and A2 point to the same node seems
to indicate that A1 and A2 have something in common, i.e. are related in
some way. However, none of the rules will directly generate a link
between A1 and A2. Yet, the repeated selection of the link A2 -> B may
actualize the link B -> A2 by symmetry. The repeated selection of the
already existing link A1 -> B followed by this new link can then
actualize the link A1 -> A2 through transitivity. Similar scenarios can
be conceived for different orientations or different combinations of the
links.
A remaining issue is the relative importance of the three above rules.
In other words, how large should the increase in strength be for each of
the rules? If we choose unity (1) to be the bonus given by the first
rule, there are two remaining parameters or degrees of freedom: t is the
bonus for transitivity, s for symmetry. Since the direct selection of a
link by a user seems a more reliable indication of its usefulness than
an indirect selection, we assume t < 1 , s < 1. The actual values will
determine the efficiency of the learning process, but it seems that this
matter cannot be settled by pure theoretical reasoning.
In order to test these ideas in practice we have set up two experiments.
We built a web consisting of 150 nodes, corresponding to the 150 most
frequent nouns of the English language (derived from the frequency list
of Johansson & Hofland, 1989). Every node was assigned 10 links to other
nodes. These links were randomly selected from the 149 remaining nodes
to initialize the web, but would then evolve according to the above
learning rules (with t = 0.5 and s = 0.3). We made the web available on
the Internet, and invited volunteers to browse through it, selecting
those links from a given node which seemed somehow most related to it.
For example, if the start node represented the noun "dog", a user would
choose a link to an associated word, such as "cat", "animal", or "fur",
but not to a totally unrelated word, such as "mathematics". Of course,
in the beginning of the experiment, there would be very few good
associations available in the lists of 10 random words, and users might
have to be satisfied with a rather weak association, such as "meat".
However, when reaching the node "meat", they might be able to select
there another association, such as "carnivore". Through transitivity, a
new link to "carnivore" might then appear in the node "dog", displacing
the weakest link in the list, while providing a much better association
than the previously best one, "meat".
In the first experiment we only used the direct and the transitive
rules. After some 6000 link selections made by several hundreds of
users, the network seemed to have settled in a relatively stable state.
The most frequented nodes had gathered a list of 10 strongest links that
quite well reflected their direct semantic environment, with words that
were near synonyms of the node name at the top of the list (see Table
1). However, this positive result was not reflected in the less
frequented nodes, because of what we termed the "attractor effect".
Nodes that had many incoming links, by accident, or because they had
many associations with other words in the list, would tend to attract
more users, which would result in increasing strength of their incoming
paths, and their replacement by even stronger direct links. In the end
almost all paths would end up in a cluster of semantically related,
strongly cross-linked nodes, forming an approximate attractor for the
network. Although new users were randomly assigned to a node when
entering the network, so that all nodes would be consulted on first
entry with the same average frequency, the subsequent moves would very
quickly end up in the attractor cluster, so that nodes and links outside
the attractor would get little chance to learn, and as a result would
remain poorly connected.
KNOWLEDGE
0 200 800 4000
view education education education
health experience experience experience
theory example development research
face theory theory development
book training research mind
line development example life
world history life theory
side view training training
government situation order thought
trade work effect interest
Table 1: self-organization of the list of 10 strongest links from the
word "knowledge", in different stages: initial random linking pattern,
after 200 steps, after 800 steps, and after 4000 steps. A step
corresponds to a user selecting a link on one of the 150 nodes, in a web
that evolves according to the direct, transitive and symmetric learning
rules (2nd experiment). Note that only one of the initial links
("theory") survives after 4000 steps, and that the evolution slows down
considerably.
In our second experiment, we added the symmetry rule to the two other
rules. This led to a faster initial learning, since a user passing
through a node with zero-strength links would immediately generate two
new links (one by symmetry and one by transitivity), which were on
average much better than the initial random links. In the longer term,
symmetry moreover attenuated the attractor effect, since strong links
leading into an attractor would necessarily produce weaker, inverse
links leading to the periphery. This gave nodes outside the attractor
the chance to develop some links of their own, generating local
attracting clusters weakly connected to other clusters. The overall
learning seemed more efficient in the sense that less time was needed to
develop good associations, and the result was more balanced, in the
sense that the differences in frequentation between nodes were less
strong. Still, the differences were still large enough to makes consider
additional mechanisms for reducing the attractor effect.
We are now trying to determine to what degree these results from the
adaptive network correlate with different word associations derived by
other means (e.g. free association experiments, or letting people judge
the degree of synonymity). We also plan to test the usefulness of the
self-organization, by checking in how far users find knowledge more
effectively in a self-organized network, as compared to a network that
did not undergo learning. This can be done by measuring the average
number of steps needed to find a node, or the average time needed to
choose a link. We are further considering additional learning rules,
such as similarity (nodes sharing several links would get stronger
cross-connections), that may make the learning more effective.
Although this research is still in its initial stage, and will need much
empirical testing to confirm its usefulness, it seems like a very
promising approach to quickly and easily develop complex knowledge webs
that are more adequate than webs built manually. It may become
especially helpful for the World-Wide Web as a whole, by allowing the
automatic creation of links between servers maintained by different
people in different parts of the world.
Conclusion
Although the system we presented is still under development, changing
almost every day, it has already become quite clear that the World-Wide
Web paradigm, augmented with dynamic restructuring of the network,
provides an extremely powerful tool for the publication and
collaborative development of complex knowledge systems. Such a tool can
be especially useful for the discipline of Systems and Cybernetics,
characterized by a very rich but unstable and ill-structured knowledge
base, which is difficult to access by traditional means. We hope that
many researchers in that domain will join these efforts, either by
contributing directly to the Principia Cybernetica Web, or by starting
parallel services, which can then be cross-linked to the PCP Web. If
that happens, the domain's historical mission of transdisciplinary
integration (Boulding, 19; von Bertalanffy, 1968) may again become a
practical issue, rather than a far-away dream."
Jack Park wrote:
> Consider these common use cases
> EMAIL
> User receive email
> User send email
> User annotate email
> User replyTo email
> OHS archive email
> OHS autoLink email
> SDS
> <note>email annotation already covered</note>
> User align records
> OHS autoLink records
> <note>I'm sure Rod will have lots more here</note>
> WEB
> User browse webpage
> User annotate webpage
> OHS autoLink webpage
> COLLABORATE
> <note>email and web fit in here</note>
> User create document
> Usser edit document
> User shareDocumentWith OtherUser
> User pose IBISQuestion
> User respondTo IBISQuestion
> OHS maintain IBISQuestion
> OHS maintain IBISResponse
> OHS autoLink IBISQuestion
> OHS autoLink IBISResponse
>
> Under these are some really primitive use cases
> OHS access webpage
> OHS access email
> User access OHS
>
> Let us examine these use cases.
> Actors:
> User, document, OHS, OtherUser, IBISQuestion, IBISResponse, email,
> records
> Verbs:
> receive, send, respondTo, archive, autoLink, align, create, edit,
> shareDocumentWith,
> pose, maintain, browse, annotate
> We can see that there is great similarity between 'create', 'pose', and
> 'send'
> 'autoLink' is a really exciting verb. Some verbs require user action,
> others are purely OHS behaviors. Some verbs need rethinking.
>
> Notice that, when we begin to flesh these use cases out, we are beginning to
> imagine the underlying mechanics of an OHS. We can now take these nouns and
> verbs, refine them, refine our use cases, develop an ontology that narrows
> the range of words we choose to those necessary to accomplish the design
> task, construct scenarios with the new ontology, perhaps refine the ontology
> and use cases, and iterate until we believe we are ready to hack some code.
>
> I recognize the fact that the use cases mentioned above appear to ignore the
> vast amount of energy this group has already put into the development of use
> cases. It is my hope that the two apparently disparate activities will
> ultimately enhance each other. It would seem that we could take my
> minimalist list and begin to flesh out an OHS.
>
> Once we get all this common stuff fleshed out, we can begin to look at the
> two specialty tracks: research collaboration (NIH), and software
> productivity. That will likely call for new iterations in the common stuff
> because ideas generated in the specialty field will be seen to have value
> across many domains.
>
> ============================================================================
> This message is intended only for the use of the Addressee(s) and may
> contain information that is PRIVILEGED and CONFIDENTIAL. If you are not
> the intended recipient, dissemination of this communication is prohibited.
> If you have received this communication in error, please erase all copies
> of the message and its attachments and notify postmaster@verticalnet.com
> immediately.
> ============================================================================
>
>
> Community email addresses:
> Post message: unrev-II@onelist.com
> Subscribe: unrev-II-subscribe@onelist.com
> Unsubscribe: unrev-II-unsubscribe@onelist.com
> List owner: unrev-II-owner@onelist.com
>
> Shortcut URL to this page:
> http://www.onelist.com/community/unrev-II
-------------------------- eGroups Sponsor -------------------------~-~>
Big News - eGroups is becoming Yahoo! Groups
Click here for more details:
http://click.egroups.com/1/10801/0/_/444287/_/977470882/
---------------------------------------------------------------------_->
Community email addresses:
Post message: unrev-II@onelist.com
Subscribe: unrev-II-subscribe@onelist.com
Unsubscribe: unrev-II-unsubscribe@onelist.com
List owner: unrev-II-owner@onelist.com
Shortcut URL to this page:
http://www.onelist.com/community/unrev-II
This archive was generated by hypermail 2b29 : Thu Dec 21 2000 - 23:51:56 PST