The Artificial Browser Agent

Reliability: Equal network with equal browsers?
The reliability with which trained, Adaptive Webs converge to one single state when being trained by the same browsers, is an indication of the stability of network development. This reliability can be compared to the required reliability of measuring apparati: does a thermometer when repeatedly used to measure the same temperature actually indicate the same temperature? The effects of the previously mentioned feedback loop could cause network development to be extremely sensitive to initial conditions and make each network settle into an equally well-structured, but different state.

Validity:do developped networks accurately resemble browsers's associations?
As mentioned, hypertext networks should not violate their browsers intuitive expectations and therefore their structure should as accurately as possible reflect their browsers knowledge structure. Reliabilty is an necessary but not sufficient condition for validity of measurement. A thermometer can e.g. repeatedly indicate the same temperature when confronted with an actually equal temperature, but might be constantly biased. The indicated temperature might be reliably measured, but it is not the temperature in question.
The question is wether our Adaptive Web Systems actually makes networks evolve so that they accurately resemble their browsers structure associative intuitions.

The Artificial Browser To test the assumptions of reliability and validity, hypertext networks needed to be repeatedly trained by exactly the same browser. Otherwise, variations among the evolved networks might be attributed to changes in browser characteristics rather than as an indication of developemental reliability.
The unpredictible nature of human behaviour led us to the development of a simulated, idealised browser.
The Artifical Browser Agent was programmed so that it could could train hundreds of adaptive networks by browsing them using a fixed map of associations (derived from our data in previous real-life experiments). The heuristic used by our Artificial Browser were based on the model of a completely serenpenditious browser. Starting from a random position in the network, it would continuously link to the associatively most related (evaluation based on fixed assocative map in memory) and best ordered link.
The resulting networks from our simulations could then be correlated to each other and to the browsers associations to respectively measure the reliability (high congruence=high reliability) and validity (high resemblence browser-network=high validity) of network development under the Adaptive Web System.

Results of the simulation

Reliability was measured as the QAP correlation [article on QAP) between 20 independently trained networks. (each 3000 jumps by Artificial Browser). This correlation averaged to 0.79, indicating a relatively high reliability
Validity was measured as the QAP correlation between 20 independently trained networks (3000 jumps by Artificial Browser each) and the browsers associative map. This correllation averaged to 0.82, indicating a high overlap and consequently a relatively high network validity.

These results indicate the Adaptive Hypertext System is in principle an accurate and reliable way of measuring browsers association, which are manifested in the resulting hypertext network structure. The hypertext networks generated by the Adaptive Hypertext System accurately represent their browsers' associative expectations and intuitions.