FOA Home | UP: Information-seeking agents

Adapting to ``spatial'' context

In order to test InfoSpiders, agents were allowed to breed in a controlled test environment previously used by Steier [REF866] [REF1112] : a portion of the Encyclopaedia Britannica (EB). The advantage is that we can make use of readily available relevant sets of articles associated with a large number of queries. The subset corresponds to the {\tt HUMAN SOCIETY} topic, roughly one tenth of the entire encyclopedia. Links to other parts of the EB are removed along with any terminal documents produced as a result. The final environment is made of $N=19427$ pages, organized in a hypertext graph (the EB is already in HTML format). 7859 of these pages are full articles constituting the {\em Micropaedia}. These, together with 10585 {\em Index} pages (containing links to articles and pointed to by links in articles), form a graph with many connected components. The remaining 983 nodes form a hierarchical topical tree, called {\em Propaedia.} These nodes contain topic titles and links to children nodes, ancestor nodes, and articles. Micropaedia articles also have links to Propaedia nodes. Propaedia and Index pages are included in the search set to ensure a connected graph and to be faithful to the EB information architecture --- an actual subset of the Web.

Now consider two agents, A and B, born at the same time and attempting to satisfy the same query, but in different ``places'' within this hypergraph of EB document pages.

As Table (FOAref) shows, the original query words were displaced from their top positions and replaced by new terms. For example, {\tt PRIVAT} and {\tt ALLEVI} had relatively low weights, while {\tt FOUNDAT} appeared to have the highest correlation with relevance feedback at this time.

A's and B's keyword vectors are shown in Table (FOAref) . In the course of the evolution leading to A and B through their ancestors, some query terms were lost from both genotypes. A was a third generation agent; its parent lost {\tt ALLEVI} through a mutation in favor of {\tt HULL}. At A's birth, {\tt PRIVAT} was mutated into {\tt TH}. B was a second generation agent; at its birth, both {\tt ALLEVI} and {\tt PRIVAT} were replaced by {\tt HULL} and {\tt ADDAM}, respectively, via mutation and crossover. These keyword vectors demonstrate how environmental features correlated with relevance were internalized into the agents' behaviors.

The difference between A and B can be attributed to their evolutionary adaptation to spatially local context. A and B were born at documents $D_A$ and $D_B$, respectively, whose word frequency distributions are partly shown in Table (FOAref) . {\tt TH} represented well the place where A was born, being the second most frequent term there; and {\tt ADDAM} represented well the place where B was born, being the third most frequent term there. By internalizing these words, the two situated agents are better suited to their respective spatial contexts.

Top of Page | UP: Information-seeking agents | ,FOA Home

FOA © R. K. Belew - 00-09-21