|positions represent a class. For instance, if the four positions immediately adjacent to the position of interest were chosen, each of these positions might be defined to have four possible states: black stone, white stone, empty, and being off the board.1In this case, there would be 44 =|
NxN 1 2 3
The radius of such a graph-based window is the number of edges that can be traversed and still remain in the subgraph which is extracted. Figure 5.8 shows the study plots obtained for graph-based extraction of small radius.
It is suspected that the poor performance of the graph-based window is that it is very susceptible to explosive complexity, because a large group can have a very large and variable number of adjacent nodes, and the graph-based window as I have proposed it is unable to simply categorize large, strong groups together; perhaps some method of dynamically controlling the extracted subgraph on a better basis than radius would improve the extraction.