pseudoitems through the rehearsal buffer also has some
useful properties. Without this rotation a buffer containing a
fixed set of pseudoitems would use (for each new item)
fewer pseudoitems more frequently, and would thus be prone
to “over-fitting” these pseudoitems and to conflicts when
pseudoitems fall very close to the new item.
It is important to note that pseudorehearsal does not
preserve the weights of a network, these can change
dramatically during learning. The method works directly at
the level of the function embodied by the network. The old
population function is preserved, and changes necessary to
accommodate the new item are restricted to being local to
the region of the new item. This localisation of changes to
the function is the essence of the pseudorehearsal method.
These and other issues are pursued in more detail in Frean
and Robins [1997] and Robins [1995].