Major change: Groups are eliminated. Features are created directly now using
the old group-creation algorithm. This in turn spawned a large number
of minor changes to support it. The end result is that features in the higher levels
of the hierarchy tend to be created more gracefully, make more sense
to a human user, and naturally terminate. The number of user-selected parameters
in BECCA also went down a bit.
• The maximum possible size for large variables (the co-activity estimate and
the model) is allocated on startup. This helps prevent running out of memory
mid-run.
• The learner was renamed actor to more accurately reflect its function
and to provide symmetry with perceiver in the sense of “perception-action
loops".