Ralf DerSee videos here
Can the self-organization of behavior be more than wishful thinking?
From our experiments with real robots we have learned that true self-organization
requires a driving force which must be rooted in the agent itself.
A useful example of this thinking is given by the principle of homeostasis
according to which the behavior of an agent in the world arises from its
desire to keep internally an equilibirum state.
We introduce a new principle - homeokinesis - which is completely unspecific
and yet induces specific, seemingly goal-oriented behaviors of an agent
in a complex external world. We view homeokinesis as the dynamical pendant
of
homeostasis. Like the latter it derives behavior from an entirely internal
perspective so that coarsely speaking behavior emerges as a by-product
of satisfying the internal needs of the agent. We give this general aim
a constructive formulation in the following way. We suppose that the agent
is equipped with an adaptive model of its behavior. A learning signal for
both the model and the controller is derived from the misfit between the
real behavior of the agent in the world and that predicted by the model.
As we could show with several examples this misfit is minimized if the
agent exhibits a smooth, controlled behavior.
In this way, a learning signal for the adaptation of the behavior is
derived from a purely internal perspective. However we have found that
using a predictive model will lead to active behavior modes only
if the robot is given a drive for activity from outside. When using
a retrospective model the system is found to develop this drive for activity
by itself due to spontaneous symmetry breaking mechanisms known from the
physics of self-organizing systems. The reason is that the model now generates
a dynamics backward in time leading to fluctuation amplification which
is a necessary prerequisit for self-organization.
The kind of behavior learned by the agent depends on the structural complexity of the behavior model. Different models generate different modes of behavior. This may be heplful in artificial evolution in that by modifying the structural complexity of the behavior model evolution may play directly with complex behavior modes which are already more or less ''reasonable'' in the world.