p.17 Even when sophisticated information processing techniques are brought to bear, however, many
problems stubbornly resist solution. The initial promises of cybernetics, and more recently, of artificial intelligence,
have proved harder than expected to attain.
p.18 Systems learn about their environment by attempting to control it, and modify their
representation of the environment as a function of the results of those attempts at control.
p.20 "Does there exist a sequence of controls that, for all possible unpredictable perturbations, drives
the system to the desired state?"
p.20 Adaptive Control
Of course, the general control problem is even harder. Usually one does not even know the dynamics
of the system that one wants to control well enough to program a computer to simulate them, let alone how to control
the system in the presence of perturbations.
To solve problems of control and stability, one needs a picture of the qualitative behavior of the
system. That is, for nonlinear systems, control requires insight into the nature of the system's dynamics.
p.20 An algorithm is a procedure for processing information. To control a nonlinear dynamical system,
an algorithm must embody a model for the system... For the algorithm to model the system successfully, it must be
an adaptive algorithm: to acquire information, one must learn.
An adaptive algorithm for control alters itself in response to information that it gets about
the system that it tries to control.
p.21 According to the particular model that an adaptive controller possesses for the system that it is to
control, some of the system's behavior is rule based, or regular, and some is not rule based, but irregular
and apparently random. When the controller adapts, it changes the algorithm that it uses to model the system, and thereby
changes what it regards as regular, and what it regards as irregular.
p.21 Since total information represents the trade-off between rule-based and random behavior,
it decreases monotonically as long as addition of extra rules to describe regularities of a system is more than compensated
for by a decrease in the system's apparent randomness. This property of total information means that a learning process that minimizes total information is in a sense optimal: arrival at a minimum of total information
implies that one has obtained a complete description of the predictable features of a system, and expressed this description
in the most compact form. In fact, the model for a system's behavior that minimizes total information is optimal
in a strict mathematical sense as well... To characterize and control our surroundings, we must identify the parts of the
world where order can be increased at the expense of disorder.
p.21 A system that is to control its environment successfully... must adapt by constructing models
that allow it to decide what information to get, and how to act on it.
