On-line techniques for behavioral programming

Many control problems of interest can be cast as an optimal hybrid system control problems, wherein an objective function represents some global goals and the input at each time instant is a choice among a finite set of control laws. There are many approaches to solving such problems in the literature, all based on dynamic programming in some form or another, and all suffering from overwhelming computational complexity. Herein, we attempt to lower this complexity by examining techniques that take advantage of the underlying properties of the individual controllers among which we are switching. We call this process "behavioral programming," since we are now attempting to perform dynamic programming at the more abstract level of behaviors of the constituent systems. We present our paradigm and discuss two arenas of its use: motion planning for autonomous agents and LQR with state and input constraints. Applications to helicopter and wheel slip control are used to illustrate problem-solving in each of these arenas, respectively.