An optimal control approach to mode generation in hybrid systems

This paper presents a solution to the problem of constructing control programs, i.e. sequences of control modes, from a given motion alphabet. In particular, techniques are developed that enable reinforcement learning to act directly at the mode level, and hence make learning applicable to continuous time control systems in a straightforward manner. Moreover, given such a control program, the issue of improving the system performance through the addition of new control laws is addressed as an optimal control problem. In fact, this is achieved through an optimal combination of recurring mode strings. A number of examples are provided that illustrate the viability of the proposed methods.