Seminars

Adaptive control has been used to provide closed-loop tracking guarantees for dynamic systems with unknown parameters. In such cases, estimates of unknown parameters are adapted over time and these estimates are used within a judiciously prescribed control law. As part of the adaptation, adaptation gains can be increased to accelerate the rate of convergence of these estimates to their true values. In prac¬tice, however, it is not prudent to simply increase learning gains as this can have the adverse effect of amplifying any unmodeled dynamics in a true system. As such, it has been the goal of adaptive control practitioners to design adaptive control structures which provide better performance for equivalent control gains. The Multi-Thread Attracting Manifold (MTAM) adaptive control methodology developed in this work seeks to provide an improved adaptive control framework through multiple design features. The first feature is the utilization of multiple threads. This allows for the unknown parameter space to be robustly sampled, rather than relying on a single estimate. Secondly, the adaptation of each individual thread is designed with an attracting manifold design which provides desirable "no-regret" learning. These features combine to provide an adaptive controller which outperforms existing single thread adaptive controllers with identical learning gains. In this work, the baseline method¬ology behind MTAM is presented first for classical adaptive control and parameter identification scenarios. Secondly, forms of MTAM which accommodate uncertainties attached to the control input in both direct adaptive control and indirect adaptive control cases are presented and shown to retain performance benefits through numer¬ical simulation. Finally, a methodology for adding and removing threads is presented which decreases the marginal computational cost of the MTAM algorithm while still providing performance benefits over existing single thread architectures.