Imperial College London
Tuesday, October 18, 2016, 11:30
Room 01-014, Georges-Köhler-Allee 103, Freiburg 79110, Germany
Conventional Model Predictive Controller (MPC) design approaches propose developing an algorithm at a high level of abstraction, followed by a low level hardware implementation. This decoupled approach may lead to suboptimal closed-loop performance, since the interaction between the software and hardware layers is not taken into account. Moreover, in practical applications performance is often traded off against computational platform resource usage, namely time, energy and space.
In order to efficiently explore these trade-offs we formulate MPC design as a multi-objective optimization (MOO) problem. We propose an algorithmic solution, for the generation of Pareto optimal MPC designs, that is based on Gaussian modelling and a hypervolume criterion described in. A numerical study of a fast gradient-based MPC controller confirms that the proposed approach allows avoiding a large number of suboptimal implementations on a field-programmable gate array.