Initialisation for Control

For nonlinear control tasks it is important that the initial estimate for control is good, otherwise the optimisation algorithm may get stuck in a local minimum or fail to converge in reasonable time. In many cases, the control signal from the previous time step can be used with quite good results. However, when a new control task starts or the goal is changed, or there are unexcepted changes in the system state, the previous control signal is often a poor choice.

A second option is to use random initialisations. If multiple different initialisations are used, this can be more robust than the first option. Unfortunately this is also much more time consuming because multiple control strategies must be computed for a single time step. If an internal forward model is available, a third option is also possible. The current system state can be propagated forward in time and the control signal from these predictions can be used as an initialisation.