Multi-unit optimization is a relatively new method for real-time optimization of dynamic systems. It assumes the availability of multiple identical units, operates them with a fixed offset between their inputs, computes the gradient by finite difference and pushes the gradient to zero. The two main problems with this scheme are: (i) the units that are physically available may not be identical and (ii) the optimization problem may have a constraints, wherein more than just gradient control is required. In this talk, two methodologies that can be used to correct for the differences between the various units, will be discussed. Also, the effect of the difference in dynamics will be studied. In addition, an adaptation of Rosen's projection algorithm will be proposed to handle constraints. It is shown that the proposed adaptation avoids jamming, i.e. getting stuck on an non-optimal solution.
Group for Research in Decision Analysis