Jyri Leskinen: Developing effective gradient-free methods for inverse
problems

An inverse problem is a task where the properties of a system are
reconstructed using information such as physical measurements on the
system. The information available is often insufficient which leads to
a so-called ill-posed problem. Without additional information the
resulting reconstruction may not be accurate. Using gradient-based
methods combined with a priori information one can reconstruct the
correct properties of the system. However, this requires that the
gradient of the objective function is available and that it leads to
the global optimum (correct
solution) which restricts the usability of these methods. However,
advanced hybrid evolutionary methods such as memetic algorithms
provide effective tools to solve inverse problems without the need of
gradients.
As an example of solving an inverse problem, the reconstruction of an
Electrical Impedance Tomography (EIT) image using adaptive
differential evolution based memetic algorithm is illustrated.