Due to the complexity of many practical applications, we encounter optimization problems with nonsmooth functions, that is, functions which are not continuously differentiable everywhere. Classical gradient-based methods are not applicable to solve such problems, since they may fail in the nonsmooth setting. Therefore, it is imperative to develop numerical methods specifically designed for nonsmooth optimization. To date, bundle methods are considered to be the most efficient and reliable general purpose solvers for this type of problems. 

The idea in bundle methods is to approximate the subdifferential of the objective function by a bundle of subgradients. This information is then used to build a model for the objective. However, this model is typically convex and, due to this, it may be inaccurate and unable to adequately reflect the behaviour of the objective function in the nonconvex case. These circumstances motivate to design new bundle methods based on nonconvex models of the objective function. 

In this dissertation, the main focus is on nonsmooth DC optimization that constitutes an important and broad subclass of nonconvex optimization problems. A DC function can be presented as a difference of two convex functions. Thus, we can obtain a model that utilizes explicitly both the convexity and concavity of the objective by approximating separately the convex and concave parts. This way we end up with a nonconvex DC model describing the problem more accurately than the convex one. Based on the new DC model we introduce three different bundle methods. Two of them are designed for unconstrained DC optimization and the third one is capable of solving also multiobjective and constrained DC problems. The finite convergence is proved for each method. The numerical results demonstrate the efficiency of the methods and show the benefits obtained from the utilization of the DC decomposition.