Method of modelless online optimization of time calculation of finite IоT devices with provision of critically important delays for all levels of interaction

The problem of managing network resources considered in terms of convex online optimization (COO) based both losses and limitations. A general wording of the COO with long-term time constraints and relevant indicators for evaluating the COO algorithm are also provided. The solution to the optimization problem is based on elements of deep machine learning. This considers the COO with different time constraints that must be met in the long run. According to this parameter, the learner first performs a certain action, not knowing a priori neither contradictory losses nor time-varying constraints, which later appear naturally. Also, the standard loss-based COO structure was generalized to take into account both competitive losses and constraints. Unlike existing work, the focus is on a setting in which some limitations are detected after action is taken, they are acceptable for instantaneous violations, but on average must be met. Next, a modified method of dynamic optimization of network resources in intelligent networks (DONR) is introduced in this new COO structure, where the learning element deals with different time losses, as well as different time but long-term constraints. In general, it is proved that the DONR method ensures that dynamic losses and pre-results increase sublinewise, if the accumulated variations of minimizers and constraints known as sublinear increase. The process of development and characteristics of the developed DONR algorithm are described in detail.