The onset of plastic strain localization in metallic materials depends on the deformation history. This history dependence has crucial role when attempting to predict localization in steel structure in the event of accidental crash on the account of infinite number of loading pathways material point can experience prior to failure. Therefore, establishing a criterion for localization experimentally is beyond the realm of feasible mechanical testing capabilities. To address this we introduced a numerically efficient physics based model to inflict arbitrary loading history on the material point. The physics based model is used to generate database of bi-linear loading paths until localization, which is used to train machine learning (ML) model. The data driven ML model is demonstrated to correctly predict localization under bi-linear loading, but also for more complex loading history. The advantages and disadvantages of different ML models is discussed in the context of history prediction.