Ultra-high-performance fiber reinforced concretes (UHPFRC) have become a construction material of great interest within the scientific community. The capacity of UHPFRCs to present a strain-hardening behavior under direct tensile test has led to it being increasingly included in the construction and rehabilitation of structures. However, its high cost and complexity in its production has limited its use in Colombian industry. Therefore, it is necessary to reduce the production cost of UHPFRC, both in raw materials and in development. In this research, machine learning algorithms such as Random Forest are carried out to predict the parameters of the UHPFRC's direct tensile behavior: g, energy absorption capacity (measured in kJ/m3) and εpc, strain under maximum post-cracking load (measured in %). Knowledge of these parameters is essential for applications that require high ductility, such as the rehabilitation and seismic retrofitting of existing non-ductile structures. Once the predictive models were developed and validated statistically and experimentally, a multi-objective optimization algorithm was used to determine the blend and fiber content that, using an optimized cementitious matrix dosage, would achieve the ductility requirements necessary for the seismic retrofitting applications of structures. (g ≥50 kJ/m3 and εpc≥0.3%) at the lowest cost. The results of this master's thesis showed that a mixture of UHPFRC with a hybrid mixture of 0.32% high strength steel fiber and 1.52% hook-shaped normal strength steel fiber by volume (1.90% of total volume fiber fraction) achieved the ductility criteria described at the lowest cost. As a conclusion, it was found that the Random Forest algorithms turned out to be an effective tool for predicting the behavior of the UHPFRC, allowing to reduce costs and research times in the development of new dosages.