We present a machine learning (ML) approach for the accelerated extraction of optical properties from the experimental reflection electron energy loss spectroscopy (REELS) spectra by the reverse Monte Carlo (RMC) method. Taking platinum (Pt) as an example, experimental REELS spectra recorded at incident electron energies of 2.0 and 1.5 keV over an energy loss range of 0-200 eV are processed by one-dimensional convolutional neural networks trained specifically for each incident energy. The training processes incorporate the physical constraints based on the perfect-screening (ps-) and oscillator-strength (f-) sum rules of ELF into the loss functions to guarantee the validity of the results. The networks are trained on datasets generated by the RMC method, which pairs the simulated REELS spectra with the corresponding energy loss functions (ELFs). This approach enables the direct prediction of the ELF and optical constants from measured REELS spectra with reduced computational cost as compared to the previous RMC method. The predicted ELFs from ML exhibit high accuracy as verified by sum rules, and the optical properties of Pt are derived from these ELFs.