Building surrogate models using trajectories of agents trained by Reinforcement Learning