Reinforcement Learning (RL) constitutes an attractive alternative for the design of control algorithms for dynamical systems. During recent years, deep RL has been successfully applied to the field of computational fluid dynamics (RL-CFD). Among others, the disciplines of active flow control [1] and turbulence modeling [2] have benefited from these tuning techniques. Structurally, RL-CFD tuning problems feature two main elements: a tunable CFD solver, and an overarching program that manages the RL loop. Lattice Boltzmann methods have established themselves as efficient, versatile CFD solvers, thanks to their locality –thus straightforward parallelization and efficiency– and their inherent conservation properties. The multiple-relaxation-time (MRT) branch of LBM solvers is of special interest, enhancing the stability and/or the accuracy of LB simulations, compared to baseline single-relaxation-time solvers. Although offline techniques have been successfully applied to LBM parameter tuning [3, 4], RL has been barely applied to LBM [5]. Here, we study the potential of using deep RL techniques to tune the free parameters of a MRT LB solver, aiming for an increased accuracy and/or stability, and comparing the results with existing, offline techniques [3, 6].