Robust Bayesian reinforcement learning through tight lower bounds