Learning sequential visual attention control through dynamic state space discretization