“…Up to now, state-of-the-art approaches [4], [27] usually planned two to three shots in advance by discretizing the search space so as to try various angles and speeds at which to hit the cue ball, simulating the shot, and then continuing at the next ply from the position reached by the cue ball. What was proposed in [16], however, was to divide the problem into two layers, such that a robust controller may be guided by a high-level planner, taking advantage of possible knowledge that can be extracted from the domain. Although the controller described in that work proved to be robust, the downside was a necessity for many simulator calls for every single evaluation of the objective function.…”