Exact Reduction of Huge Action Spaces in General Reinforcement Learning