Optimizing low-energy anti-fibrillation pacing: Lessons from a cellular automaton model