Action-bisimulation encoding improves representation learning for reinforcement agents by capturing multi-step controllability.
