Using chains of bottleneck transitions to decompose and solve reinforcement learning tasks with hidden states