JSD is just symmetrized KL, it's the forward KL + reverse KL.

In reinforcement learning, usually what we want is to find the optimal action, i.e. action that maximizes the reward, this translates to the so-called "mode-seeking" optimization, which is the reverse KL.