How to get gradients w.r.t the reward value prediction error ?

[shamanez]

Here we predict the reward value by multiplying immediate state representations with reward prediction vector W. When getting the mean squared error w.r.t above loss we get gradients w.r.t reward vector prediction variables right ?