Add the actor-critic algorithm

[s-sd]

Can do this only inside the update function with the help of an additional value function which learns to approximate rewards. First update value function using a simple loss, then update model.