Learning Continuous Control Policies by Stochastic Value Gradients