Direct Control (DC)

In direct control schemes, the neural network itself acts as the controller. Many such schemes exists, including direct inverse control, optimal control, and feedforward control [13]. Direct control can only mimic the control done in the data that has been used for learning. It therefore requires examples of correct control aiming at the same goal. Equation (8) provides a prediction of the control signal $\mathbf{u}(t_0)$ based on the previous control signal $\mathbf{u}(t_0-1)$ and the previous estimate of the hidden state $\mathbf{s}(t_0-1)$. The prediction mapping is called the policy in Figure 1. A control method that we simply call direct control (DC), chooses the control signal by collapsing the inferred probability distribution $q(\mathbf{u}(t_0))$ to its expected value. When the control signal $\mathbf{u}(t_0)$ is selected and the observation $\mathbf{x}(t_0)$ is made, the two probability distribution collapse and these changes affect the estimates of the states $\mathbf{s}(t)$ that are then re-inferred. This works as the error feedback mechanism.