Policy-based Control

\[\newcommand{\vect}[1]{\boldsymbol{#1}} \newcommand{\dvect}[1]{\dot{\boldsymbol{#1}}} \newcommand{\ddvect}[1]{\ddot{\boldsymbol{#1}}} \newcommand{\mat}[1]{\boldsymbol{#1}}\]

Policy based controllers do not aim to follow a singular trajectory. Instead, they define a funtion(the policy) \(\pi\) over the entire state space and return a control signal according to this policy:

\[\vect{u} = \pi(\vect{x})\]

the policy can be either analytically defined by reasoning about the system (e.g. about the system’s energy) or it can be learned with reinforcement learning.

• Linear Quadratic Regulator (LQR)
• Partial Feedback Linearization (PFL)
• Soft Actor Critic (SAC)
• Deep-Q Learning (DQN)