Random Shooting MPC

Topics: Model Predictive Control

Concept

Random Shooting is a simple sampling-based model predictive control method where the controller randomly generates many candidate control sequences, simulates the system's future states for each sequence using the predictive model, evaluates their performance via the cost function, and picks the single sequence that yields the lowest cost. The sampling method for the controls is typically uniform or Gaussian sampling.
It is basically the Monte Carlo approach for solving the optimization problem of MPC.

The advantages of random shooting MPC are:

• very simple and easy to implement
• does not require gradients nor solving the optimization problem
• works with nonlinear dynamics
• works with non-convex cost functions
• parallelizable because each sampled trajectory is independent

Its main disadvantage is the "curse of dimensionality". As the control dimension or the planning horizon grows, it requires many more samples to find a good solution. And, since it is a sampling-based method, there is no guarantee of finding the optimal trajectory.