Learning-based model predictive control (LBMPC) is a new technique that rigorously combines statistics and learning with control engineering, while providing levels of guarantees about safety, robustness, and convergence. It achieves this by decoupling the notion of safety and robustness from the notion of performance. An approximate model of a system with knowledge of its uncertainty provides safety, and statistical and learning methods refine the model and improve performance.
A useful analogy for describing what LBMPC does is the difference between driving a normal car and a sports car. When a regular driver uses a sports car, they use their mental model of a normal car to make sure they do not crash. As they spend more time in the sports car, they improve their mental model of the car and can drive it more precisely.
LBMPC safely improves the operational characteristics of different engineered systems, and it has been applied to energy-efficient building automation and high-performance semi-autonomous systems. An implementation in C++ source code is available from lbmpc.bitbucket.org.
LBMPC identifies physical phenomena like ground effect and uses this to refine the model of a quadrotor helicopter, which improves flight performance while providing robustness and safety. This enables the precise movements necessary for tasks such as catching a ball.
LBMPC uses statistical methods to estimate heating load due to occupants and equipment in a room using only temperature measurements. This allows it to compensate for these highly varying effects and leads to significant reductions in energy consumption for heating, ventilation, and air-conditioning (HVAC) systems on experimental testbeds.
Berkeley Retrofitted and Inexpensive HVAC Testbed for Energy Efficiency (BRITE)