Model-free Learning to Avoid Constraint Violations: An Explicit Reference Governor Approach
Kaiwen Liu, Nan Li, Denise M. Rizzo, Emanuele Garone, Ilya V. Kolmanovsky, Anouck Girard · 2019
Constraints, including thermal, power, traction and rollover limits, as well as actuator range and rate limits, are ubiquitous in advanced ground vehicles and propulsion systems, and in their components, especially as these systems are downsized. These vehicles and systems will be operating in unknown environments where the recognition and avoidance of degradation or damage will be required. This paper proposes a model-free learning algorithm that over time modifies the parameters of an explicit reference governor (ERG) scheme so that violations of pre-specified constraints are avoided after a sufficiently informative learning phase. The ERG modifies setpoint commands to a nominal closed-loop system. Our learning algorithm modifies the ERG parameters based on observed constraint violations during a learning phase so as to eliminate constraint violations after learning is completed. Theoretical properties of the algorithm are analyzed and several examples that illustrate its effectiveness are presented.