Current students


Section: Systems and Control

Major Research topic:
Black-box optimization techniques for (semi-)automatic control calibration in vehicle applications

This PhD project investigates frameworks for optimal calibration in vehicle applications. The problem is formulated as a black-box global optimization task that aims at minimizing an objective function (OF). To tackle this issue data-driven methods have gain particular interest in the last decades. Their idea relies on available measured data to spot connections and correlations among system’s variables, without the need of physical understanding. This class of methods is commonly adopted for calibration purposes, both for real system and simulation model. However, the real implementation of such calibration techniques might arise some issues that are negligible in simulation.  

The first issue under consideration appears when dealing with optimization problems that aims at minimizing a known OF affecting the closed-loop performance of the system. This global optimization problem is typically tackled through a well-known tool: bayesian optimization (BO). This algorithm requires an exploration phase, where samples within parameters’ space are mapped as function of the returned OF observation, then it focuses on the parameters’ sub-space where the global optimum lays with most probability, referred to as exploitation phase. The problem arises for those systems where certain calibration tunings affect the closed-loop stability, leading to an unstable system response. This behavior doesn’t need to be addressed in simulation environment, however, when dealing with the real implementation the unsafe parameters during the optimization process lead to safety-critical system failures that can cause serious damage to the system and its environment. This occurs because the exploration phase of BO algorithm doesn’t penalize samples that makes the closed-loop system unstable. This work investigates a solution to this issue that incorporates safety margins to limit the amount of unstable explored tunings. This optimization tool, named Safe-BO, is developed and employed for the automatic calibration of control parameters in an electric vehicle transmission system.                                                                                                                              
The second addressed question deals with those black-box optimization problems where the OF is hardly quantifiable, either because it is of qualitative nature or because it involves several goals. The pursued idea relies that sometimes the goodness of a certain combination of decision variables can only be assessed by a human decision maker. The calibration problem is here formulated as a preference-based optimization and solved via Active Preference Learning (APL). This class of problems are characterized by completely unknown OF that cannot be evaluated, and the decision-maker is only able to express a preference such as “this is better than that” between two calibration candidates. This strategy drives a trial-and-error procedure, where the APL algorithm actively proposes pair-wise comparisons to the decision-maker, which, in turn, expresses its subjective preference. In this work, we propose a purely model-free comfort-oriented calibration strategy for vehicle suspensions based on APL. The implementation of such methodology and the experimental results have been gathered in a journal paper (submitted).