Batch Reinforcement Learning for semi-active suspension control

Abstract

The object of this work is the design of a control strategy for semi-active suspension. In particular this paper explores the application of batch reinforcement learning (BRL) to the design problem of optimal comfort oriented semiactive suspension. BRL is an artificial intelligence technique able to provide an approximate solution of optimal control problems. The resulting control rule is a multidimensional relation which maps the measurable states of the system to the control action (reference damping). Recently a quasi optimal strategy for semi-active suspension has been designed and proposed: the Mixed SH-ADD algorithm, herein recalled for benchmarking purposes. This paper shows that an accurately tuned BRL provides a policy able to guarantee the overall best performances, which are paid in terms of complexity of both the training phase and the resulting control rationale.