USE OF REINFORCEMENT LEARNING FOR MULTIFIDELITY OPTIMIZATION IN MULTIPLY-CONNECTED TRUST REGIONS

The expected increase of the air traffic and the emergence of new paradigms of air transportation induced by the evolution of the market demand are going to significantly worsen the community noise scenario of our cities. An accurate and timely prediction of the noise emission is, therefore, a mandatory aspect of the design process of any innovative air transport solution. Unfortunately, aeroacoustic simulations are extremely expensive and a popular approach to circumvent the problem is the development of surrogate models equally accurate yet computationally cheaper for the evaluation of the objective functions. This work deals with a multi-fidelity approach to surrogate modeling for the identification of optimal engine installation to maximise noise shielding. The paper explores the possibility to couple reinforcement learning techniques with a trust-region-based multi fidelity approach to meta modelling. Specifically, an accelerated variant of the well known Q-learning algorithm, namely the Evolutionary Q-learning (EVQL) is used to train a group of agents, each one deployed in one of a set of trust subregions, defined around the high-accuracy samples. The subregions of confidence are updated according to the exit directions of each agent. These directions are also used to enrich the set of high-accuracy samples by evaluating the meta model uncertainties along them in the untrusted regions of the domain.