DEVI seminar- Clément Bouttier

Date - Heure / Date - Hour
Date(s) - 12/10/2017
11h00 - 12h00

Emplacement / Location
Amphi Breguet, Building Breguet

Global stochastic optimization: stochastic and bandit algorithms with application to the stochastic flight planning problem.

In this talk, we will present the theoretical and numerical analysis of stochastic algorithms for the stochastic flight planning problem. Optimizing the fuel consumption and flight time is a key factor for airlines to be competitive. These companies thus look for flight optimization tools with higher and higher accuracy requirements. However, nowadays available methodologies for flight planning are based on simplified aircraft performance models. In this thesis, we propose to fulfill the accuracy requirements by adapting our methodology to both the constraints induced by the utilization of an industrial aircraft performance computation code and the consideration of the uncertainty about the real flight conditions, i.e., air traffic and weather conditions. Our proposal is supported by three main contributions. First, we design a numerical framework for benchmarking aircraft trajectory optimization tools. This provides us a unified testing procedure for all aircraft performance models. Second, we propose and study (both theoretically and numerically) two global derivative-free algorithms for stochastic-optimization problems. The first approach, the NSA algorithm, is highly generic and does not use any prior knowledge about the aircraft performance model. It is an extension of the simulated annealing algorithm adapted to noisy cost functions. We provide an upper bound on the convergence speed of NSA to globally optimal solutions. The second approach, the SPY algorithm, is a Lipschitz bandit algorithm derived from Piyavskii’s algorithm. It is more specific as it requires the knowledge of some Lipschitz regularity property around the optimum, but it is therefore far more efficient. We also provide a theoretical study of this algorithm through an upper bound on its simple regret.

Clément Bouttier, Airbus, Toulouse.