|Ph.D Student||Dobrovinsky Maya|
|Subject||Cruise Flight Throttle Optimization by Neuro-Dynamic|
|Department||Department of Aerospace Engineering||Supervisor||Professor Yoseph Ben-Asher|
|Full Thesis text|
During commercial flights (and also in many military
operations) the cruise phase represents the largest part of the flight. The
efficiency of the cruise flight is, therefore, especially important. Cruise
optimization at constant altitude problem is studied in the thesis.
The problem can be formulated as a singular optimal control problem. Analytical solution of this problem (optimal throttle control in the closed loop) is very complicated, and includes singular arcs finding.
Consequently, this thesis focuses on developing numerical solutions for the throttle control system based on the principle of Reinforcement Learning. The system learns on-line from its own mistakes through the reinforcement signal from the environment and tries to reinforce its action to improve future performance.
The full problem is decomposed into three sub-problems: finding the optimal cruise speed, reaching the optimal cruise speed from non-optimal conditions and maintaining the optimal cruise speed.
The sub-problems can be solved by Dynamic Programming but due to the so-called “curse of dimensionality” approximate dynamic programming methods are motivated.
The throttle control system implementation by Approximate Dynamic-Programming via a Reinforcement Learning mechanism based on SARSA or Q-learning will render the solution feasible for practical use in the aviation industry.
Even a small percentage in fuel saving might be of a significant value from the economic point of view.