Haupt-Navigation

Publikationsliste

Linear approaches to a stochastic mechanical control problem
Zitatschlüssel	Mabrouk:2010:LAS
Autor	Mahmoud Mabrouk
Jahr	2010
Schule	TU Berlin
Zusammenfassung	This thesis discusses a new method to linearize the Bellman equation for a special class of problems and tests its resulting algorithm with the state-of-the-art solutions. Reinforcement learning and Dynamic programming are presented and the state-of-the-art algorithms are discussed. The new framework and its mathematical foundations are then introduced. It results in a linear solution to the optimal action both in discrete and continuous domains, and in a new formulation of the cost-to-go function which exchanges the exhaustive search over actions with a linear solution. Later, an online and an oﬄine algorithm are developed from the last results. They are tested against Policy Iteration and Q-Learning in a stochastic variant of the Mountain car problem. Results show a great improvement brought by the new algorithms both in speed and eﬃciency. Last, the limitations of the new framework are discussed.
Typ der Publikation	Bachelor Thesis