Loading...
 
[Show/Hide Left Column]

Sujets Help

View Item

Domain Machine Learning-Robotics
Domain - extra
Year 2010
Starting as soon as possible
Status Open
Subject Monte-Carlo Tree Search (MCTS) is a revolution in high-dimensional planning.
This ph.D. is devoted to:
(i) improvement of MCTS algorithm;
(ii) experiments in the Iomca platform.
Thesis advisor TEYTAUD Olivier
Co-advisors
Laboratory INRIA-Saclay TAO
Collaborations NUTN (Taiwan)
Artelys (www.artelys.com)
Contact: olivier.teytaud at inria.fr
Abstract The IOMCA project involves several partners and includes a platform for dynamic optimization.
- several artificial benchmarks will be included (some of them to be chosen/developed
by the ph.D. student)
- some real-world benchmarks developed by the company involved in the project (not to be
developped by the ph.D. student, but to be used in experiments).
- a MCTS implementation (to be developped and improved by the ph.D. student)
- other implementations (to be included in tests, but not developped by the ph.D. student)

The work will include tests on the very important (economically and ecologically) electricity production management problem.

Objectives:
- finding significant and generic improvements to the MCTS algorithm;
- testing MCTS versus other algorithms depending on the size of the problem; the energy
management problem is necessarily included in tests.

Contact: olivier.teytaud at inria.fr

Context Main focus of MCTS = high-dimensional planning when no evaluation function is available.

Many successful applications in games (in particular the game of Go http://www.lri.fr/~teytaud/mogo.html) , but also the important case of general game playing.

Some publications on difficult high-dimensional planning problems:
- industrial application published in ICML: http://hal.inria.fr/inria-00379523/
- application to active learning published in ECML: http://hal.inria.fr/inria-00433866/
- application to non-linear optimization in Algorithmica: http://hal.inria.fr/inria-00369788

Contact: olivier.teytaud at inria.fr
Objectives Objectives:
- finding significant and generic improvements to the MCTS algorithm;
- testing MCTS versus other algorithms depending on the size of the problem; the energy
management problem is necessarily included in tests.
Work program Understanding of the field of dynamic optimization.

Understanding of MCTS algorithms.

Writing of a state of the art.

Development of a MCTS implementation in the Iomca platform.

Finding new ideas for MCTS.

Comparing MCTS and other techniques (developped by other teams/persons) on artificial scalable benchmarks (developed by other teams/persons and/or by the ph.D. student).

Testing MCTS on energy management problems.
Extra information Contact: olivier.teytaud at inria.fr
Prerequisite Programming skills. C or java or C++ is ok.
Mathematical programming or statistics are a bonus.
Ability to work in a team.
Contact: olivier.teytaud at inria.fr
Details
Expected funding international ANR funding
Status of funding Confirmed
Candidates
user
Created Wednesday 10 of March, 2010 11:11:00 CET
LastModif Wednesday 10 of March, 2010 11:12:34 CET


Ecole Doctorale Informatique Paris-Sud


Directrice
Nicole Bidoit
Assistante
Stéphanie Druetta
Conseiller aux thèses
Dominique Gouyou-Beauchamps

ED 427 - Université Paris-Sud
UFR Sciences Orsay
Bat 650 - aile nord - 417
Tel : 01 69 15 63 19
Fax : 01 69 15 63 87
courriel: ed-info at lri.fr