Homepage --- Olivier Teytaud --- TA
Responsible of the TAO-UCTSIG
0169156607

Saclay (unfortunately)

Here: selected publications and realizations.

Publications:

video sur UCT/MCTS en francais (seminaire Le Modele et L'algorithme, Inria)
Publications (since October 2004):
- Complete list since 2005 (one can also see my old publications from my ph.D.)
- International publications (HAL)
- International peer-reviewed journals (HAL) (for some unknown reason, HAL does not find my publication in physics)
- International peer-reviewed conferences
I maintain popularization papers here. This is made both for scientists interested in having a first idea of my research, and for all people willing to have more information. The pages dedicated to research are here and a complete list is here. L'Inria, le CNRS et l'universite n'ont aucune responsabilite quant a ce site exterieur.

Downloads:

gsll/gslp: two open source programs (with libsvm/liblinear compliant formats) for linear regression (added June 2009).
librks: an open source program (with gsll/libsvm/liblinear compliant formats) for random kitchen sinks (added June 2009).
un script pour construire une liste de publications en html depuis HAL (remplacer Olivier par votre prenom et Teytaud par votre nom et c'est bon

Long version here
.

Scientific claims:

At last some wins as black in 9x9 Go komi 7.5 against top players.
Bernstein races are great for adaptive noisy optimization.
The optimal speed-up of many optimization algorithms is tightly known thanks to this paper .
Many algorithms have suboptimal speed-up and simple modifications can correct this for parallel optimization
(randomized) quasi-random works for (i) integration (ii) active learning (iii) optimization (iv) robust optimization (v) discrete sampling (vi) covariance matrix adaptation
high dimensional comparison-based optimization is slow
choose mu=lambda/2 for increasing the convergence rate in ES in many cases (algorithms which do not use the complete algorithms); but for lambda large and for many classical algorithms mu = dimension is good.
the population size should increase logarithmically with the number of iterations in evolution strategies for monomodal non-convex optimization
highly multi-objective comparison-based optimization is very very slow
comparison-based optimization is robust
UCB-like formulas work for computer-go but only if your formula is not very well optimized (in that case there's much better)
randomness is necessary in (i) robust optimization and (ii) active learning
genetic programming requires bloat penalization
genetic algorithms can program better than humans in computer-Go
NFL-theorems are false in continuous domains

Scientific policy claims:

Events:

Board games in Academia: I'm not an organizer of it, but as this symposium looks really interesting and as it is seemingly unsufficiently well done I try to advertise it.
ThRaSH : Theory of randomized search heuristics 2010 in Paris!
WCCI: humans vs computers in Go in Barcelona
Workshop on emergent applications of Computer Go at IEEE Fuzz 2010
Le jeu de Go et la recherche en intelligence artificielle: Sciences en Fete 2009
The "type I and type II errors in multiple simultaneous hypothesis testing" workshop

A titre personnel, n'engageant pas le Lri, l'Inria ou le Cnrs, je maintiens divers sites dont celui-ci , celui-ci et celui-la, ainsi que les blogs-pages qui suivent: