1: \begin{abstract}

2: This paper presents a new algorithm to perform regression

3: estimation, in both the inductive and transductive setting. The

4: estimator is defined as a linear combination of functions in a given

5: dictionary. Coefficients of the combinations are computed

6: sequentially using projection on some simple sets. These sets are

7: defined as confidence regions provided by a deviation (PAC)

8: inequality on an estimator in one-dimensional models. We prove that

9: every projection the algorithm actually improves the performance of

10: the estimator. We give all the estimators and results at first in

11: the inductive case, where the algorithm requires the knowledge of

12: the distribution of the design, and then in the transductive case,

13: which seems a more natural application for this algorithm as we do

14: not need particular information on the distribution of the design in

15: this case. We finally show a connection with oracle inequalities,

16: making us able to prove that the estimator reaches minimax rates of

17: convergence in Sobolev and Besov spaces.

18: \end{abstract}

19: