1: \begin{abstract}

2:     We want to recover the regression function in the single-index

3:     model. Using an aggregation algorithm with local polynomial

4:     estimators, we answer in particular to the second part of

5:     Question~2 from \cite{stone82} on the optimal convergence

6:     rate. The procedure constructed here has strong adaptation

7:     properties: it adapts both to the smoothness of the link function

8:     and to the unknown index. Moreover, the procedure locally adapts

9:     to the distribution of the design. We propose new upper bounds for

10:     the local polynomial estimator (which are results of independent

11:     interest) that allows a fairly general design. The behavior of

12:     this algorithm is studied through numerical simulations. In

13:     particular, we show empirically that it improves strongly over

14:     empirical risk minimization.

15:   \end{abstract}

16: