1: \begin{abstract}

2: The Adaptive Multiple Importance Sampling

3: (AMIS) algorithm is aimed at an optimal recycling of past simulations in an iterated importance sampling scheme. The difference with earlier adaptive

4: importance sampling implementations like Population Monte Carlo is that the importance weights of all 

5: simulated values, past as well as present, are 

6: recomputed at each iteration, following the technique of the deterministic multiple mixture estimator 

7: of Owen and Zhou (2000). Although the convergence

8: properties of the algorithm cannot be investigated, we demonstrate through a challenging banana shape target distribution

9: and a population genetics example that the improvement brought by this technique is substantial.

10: 

11: \noindent {\bf Keywords:} adaptive importance sampling, sequential Monte Carlo,

12: population Mon\-te Carlo,

13: particle filters,

14: deterministic mixture weights,

15: banana shape target,

16: population genetics.

17: \end{abstract}

18: