1: \begin{abstract}

2: One central theme in

3: machine learning is function estimation 

4: from sparse and noisy data.

5: An example is supervised learning where

6: the elements of the training set are couples, each containing 

7: an input location and an output response.

8: In the last decades, a substantial amount of work has been devoted to

9: design estimators for the unknown function and 

10: to study their convergence

11: to the optimal predictor, also characterizing the learning rate. 

12: These results typically rely on stationary assumptions where input locations are drawn from 

13: a probability distribution that does not change in time.

14: In this work, we consider kernel-based ridge regression and derive convergence conditions

15: under non stationary distributions, addressing also cases where 

16: stochastic adaption may happen infinitely often. 

17: This includes the important exploration-exploitation

18: problems where e.g. a set of agents/robots has to monitor an environment to reconstruct

19: a sensorial field and their movements rules are

20: continuosuly updated on the basis of the acquired knowledge on the field and/or the surrounding environment. 

21: \end{abstract}

22: