1: \begin{abstract}

2: An additive noise channel is considered, in which the distribution

3: of the noise is nonparametric and unknown. The problem of learning

4: encoders and decoders based on noise samples is considered. For uncoded

5: communication systems, the problem of choosing a codebook and possibly

6: also a generalized minimal distance decoder (which is parameterized

7: by a covariance matrix) is addressed. High probability generalization

8: bounds for the error probability loss function, as well as for a hinge-type

9: surrogate loss function are provided. A stochastic-gradient based

10: alternating-minimization algorithm for the latter loss function is

11: proposed. In addition, a Gibbs-based algorithm that gradually expurgates

12: an initial codebook from codewords in order to obtain a smaller codebook

13: with improved error probability is proposed, and bounds on its average

14: empirical error and generalization error, as well as a high probability

15: generalization bound, are stated. Various experiments demonstrate

16: the performance of the proposed algorithms. For coded systems, the

17: problem of maximizing the mutual information between the input and

18: the output with respect to the input distribution is addressed, and

19: uniform convergence bounds for two different classes of input distributions

20: are obtained.

21: 

22: \looseness=-1

23: \end{abstract}