Most classical population genetic inference methods assume that genome wide genetic diversity is mostly influenced by neutral processes (commonly named as “demography”) while selective processes affecting few isolated loci. Thus, changes in allele frequencies through time for a large number of loci can be used to infer the amount of drift (i.e. the effective population size,Ne) and the presence of some selection is assumed to have a negligible impact on that estimate. Loci under selection can then be identified as the ywill present allele frequency changes larger or more often in the same direction, than expected under pure drift. In some cases these assumptions might not hold. If the action of selection is widespread along the genome or recurrent in time, the high proportion of loci affected by selection and hitch-hiking could significantly bias the estimates of effective population size. Removing outlier loci would not solve this problem because only loci with large effects will be detected but manyother loci under weaker selection (e.g. selection on polygenic characters) could be present. We propose to make population genetic inference using models that include both the neutral and adaptive processes on the genome scale. This will allow to estimate demographic and selection parameters taking into account their interaction. Because these models are difficult to address under a likelihood framework we recourse to approximate Bayesian computation via Random Forest (ABC-RF). ABC-RF uses simulations to generate a training data set from which RF can learn and make inferences from real data.