Approximate Bayesian computation is a family of computational techniques in Bayesian statistics. These techniques allow to fit a model to data without relying on the computation of the model likelihood. A number of improvements to the original ABC scheme have been proposed to speed it up. They include the coupling to Markov chain Monte Carlo to explore the parameter space and the sequential improvement of posterior distributions inspired by sequential Monte Carlo methods. These algorithms present severals shortcomings which are particularly problematic for costly to simulate complex models. In this article an adaptive appoximate Bayesian computation for complex models is proposed which offers improvements to circumvent the limitations of the application of the ABC methods to complex models. We demonstrate our algorithm on a toy example and on a complex social model, the PRIMA model.