Dispersion scan is a self-referenced measurement technique for ultrashort pulses. Similar to frequencyresolved optical gating, the dispersion scan technique records the dependence of nonlinearly generated spectra as a function of a parameter. For the two mentioned techniques, these parameters are the delay and the dispersion, respectively. While dispersion scan seems to offer a number of potential advantages over other characterization methods, in particular for measuring few-cycle pulses, retrieval of the spectral phase from the measured traces has so far mostly relied on the Nelder-Mead algorithm, which has a tendency of stagnation in a local minimum and may produce ghost satellites in the retrieval of pulses with complex spectra. We evaluate three different strategies to overcome these retrieval problems, namely regularization, use of a generalized-projections algorithm, and an evolutionary retrieval algorithm. While all these measures are found to improve the precision and convergence of dispersion scan retrieval, differential evolution is found to provide best performance, enabling the near-perfect retrieval of the phase of complex supercontinuum pulses within less than ten seconds, even in the presence of strong detection noise and limited phase-matching bandwidth of the nonlinear process.