Recently the double haploid (DH) lines are widely used in maize breeding. To derive them, haploids are first obtained from the heterozygous genotype, then haploids are doubled with colchicine or otherwise. Haploids are usually obtained by pollinating the source material with a special inducer. We use our inducer LHI, which gives 15% of haploid kernels. DH lines are recombinant homozygous forms in which the ancestor genes are randomly intermixed. In this respect, they are similar to recombinant inbred lines, but are derived in a shorter time. This makes DH lines a convenient tool for combining genet, and allows them to be used in various breeding and genetic aims. Our aim was to obtain a recombinant line in which as many as possible positive alleles from both parents would come together. Rf7 and Ku123 lines were chosen as ancestors. In first cycle, 41 DH lines were derived from Rf7xKu123 hybrid. Only one of them exceeded the best parent Rf7 for grain yield of first ear (113 versus 103 gm per plant). DH lines were tested in different years, but here and below we will give data 2017 for equability, because only this year all the cycles were tested simultaneously. The productivity will be estimated by the first ear, since this trait proved to be less variable than the total productivity. The second cycle was carried out in two variants. In first variant, the line rk-6 from first cycle was chosen. It was not the most productive, but it had the best combining ability in cross with Rf7. Hence, we concluded that it had accumulated a sufficient number of positive alleles from Ku123, and it needs additional alleles from Rf7. To add new alleles, a cross rk-6 x Rf7 was made and 49 DH lines were obtained from this cross. This operation leads not only to the addition of new positive alleles, but also to a partial loss of previously accumulated ones. With a single DH line would be a risk, but since there are many lines, it can be hoped that the best line will have fewer losses and more profit. As a result, 6 of 49 lines exceeded the Rf7 level, while the best of them had a productivity of 120 gm. For the second variant, the best line of the first cycle rk-5 was chosen. To add new positive alleles, it was crossed with F1 and 41 DH lines were derived from this cross. The result was similar to the first variant. The Rf7 level was exceeded in 9 of 49 lines with maximum value of 118 gm. For the third cycle, the line rk-148 from the first variant of the second cycle was chosen, because it had the best combining ability in cross with Rf7. For the same reason as before, it was crossed with Rf7, and 24 DH lines were obtained. 4 of them exceeded the level of Rf7, the best line rk-433 had productivity of 128 gm. Thus, for three cycles of dihaploid selection, the productivity of the first ear has increased by 25%, from 103 to 128 gm, on average by 8% per cycle. The distance between the best parent and F2 (137 gm) has passed by 3/4. At such rates, we can expect to overcome the level of F2 in another 1-2 cycles, which will mean a partially fixed heterosis.