Dynamic Software Product Lines (DSPLs) are a well-accepted approach for self-adapting Cyber-Physical Systems (CPSs) at run-time. The DSPL approaches make decisions supported by performance models, which capture system features' contribution to one or more optimization goals. Combining performance models with Multi-Objectives Evolutionary Algorithms (MOEAs) as decision-making mechanisms is common in DSPLs. However, MOEAs algorithms start solving the optimization problem from a randomly selected population, not finding good configurations fast enough after a context change, requiring too many resources so scarce in CPSs. Also, the DSPL engineer must deal with the hardware and software particularities of the target platform in each CPS deployment. And although each system instantiation has to solve a similar optimization problem of the DSPL, it does not take advantage of experiences gained in similar CPS. Transfer learning aims at improving the efficiency of systems by sharing the previously acquired knowledge and applying it to similar systems. In this work, we analyze the benefits of transfer learning in the context of DSPL and MOEAs testing on 8 feature models with synthetic performance models. Results are good enough, showing that transfer learning solutions dominate up to 71% of the non-transfer learning ones for similar DSPL.