This paper investigates the optimization of phase shifts at intelligent reflecting surface (IRS)-assisted multiple input single output (MISO) systems with imperfect channel state information (CSI). By utilizing the channel statistical expressions, a closed-form ergodic achievable rate expression is derived by considering channel estimation errors of the base station (BS)-user channel, BS-IRS channel and IRS-user channel for semi-passive IRS systems in which only a portion of all IRS elements has been equipped with active sensors. We further propose to apply the genetic algorithm to tackle the system ergodic achievable rate maximization problem. Simulation results show the effectiveness of our proposed scheme.