The time average expected age of information (AoI) is studied for status updates sent over an error-prone channel from an energy-harvesting transmitter with a finite-capacity battery. Energy cost of sensing new status updates is taken into account as well as the transmission energy cost better capturing practical systems. The optimal scheduling policy is first studied under the hybrid automatic repeat request (HARQ) protocol when the channel and energy harvesting statistics are known, and the existence of a threshold-based optimal policy is shown. For the case of unknown environments, average-cost reinforcement-learning algorithms are proposed that learn the system parameters and the status update policy in real-time. The effectiveness of the proposed methods is demonstrated through numerical results.