Accurate channel estimation and allocation are vital in the provision of reconfigurable intelligent surfaces (RIS)-aided wireless network services to mobile users. Typically, channel estimation is carried out using a pilot signal. However, RIS elements cannot transmit or receive pilot signals because they are passive elements. Therefore, to maximize the gain of using the RIS, it is essential to accurately estimate a cascaded channel using a pilot signal between a base station (BS) and a terminal through an RIS. Moreover, although using a large number of pilot signals can guarantee accurate channel estimation performance, this can also drastically lower the wireless communication system's efficiency. Thus, in this letter, a new paradigm for learning-based pilot allocation and channel estimation in RIS systems is proposed. A masked autoencoder (MAE) is trained to achieve high channel estimation accuracy with a limited number of pilots. Then, a deep reinforcement learning(DRL) agent learns pilot allocation policies through MAE. Simulation results show that the MAE channel estimator has almost the same channel estimation performance even though it uses up to 33% fewer pilots than the autoencoder (AE)-based channel estimator. Furthermore, the proposed DRL-based pilot optimization method achieves higher channel estimation performance with 20% fewer pilots than the general autoencoder and other learning algorithms without the proposed RL-based pilot optimization algorithm.