In order to overcome the complexity and variability of underwater working environments, as well as the difficulty of controlling the flapping motion due to the significant nonlinear characteristics and numerous variables involved, a direct exploration approach is proposed to search for the optimal flapping foil propulsion parameters in the environment. The Latin hypercube sampling technique is utilized to obtain the samples of multi-dimensional flapping parameters in actual water pool data, and a Gaussian process regression (GPR) machine learning model is established based on these samples to generalize the working environment. Under different propulsion performance requirements, the TD3 algorithm in deep reinforcement learning (DRL) is trained for maximizing rewards and obtaining the optimal combination of multiple parameter actions in continuous intervals. The experimental results demonstrate that the GPR-TD3 method is capable of learning the globally optimal solution for flapping propulsion in the experimental environment, including maximum speed and maximum efficiency. Furthermore, the accuracy of this optimal solution can be intuitively verified through a two-dimensional contour plot in the GPR. Meanwhile, with 290 sets of real samples provided for any given propulsion speed requirement, the agent can recommend a set of action combinations with an error range of 0.23% to 6.68%, which can provide reference for practical applications.