Polar codes with large kernels can achieve improved error exponents but are challenging to design with low decoding complexity. This work investigates kernel construction under recursive maximum likelihood decoding (RMLD) using a reinforcement learning framework based on the Gumbel AlphaZero algorithm. The proposed method efficiently explores the design space and identifies large-size kernels that satisfy a given error exponent while minimizing decoding complexity. For a size-16 kernel, it achieves 17% lower decoding complexity than handcrafted designs while reaching an error exponent of 0.5183 compared to 0.5 for Arikan's kernel, demonstrating the effectiveness of the learning-based approach for practical polar code construction.