Decentralized finance (DeFi) has seen a tremendous increase in interest in the past years with many types of protocols, such as lending protocols or automated market-makers (AMMs) These protocols are typically controlled using off-chain governance, where token holders can vote to modify different parameters of the protocol. Up till now, however, choosing these parameters has been a manual process, typically done by the core team behind the protocol. In this work, we model a DeFi environment and propose a semi-automatic parameter adjustment approach with deep Q-network (DQN) reinforcement learning. Our system automatically generates intuitive governance proposals to adjust these parameters with data-driven justifications. Our evaluation results demonstrate that a learning-based on-chain governance procedure is more reactive, objective, and efficient than the existing manual approach.