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.

翻译：去中心化金融（DeFi）在过去几年中引起了极大的兴趣，涌现出多种类型的协议，例如借贷协议或自动做市商（AMMs）。这些协议通常通过链下治理进行控制，代币持有者可以投票修改协议的不同参数。然而，截至目前，这些参数的选择一直是一个手动过程，通常由协议核心团队完成。在这项工作中，我们对DeFi环境进行了建模，并提出了一种基于深度Q网络（DQN）强化学习的半自动参数调整方法。该系统自动生成直观的治理提案，并附带数据驱动的理由，以调整这些参数。我们的评估结果表明，与现有的手动方法相比，基于学习的链上治理过程更具反应性、客观性和高效性。