BFTBrain - Adaptive BFT Consensus with Reinforcement Learning

Learn about BFTBrain, a reinforcement learning-based Byzantine fault-tolerant system that adapts to changing network conditions and workloads in real-time through this 17-minute conference presentation from NSDI '25. Discover how this plug-and-play system switches between different BFT protocols dynamically based on systematic performance modeling and decentralized reinforcement learning coordination. Explore the key technical advances including comprehensive metrics modeling that correlates BFT protocol performance with fault scenarios and workloads, and examine how the decentralized RL approach remains resilient to adversarial data pollution while enabling nodes to reach consensus on learning outputs. Understand the significant performance improvements achieved, with throughput gains of 18% to 119% over fixed protocols under dynamic conditions and 44% to 154% improvements over existing learning-based approaches, making it suitable for diverse hardware and network configurations across various operational scenarios.