BFTBrain: Adaptive BFT Consensus with Reinforcement Learning
summary
System adapts to live conditions by measuring metered features and using a decentralized deterministic RL agent on each node to agree on protocol changes, then switches protocols at a high watermark to preserve log correctness. It feeds throughput and fault features to the RL engine and selects among existing BFT protocols (PBFT, Ziza, Prime, CheapBFT, etc.). Evaluation shows 18% to 119% throughput gains over fixed protocols under dynamic conditions, converges in 1–6 minutes to within ~5% of the best protocol, and resists adversarial data pollution.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:21Application design planes
- 3:26Crash vs Byzantine faults
- 5:54Fixed BFT limits
- 8:42Need for adaptive BFT
- 10:24Reinforcement learning
- 11:45How BFTBrain works
- 14:11Results and performance
- 16:19Dynamic testing
- 17:01Conclusion & Q&A
related talks
15mAlea-BFT: Practical Asynchronous Byzantine Fault Tolerance
29mBuilding Probabilistic BFT Protocols Through Subsampled Voting
33mLibraBFTv2: Optimistically-linear BFT Consensus with Concrete Latency Bounds
17mThe Bedrock of Byzantine Fault Tolerance: A Unified Platform for BFT Protocols...
17mIn Link We Trust: BFT at the Speed of CFT using Switches
32mByteDance's Platform for Reinforcement Learning from Human Feedback
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
18mChitu: Avoiding Unnecessary Fallback in Byzantine Consensus