Researchers address liquidity placement in the Bitcoin Lightning Network by casting it as a budget-constrained combinatorial optimization problem. They propose MPFlow, a method that selects channel additions to maximize s-t max-flow using graph reinforcement learning.
- The agent combines a message-passing policy network with proximal policy optimization (PPO) and action masking.
- Training employs a hub-exclusion curriculum, removing top hubs to force capacity-aware placement rather than hub attachment.
- Experiments on real Lightning Network snapshots show consistent outperformance of strong heuristic baselines.
- The agent is deployed in production for peer recommendations, executing 4640 channel-open decisions allocating 267.3 BTC across 30 nodes.
The method provides a theory-grounded approach to routing capacity that has been validated through extensive real-world deployment.