RL Agent Optimizes Bitcoin Lightning Network Liquidity
A new graph reinforcement learning approach maximizes routing capacity by strategically placing liquidity under fixed budget constraints.
Strategic liquidity placement in the Bitcoin Lightning Network determines how effectively a node can route payments. MPFlow uses graph reinforcement learning to solve this as a budget-constrained optimization problem, selecting the specific channel additions that maximize the network's max-flow routing capacity.
https://arxiv.org/abs/2607.08703
The system employs a message-passing policy network and proximal policy optimization. To ensure the agent learns actual capacity-aware placement rather than simply connecting to the largest existing hubs, the researchers used a hub-exclusion curriculum that removes top hubs from training subgraphs.
This approach has moved beyond simulation into production. The agent has executed 4,640 channel-open decisions across 30 managed nodes, allocating 267.3 BTC—valued at over $16 million—to optimize peer recommendations.
By outperforming standard heuristic baselines on unseen graphs, the work demonstrates that deep RL can manage the combinatorial complexity of blockchain liquidity more efficiently than manual or rule-based strategies.