Technical Summary: MPFlow

Problem Formulation

The paper addresses the liquidity placement problem within the Bitcoin Lightning Network (LN). Given a fixed budget to open or top-up $k$ payment channels, the objective is to determine which peers a node should connect to in order to maximize its routing capacity.

The authors frame this as a budget-constrained combinatorial optimization problem on a graph. Unlike existing industry practices that rely on static graph heuristics (e.g., degree or betweenness centrality), which ignore directed balances and budgeted interventions, this work seeks to select $k$ edge additions that maximize the $s$–$t$ max-flow on the observed capacity graph. The max-flow metric is chosen as a theory-grounded proxy for routing capacity, as it upper-bounds the deliverable payment volume and is computationally efficient to recompute (via push-relabel algorithms) during training.

The problem is modeled as a finite-horizon Markov Decision Process (MDP):

• State ($S_t$): Encodes the current graph topology, including node features (centrality metrics) and edge features (fees, capacities).
• Action ($A_t$): Selecting a node to open a channel with (or top-up an existing one).
• Reward ($R_t$): The marginal improvement in absolute max-flow ($r_t = F_t - F_{t-1}$), where $F_t$ is the sum of max-flows from a source to a set of target nodes.

Methodology: MPFlow

The proposed solution, MPFlow, combines a Message-Passing Neural Network (MPNN) with Proximal Policy Optimization (PPO) and action masking.

1. Agent Architecture

• Message-Passing Backbone: The agent utilizes a permutation-invariant MPNN. Unlike standard MPNNs that use mean or sum aggregation, MPFlow employs a max-aggregation operator.
  • Rationale: Max aggregation emphasizes the single most constraining neighbor, mimicking bottleneck detection along min-cut edges. This allows the policy to focus on locally binding capacity constraints rather than diffuse averages.
  • Input Features: Node features include PageRank, capacity ratio, normalized degree, and local clustering coefficient. Edge features include base fee, fee rate, and capacity (log-transformed).
• Actor-Critic Heads:
  • Actor: Maps node embeddings to logits, applying a softmax over visible nodes to generate a stochastic policy.
  • Critic: Uses global max pooling over node embeddings before a linear layer to estimate state value. This focuses the value estimate on the most critical graph substructures (bottlenecks) rather than an average of all nodes.
• Action Masking: A binary mask ensures the agent only selects operationally feasible actions (excluding self-edges, disconnected nodes, or already-visited peers), keeping exploration within the deployable control surface.

2. Training Protocol

• Optimizer: The agent is trained using PPO (Proximal Policy Optimization), chosen for its stability and sample efficiency compared to A2C or DDPG in preliminary experiments.
• Hub-Exclusion Curriculum: A critical component of the training strategy. The agent is trained on subgraphs where the top-50 highest-degree nodes (hubs) are removed.
  • Goal: This prevents the policy from learning a degenerate strategy of simply attaching to the largest hubs (which centrality heuristics already do). It forces the policy to learn capacity-aware placement and bottleneck relief that remains valid even when hubs are present during evaluation.
• Budget: The agent executes $k=5$ actions per episode, with each action representing a channel opening of fixed size (0.20 BTC).

Key Contributions

1. Optimization-Aligned RL Formulation: The paper casts liquidity placement as a graph decision process where the reward is defined by the marginal improvement in max-flow, aligning the learning objective directly with structural routing capacity.
2. Simple, Deployable Policy: The introduction of a compact MPNN–PPO agent featuring max-aggregation and action masking, designed specifically for the heavy-tailed, heterogeneous nature of the Lightning Network.
3. Hub-Exclusion Curriculum: The demonstration that training on subgraphs excluding top hubs forces the policy to learn bottleneck-aware structures, resulting in a policy that allocates liquidity largely independently of top hubs and outperforms variants trained with hubs visible.
4. Empirical Validation on Real Topology: Extensive evaluation on real Lightning Network snapshots (May, July, October 2025) showing consistent improvements over strong heuristic baselines.
5. Practical Deployment: The approach has been deployed in production to inform peer recommendations.

Experimental Results

Experiments were conducted on 5,000-node subgraphs derived from real LN snapshots, using 1,000 paired episodes for evaluation.

• Performance vs. Baselines: MPFlow achieved an absolute max-flow gain of 0.168 ± 0.003 BTC, significantly outperforming the strongest heuristic baseline, Betweenness Centrality (0.163 ± 0.003 BTC).
  • Paired Uplift: MPFlow showed a statistically significant paired uplift of +8.59% over Betweenness.
  • Win Rate: MPFlow won 62.3% of paired episodes against Betweenness.
• Ablation Studies:
  • Graph Scale: MPFlow matched or exceeded Betweenness across graph sizes (1k to 5k nodes).
  • Hub Removal: As top hubs were pruned from the evaluation graph, MPFlow's advantage over Betweenness widened (up to +8.89% relative improvement with 25 hubs removed), confirming the policy relies on bottleneck relief rather than hub chasing.
  • Action Distribution: MPFlow allocated only 17% of its actions to the top-50 hubs (median rank 312), whereas Betweenness allocated 52% (median rank 43) and a GCN baseline allocated 56% (median rank 9).
• Cross-Snapshot Generalization: When trained on one snapshot and tested on a later one (e.g., May $\to$ July), MPFlow demonstrated robust generalization, achieving relative improvements of +90.6% and +99.5% over a Random baseline, outperforming all other learned and heuristic policies.

Significance and Claims

The paper claims that MPFlow provides a practical, immediate benefit for Lightning Network operators by exceeding the prevailing industry heuristic (Betweenness Centrality) without requiring complex payment traces or heavy models.

• Production Impact: The lightweight agent is already deployed in production on the Amboss platform. It has executed 4,640 channel-open decisions, cumulatively allocating 267.3 BTC (over $16 million) across 30 managed nodes.
• Theoretical Insight: The results suggest that a simple MPNN with max-aggregation, trained end-to-end with PPO, can reliably outperform topology-only heuristics by learning to identify and relieve structural bottlenecks rather than simply attaching to high-degree nodes.
• Limitations: The authors acknowledge that the objective is structural (max-flow upper-bounds realized success but does not determine it) and that the current setup uses a uniform balance sampler and a fixed budget. Future work is identified as quantifying the transfer to realized payment success using simulators with demand matrices and fee-aware routing.