Increasing energy efficiency of bitcoin infrastructure with reinforcement learning and one-shot path planning for the lightning network

Increasing energy efficiency of bitcoin infrastructure with reinforcement learning and one-shot path planning for the lightning network

연구 분야: Networking

학회: Neural Computing and Applications

The lightning network (LN) is a technological solution designed to solve the bitcoin blockchain transaction speed problem by introducing off-chain transactions. Since LN is a sparse and highly distributed network with three predominant routing protocols, its native pathfinding algorithms can potentially find multi-hop payment paths similar from the payment sender’s perspective, but the algorithms themselves have different performance, computational cost, energy consumption, and ultimately different CO2 emissions per step in the pathfinding phase. Bitcoin itself generates approximately 61.4 million tons of CO2 eq. per year. Since the LN is built on top of bitcoin, every small change in its energy consumption can have a significant impact on overall pollution. In this paper, we show that the reinforcement learning (RL) approach can reduce these costs and achieve better performance in terms of energy consumption at each pathfinding step. We introduce one-shot path prediction and propose a RL solution for a network agent that learns its neighborhood and uses local knowledge to cleverly solve the pathfinding problem and outperform native pathfinding algorithms.