EBMICQL: Improving Efficiency of Blockchain Miner Pools via Incremental and Continuous Q-Learning Framework

Abstract

Blockchain mining pools assist in reducing computational load on individual miner nodes via distributing mining tasks. This distribution must be done in a non-redundant manner, so that each miner is able to calculate block hashes with optimum efficiency. To perform this task, a wide variety of mining optimization methods are proposed by researchers, and most of them distribute mining tasks via statistical request processing models. These models segregate mining requests into non-redundant sets, each of which will be processed by individual miners. But this division of requests follows a static procedure, and does not consider miner specific parameters for set creation, due to which overall efficiency of the underlying model is limited, which reduces its mining performance under real-time scenarios. To overcome this issue, an Incremental & Continuous Q-Learning Framework for generation of miner-specific task groups is proposed in this text. The model initially uses a Genetic Algorithm (GA) method to improve individual miner performance, and then applies Q-Learning to individual mining requests. The Reason for selecting GA model is that it assists in maintaining better speed-to-power (S2P) ratio by optimization of miner resources that are utilized during computations. While, the reason for selecting Q-Learning Model is that it is able to continuously identify miners performance, and create performance-based mining pools at a per-miner level. Due to application of Q-Learning, the model is able to assign capability specific mining tasks to individual miner nodes. Because of this capability-driven approach, the model is able to maximize efficiency of mining, while maintaining its QoS performance. The model was tested on different consensus methods including Practical Byzantine Fault Tolerance Algorithm (PBFT), Proof-of-Work (PoW), Proof-of-Stake (PoS), and Delegated PoS (DPoS), and its performance was evaluated in terms of mining delay, miner efficiency, number of redundant calculations per miner, and energy efficiency for mining nodes. It was observed that the proposed GA based Q-Learning Model was able to reduce mining delay by 4.9%, improve miners efficiency by 7.4%, reduce number of redundant computations by 3.5%, and reduce energy required for mining by 7.1% when compared with various state-of-the-art mining optimization techniques. Similar performance improvement was observed when the model was applied on different blockchain deployments, thus indicating better scalability and deployment capability for multiple application scenarios.