A scalable bitcoin blockchain network by dynamic block size adjustment using reinforcement learning

Abstract

The revolutionary technology known as blockchain has revolutionized the way we handle and keep confidential information. In a blockchain network, transactions are organized into blocks that are connected to one another in a sequential and immutable chain. So, the size of the block is a crucial element in the blockchain technology. It influences the number of transactions handled per second, the confirmation time and the general performance of the network. However, the size of each block is fixed to 1 MB in the current technology and we cannot simply alter the block size of the blockchain which causes network congestion and transaction delay. So, we have proposed a model to enhance network performance by dynamically estimating the optimal block size using the Soft Actor-Critic (SAC) algorithm of Reinforcement Learning (RL). This model can predict appropriate block size in real-time by taking account into 4 different features of the network that effects the network performance greatly and resolve the 1 MB fixed block size issues. Soft Actor-Critic model enhances network performance by speeding up transaction confirmation times and making the most use of each block. By using the Soft Actor- Critic model, a scalable improvised network will not only reduce network congestion but it will guarantee consistent profits for the miners. Initially, we constructed a baseline block size prediction model using Long Short-Term Memory (LSTM) networks. However, due to LSTM’s limitations in reflecting the dynamic nature of the blockchain, we implemented a better solution using Reinforcement learning. We have applied both online and offline reinforcement learning techniques using a simulator and a real-world dataset. Offline reinforcement Learning was not impressive as well due to the static nature of the data. But using online reinforcement learning gave us the desired results. In online RL, Gym library-based blockchain simulation environment have been utilized for real-time training and assessment. Later we have applied Inverse reinforcement learning (IRL) for reward shaping using Generative Adversarial Imitation Learning (GAIL). Nevertheless, our reward function and Soft Actor-Critic (SAC) model were most efficient to build a scalable improvised network that guaranteed optimal block utilization, minimum confirmation time and miners consistent profits.