A Reinforcement Learning-Based Node Selection for PBFT Consensus

Blockchain is nowadays a flourishing technology that served many domains (finance, supply chain, smart cities, healthcare, etc.) thanks to the immutability of data stored on it. In fact, before validating a block and appending it to the blockchain, most of the system nodes should agree on it and reach consensus. One of the most used consensus protocols is the Practical Byzantine Fault Tolerance which tolerates byzantine nodes and works upon asynchronous systems. However, it is not scalable enough because of the increasing number of messages exchanged through the network. Our proposition uses reinforcement learning to adjust the number of participating nodes depending on their behavior and on the actual network state. The number of participating nodes is reduced without affecting the protocol’s safety and liveness. Only honest and rapid nodes are included in the consensus set. Malicious nodes are removed and slow nodes are kept in a set until needed.