MnemoSys: A Conditional Probability Estimation Protocol for Blockchain Audited Reputation Management

Abstract

Reputation systems have been one method of solving the unique challenges that face distributed networks of independent operators. Fundamentally, historical performance must be considered in a way that attempts to predict future behav-ior, optimize present functionality, and provide some measure of immutable recording. In this paper, a three-part system, MnemoSys, is proposed to solve this diverse set of problems. First, historical performance is dynamically weighted and scored using geometrically expanding time windows. Second, a quorum is abstracted as a restricted Boltzmann machine to produce a conditional probability estimate of log-normal likelihood of good-faith behavior. Third, all rewards and punishments are recorded on an immutable, decentralized ledger. Our experimentation shows that when applied iteratively to an entire network, consistently under-performing nodes are removed, network stability is maintained even with high percentages of simulated error, and global network parameters are optimized in the long-term.