Systematic Search for Optimal Hyper-parameters of the Tsetlin Machine on MNIST Dataset

Modern Machine Learning (ML) models have a significant number of hyper-parameters that need adjusting to leverage performance and energy efficiency for a given model configuration during training. This becomes a considerable design challenge with increasing complexity requiring larger models. This paper explores the Tsetlin Machine (TM) - a new logic-based ML approach with only four hyper-parameters regardless of the problem space. Two of these hyper-parameters influence the TM architecture while the remaining two impact the learning efficacy. This work focuses on the systematic search for optimal hyper-parameters for the TM and aims to understand how hyper-parameter values affect performance and prediction accuracy using MNIST dataset as a case study.