An Improved Method for Tsetlin Machine Based on Multi-level Feedback and Dynamic Parameter

Abstract

Addressing the issues of slow convergence and insufficient stability in the training process of Tsetlin Machine (TM), this paper proposes an innovative approach that integrates a multi-level feedback mechanism with dynamic parameter adjustment. By introducing a four-level feedback mechanism, this method accelerates the reinforcement of high-contribution clauses and the correction of significant errors through fine-grained feedback intensity allocation. Additionally, a dynamic parameter adjustment strategy for s is proposed, which optimizes state transition probabilities across different training stages using periodic steps d, thereby enhancing the model's exploration-exploitation balance capabilities. Experimental results demonstrate that this method improves training speed by approximately two times on the MNIST dataset, with an accuracy enhancement of 0.5%. These findings confirm that the proposed method significantly enhances model convergence efficiency and generalization performance through graded feedback intensity and adaptive parameter mechanisms, offering new insights for the deployment of interpretable AI in resource-constrained scenarios.