Design of Event-Driven Tsetlin Machines Using Safe Petri Nets

In the last decade, there has been a significant shift towards the use of machine learning (ML) within the technology industry. One prominent ML algorithm is the Tsetlin Machine (TM), where it uses a collection of learning automata to learn new patterns through propositional logic. While TMs are considered computationally simpler and more efficient than neural networks (NNs), there is difficulty in how TMs can be better understood by industrial practitioners. Although many approaches help demonstrate the benefits of TMs, there is however no approach that helps better explain the behaviour of TMs, e.g. how the TM’s decision is influenced by the initial states of their learning automata and how the TM’s learning is determined by the calculations made from its inference and feedback components. In this paper, we present the concept of event-driven TMs, where we model the complete behaviour of TMs using 1-safe Petri nets. The key aspects of Petri nets are their flexibility to model many types of specifications including distributed systems and concurrent systems, and their rich support from many well-established tools including Petrify, MPSat, and Workcraft. To highlight the benefits of our approach, we conduct a simple experiment where we showcase our Petri net specifying the complete behaviour of a TM, analyse its behaviour through a set number of epochs, and most importantly evaluate its accuracy.