Fuzlearn: A Fuzzy Clusterization based Machine Learning using Learning Automata

Existing machine learning (ML) algorithms, such as Neural Networks (NNs), typically consist of complex data encoding and preprocessing on real-world data followed by extensive arithmetic. As such when implemented in hardware or software they incur high energy consumption as well as increased data-to-decision latency. Recently, the Tsetlin Machine (TM) was proposed as a promising new alternative which takes advantage of logic propositions founded on the principle of Tsetlin Automata, leading to higher energy efficiency, faster training convergence and lower latency. However, TM interfaces require complex data encoding procedures which can add significant overheads. In this paper, we propose a new machine learning approach that is built on the principle of fuzzy clusterization of decision boundaries directly from analog data. This approach effectively excludes the usual booleanization or analog-to-digital (A2D) conversion necessary in conventional approaches, thus offering low-complexity and potentially energy efficiency gains. Fundamental to this algorithm is a team of Learning Automata constituting a clusterization unit termed Clause that act as an Analog-to-Boolean classifier. Depending on the complexity of the ML dataset, a team of clauses derive the classification rules with minimal algorithmic and arithmetic complexities. The software implementation of the algorithm is tested against several problems. Experiments show that the proposed algorithm manages to achieve comparable accuracy as the Tsetlin Machine with fewer clauses. To study the impact of noise in the input features, we subsequently validate our algorithm with a few variants of noisy XOR datasets. Experiments with these datasets show that our approach is capable of mitigating noise effectively due to fuzzy classification boundaries.