Learning Automata: An Alternative to Artificial Neural Networks

Abstract

This paper describes the research done by the author in three areas of study; all of them are concerned with machine learning. Although this paper relates mainly to learning automata, the related fields of neural networks and adaptive learning are also discussed. Learning Automata was introduced by Tsetlin and later by Narendra. Collective Models have been proposed in different fashion by authors like by Michie, Barto and Bock. Briefly, the Learning Automata approach is based on the use stochastic matrices to achieve learning by means of different reinforcement schemes. The main distinction of this research with other learning automata approaches is that of taking a collection of actions before the environment delivers an evaluation value. Much of the work described here has been published under the name of collecting learning systems by others authors. A organized structure of stochastic matrices can be used to obtain machine learning in a step by step fashion, a collective fashion and even in a hierarchical manner, as discussed in the paper. The examples discussed lie within the realm of pattern recognition and game playing. Neural network terminology and concepts relate to the work discussed here. A large number of cells arranged in a network fashion are organized to produce the desired machine learning outputs. The use of learning automata rather than a deterministic backpropagation results in faster convergence times. One must take into account that neural networks are slow but concentrate on obtaining adequate behavior; yet, in the research presented here, the speed in which the systems leams is also of main concern.