Automata for learning sequential tasks

Abstract

This dissertation describes a system that is capable of learning both combinational and sequential tasks. The system learns from sequences of input/output examples in which each pair of input and output represents a step in a task. The system uses finite state machines to represent learned models. The resultant learned models are often deterministic finite state machines which often contain minimal number of states. Thus, the resultant learned models can easily and economically be implemented in VLSI. On the contrary, neural network models are more difficult to be implemented in hardware. This dissertation proposes a method for inferring finite state machines from sequences of input/output examples. The system is often capable of inferring a reasonable finite state machine from just one sequence of examples. On the contrary, recurrent neural networks often require thousands of sequences. New algorithms are developed for modifying the finite state machines to allow the system to learn new information when additional examples are available. In addition, new algorithms are developed to allow the system to handle inconsistent information that may result from noise in the training examples. A simulation program is developed to test the system and several application examples are provided.