On the Inference of Turing Machines from Sample Computations A. W. Biermann

This paper will be concerned with the problem of obtaining this performance from the machine by giving it examples of the desired computation and having it program itself. We will be concerned with designing a trainable Turing machine although the concepts presented are applicable in a much more general context as discussed in Section 4. The Turing machine to be discussed here will have an infinite one dimensional tape and will have the capability in one move to read a symbol on the tape, print a new symbol to replace the one just read, and step right or left one increment on the tape. It will have a deterministic finite-state controller with a designated initial state which will upon receiving an input symbol read from the tape, yield the symbol to be printed and the step direction (right or left) to be made.