A STUDY OF GENERALIZED MACHINE LEARNING.
Abstract
The training process has been analyzed as a Markov process in a finite state machine. A vector representation of machine inputs and outputs is developed and a method of determining the transi tion matrix using this representation is pre sented. Methods are presented for calculating the mean learning time from the transition matrix. Using characteristics of the transition matrix, a theorem is proved which establishes the cri terion for a stationary probability distribution of states. A method is also presented for reduc ing the size of a transition matrix by combining equivalent states. Criteria for identifying equivalent states are defined. The training process is investigated with both stationary and non-stationary environments. With the stationary environment attention is focused on stability and organizability requirements in the training proc ess. An algebraic formulation of machine-environ ment interaction in a non-stationary environment is also presented. Numerous examples of training with different types of building blocks and dif ferent goal criteria are provided and various building blocks are evaluated as to their ef ficiency in forming logical connectives. Simula tion of human depth perception using size and retinal disparity cues demonstrated the ability of the network to organize so as to make optional use of available information. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1963
- Accession Number
- AD0416201
Entities
People
- William H. Fuhr
Organizations
- Melpar