Dissertation/Thesis Abstract

Computational and Design Techniques for a Semi-Autonomous Computerized Dog-Training System with Timing and Accuracy Performance Comparable to a Professional Dog Trainer
by Majikes, John J., Ph.D., North Carolina State University, 2018, 98; 10970031
Abstract (Summary)

Humans spend a great deal of time and effort to train dogs to perform specific tasks when given a command. Areas of research, such as canine cognition and Applied Behavior Analysis, have been developed to aid in this dog training. From this body of knowledge and the canine-human relationship, professional dog trainers have come to rely on timing, accuracy, and repetition to place the desired task under stimulus control such that a dog provides the desired behavior only when the stimulus is given.

For example, after observing a dog raising its paw possibly to scratch, a trainer might recognize this paw motion, exploit the dog's innate ability to recognize human gestures, and quickly offer positive reinforcement. The trainer continues shaping the behavior until it resembles a hand shake, associates the behavior with a cue of “shake”, and places the hand-shake behavior under stimulus control such that it's only offered when the trainer says “shake”.

My research relies on the canine-human relationship, three fundamental animal training concepts of timing, accuracy, and rate of reinforcement, and codifies the dog training techniques of capturing, shaping, cue association, and stimulus control into a semi-autonomous computerized canine training system with timing comparable to, but more consistent than a professional dog trainers. The difficulties in the system design are threefold: First, the system must balance timely behavior recognition with sensor noise reduction. Second, the system must consistently recognize and reinforce the behavior when it is offered. Third, the system must employ a reinforcement schedule that encourages the animal to repeatably provide the behavior when requested. Training success requires the system to identify when the dog is attentive, offer the cue, and reinforce the behavior at a rate of reinforcement comparable to a professional dog trainer.

The success of the system will be shown in three experiments and one pilot by using existing analytical techniques for noise/latency trade-off, for behavior recognition accuracy, and for achieving a rate of reinforcement comparable to a professional dog trainer. Experiment 1 will examine accuracy and latency of the system. Experiment 2 will compare a professional dog trainer to the system using a novel and an existing posture classification system. Experiment 3 will show some evidence of canine learning when using the system. And the Discriminative Stimulus Pilot will combine the experience from the first three experiments to demonstrate a system that can begin to do discriminative stimulus training.

Since training relies on the canine-human relationship, a computer system cannot replace a professional dog trainer. Professional dog trainers rely on years of experience and understanding of behavioral signs to effectively communicate human requests and canine responses. A system that codifies a professional trainer's knowledge and facilitates human-canine communications has the potential to positively effect the lives of both canines and humans. Therefore the goal of this work is a system that facilitates canine training, algorithmically encapsulates training techniques, and leverages the advantages of computers: timing and consistency.

Indexing (document details)
School: North Carolina State University
Department: Computer Science
School Location: United States -- North Carolina
Source: DAI-B 80/01(E), Dissertation Abstracts International
Subjects: Behavioral Sciences, Computer science
Keywords: Animal-computer interaction, Classification algorithm, Machine learning, Stimulus control
Publication Number: 10970031
ISBN: 978-0-438-28488-3
Copyright © 2020 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy