Humans teaching intelligent agents with verbal instruction
MetadataShow full item record
The widespread integration of robotics into everyday life requires significant improvement in the underlying machine learning (ML) agents to make them more accessible, customizable, and intuitive for ordinary individuals to interact with. As part of a larger field of interactive machine learning (IML), this dissertation aims to create intelligent agents that can easily be taught by individuals with no specialized training, using an intuitive teaching method such as critique, demonstrations, or explanations. It is imperative for researchers to be aware of how design decisions affect the human’s experience because individuals who experience frustration while interacting with a robot are unlikely to continue or repeat the interaction in the future. Instead of asking how to train a person to use software, this research asks how to design software agents so they can be easily trained by people. When creating a robotic system, designers must make numerous decisions concerning the mobility, morphology, intelligence, and interaction of the robot. This dissertation focuses on the design of the interaction between a human and intelligent agent, specifically an agent that learns from a human’s verbal instructions. Most research concerning interaction algorithms aims to improve the traditional ML metrics of the agent, such as cumulative reward and training time, while neglecting the human experience. My work demonstrates that decisions made during the design of interaction algorithms impact the human’s satisfaction with the ML agent. I propose a series of design recommendations that researchers should consider when creating IML algorithms. This dissertation makes the following contributions to the field of Interactive Machine Learning: (1) design recommendations for IML algorithms to allow researchers to create algorithms with a positive human-agent interaction; (2) two new IML algorithms to foster a pleasant user-experience; (3) a 3-step design and verification process for IML algorithms using human factors; and (4) new methods for the application of NLP tools to IML.