Abstract
Humans often teach procedures through tutorial instruction to other humans. For computers, learning from natural human instruction remains a challenge as it is plagued with incompleteness and ambiguity. Instructions sre often given out of order and are not always consistent. Moreover, humans assume that the learnerhas a wealth of knowledge and skills, which computersdo not always have. Our goal is to develop an electronic student that can be made increasingly capable through research to learn from human tutorial instruction. Initially, we will provide our student with human understandable instruction that is extended with many scaffolding statements that supplement the limited initialcapabilities of the student. Over time, improvements to the system are driven and quantified by the removal of scaffolding instructions that are not consideredto be natural for users to provide humans. This paper describes our initial design and implementationof this system, how it learns from scaffolded instruction in two different domains, and the initial relaxations of scaffolding that the system supports.
Original language | English |
---|---|
Title of host publication | AAAI 2009 Spring Symposium on Agents that Learn from Human Teachers |
Publisher | The AAAI Press |
Publication status | Published - 2009 |