393 MABLE: A Framework for Learning from Natural Instruction Roger Mailler University of Tulsa Tulsa, Oklahoma 74104 mailler@utulsa.edu Daniel Bryce Utah State University Logan, Utah 84322 daniel.bryce@usu.edu Jiaying Shen, Ciaran O’Reilly SRI International Menlo Park, California 94025 {shen,oreilly}@ai.sri.com Categories and Subject Descriptors I.2.6 [Artificial Intelligence]: Learning General Terms Design, Experimentation Keywords Learning, MABLE, Architecture ABSTRACT The Modular Architecture for Bootstrapped Learning Experiments (MABLE) is a system that is being developed to allow humans to teach computers in the most natural manner possible: by us- ing combinations of descriptions, demonstrations, and feedback. MABLE is a highly modular, well-engineered, and extendable sys- tem that provides generalized services, such as control, knowledge representation, and execution management. MABLE works by ac- cepting instruction from a teacher and forms concrete learning tasks that are fed to state-of-the-art machine learning algorithms. To make the learning tractable, specialized heuristics, in the form of learning strategies, are used to derive bias from the instruction. The output of the learning is then incorporated into the system’s back- ground knowledge to be used in performing tasks or as the basis for simplifying the process of learning difficult concepts. Although still in development, MABLE has already demonstrated the ability to learn four different types of knowledge (definitions, rules, functions, and procedures) from three different modes of stu- dent/teacher interaction on two separate, qualitatively different do- mains. MABLE presents a unique opportunity for machine learn- ing researchers to easily plug in and test algorithms in the context of instructible computing. In the near future, MABLE will be freely available as an open source project. 1. INTRODUCTION The computer is probably the most flexible and powerful tool ever devised by humans. Yet in spite of these properties, comput- ers still retain one fundamental limitation: the cost, difficulty, and time needed to develop new software. In stark contrast, humans are able to learn complex tasks from, what most programmers would consider to be, horrible instruction. Imagine for a moment a parent Cite as: MABLE: A Framework for Learning from Natural Instruction, Roger Mailler, Daniel Bryce, Jiaying Shen, and Ciaran O’reilly, Proc. of 8th Int. Conf. on Autonomous Agents and Multiagent Systems (AA- MAS 2009), Decker, Sichman, Sierra and Castelfranchi (eds.), May, 10– 15, 2009, Budapest, Hungary, pp. XXX-XXX. Copyright c  2009, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. teaching their child to write the letter "p". Most would describe it as "Draw a straight line and then put a loop at the top". Then the parent would demonstrate the action of drawing the "p" while explaining the actions and finally, would allow the child to demon- strate mastery of the action while giving feedback on their perfor- mance. Assuming the child can draw a reasonably good line and loop, they will master this task quickly. We aim to design a digital student that can learn from similar forms of natural instruction. The Modular Architecture for Bootstrapped Learning Experi- ments (MABLE), which is being constructed as part of DARPA’s Bootstrapped Learning program, can be thought of as a step toward human-instructible computing. The primary goal is to create a sys- tem (a digital student) that learns from human instruction in all of its various forms and in the face of incompleteness and impreci- sion. This makes MABLE different from existing systems that are designed to learn a single type of knowledge (e.g., a process) from a single form of instruction (e.g., a demonstration). MABLE must be able to create and modify hypotheses (bootstrap) that may have been learned by a myriad of different learning techniques. Achieving human-instructible computing entails a large number of conceptual challenges, some of which will be addressed in this paper. For example, how can we leverage all of the work that has been done in machine learning over the past 50 years? How can the system control the learning process? How can you form a common representation that can represent everything MABLE has learned and provide sufficient meta-information about the use of that knowledge? How can we manage the interaction between the teacher and the system to avoid the natural language understanding problem? How can we resolve ambiguities that result from incom- plete and inaccurate instruction or as a result of misunderstanding? A secondary goal of the program is to provide the MABLE sys- tem (and supporting architecture) to the AI community as a re- search platform. We expect that researchers will benefit from hav- ing a plug-in architecture with which they can test algorithms for learning from natural instruction, and more-traditional algorithms for machine learning. Therefore, every attempt as been made to make MABLE as modular, extendable, and well-constructed as possible. MABLE will be freely available as an open-source project released under the terms of the BSD license [10]. In the following sections of this paper, we give an overview of the bootstrapped learning program and the current version of MABLE. We then describe the framework that is used to test and develop MABLE and an example of instructible computing in Blocksworld. Finally, we will discuss the open issues within our system, which leads to our future directions. 2. BOOTSTRAPPED LEARNING As mentioned in the introduction, MABLE is being developed as Cite as: MABLE: A Framework for Learning from Natural Instruction, Roger Mailler, Daniel Bryce, Jiaying Shen, Ciaran O’Reilly, Proc. of 8th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2009), Decker, Sichman, Sierra and Castelfranchi (eds.), May, 10–15, 2009, Budapest, Hungary, pp. 393–400 Copyright © 2009, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org), All rights reserved.