PLUTO: Designing Learning Problems and Plans

Sponsors	Gordon Shippey shippey@cc.gatech.edu	Ashwin Ram ashwin@cc.gatech.edu
Area	Intelligent Systems

Problem
A major aim of Artificial Intelligence is to create programs that can learn independently rather than rely on instructions from a programmer. The PLUTO (Planning to Learn Using Transmutation Operators) architecture is one attempt to design such a learning system. The key idea behind PLUTO is that of planning to learn using small "chunks" of reasoning called "knowledge transmutations." PLUTO's task is to create plans which
compose these knowledge transmutations into useful networks of inference in order to learn.

While the implementation of this system is not complete, creating detailed examples which PLUTO can run will help to guide the design as well as give us extra ways to evaluate the system.

Background
To get up to speed with the planning-to-learn approach and multistrategy learning, read the following papers. Talk with Gordon Shippey for the specifics of how PLUTO embodies some of these ideas.

Chapter 1 of Ashwin Ram's Goal-Driven Learning book.
Learning as Goal-Driven Inference: http://www.cc.gatech.edu/faculty/ashwin/papers/er-95-05.ps.Z
Multistrategy Learning: http://www.cc.gatech.edu/faculty/ashwin/papers/git-cc-92-19.ps.Z

Deliverables
Since PLUTO isn't a functional system yet, all your work will be done "on paper." Note that this isn't an excuse to be less specific, but you don't have to encode your work in any formalism (unless you're going for extra credit, in which case you do.)

A one-page summary (or longer) of the planning-to-learn approach and how PLUTO works.
One learning problem which you think PLUTO should be able to do. Be specific and detailed about how the problem is represented.
A detailed plan of how PLUTO might solve this problem, including all the knowledge transmutations, knowledge goals and background knowledge required.
Extra credit: Come up with a concise, machine-readable specifications for PLUTO knowledge goals, transmutations, and plans.

Evaluation
Evaluation is based on how well your summary communicates the principles of the planning-to-learn approach and the significant design choices of the PLUTO architecture. The problem and solution parts of the deliverableswill be evaluated both on how ambitious the goals are, the level of detail in the learning plan, and the overall plausibility of the solution.