Tristan Andrew Fraser Thrush

Tristan Andrew Fraser Thrush

Scholar Title

MIT EECS Undergraduate Research and Innovation Scholar

Research Title

Learning Intuition for Problem Solving with K-Line Theory

Cohort

2017–2018

Department

EECS

Research Areas
  • Artificial Intelligence & Machine Learning
Supervisor

Patrick H. Winston

Abstract

The Genesis Story Understanding System has a variety of mental resources that allow it to solve problems, but it still cannot learn anything from the problems that it has seen before. This faculty is essential if Genesis is to provide a computational model of high-level symbolic human intelligence, which is the overall goal of our research group. Currently, Genesis attempts to solve a problem by randomly looking up an insight, which provides a mapping from a question structure to an action (known as an intention) that may be useful to answer the question. Genesis can have an unbounded number of insights for the same structure of question; additionally, a single insight may lead to multiple intentions, and many intentions may not lead to a solution. Sometimes, it may also be necessary for Genesis to follow a path of insights before it can solve a problem. If Genesis has many insights for a certain question, randomly finding a path of insights (or even a single insight) that leads to a solution may be infeasible. The goal of my project is to enable Genesis to learn a path of insights (Minsky would call them agents and agencies ) to use for a novel problem, by recovering partial mental states from problems that it has already solved. Essentially, I hope to implement Minsky’ s notion of K-lines in Genesis, and to produce the final deliverable of a robust reinforcement learning system that takes several example problems and insights as an input, bootstraps a learning process by experimenting with these example problems until it can solve them efficiently, and then learns from and solves novel problems that have the same question structure as the examples.

Quote

I am fascinated with the idea of creating a human-level artificial intelligence (AI) and believe that developing a robust problem-solving apparatus is essential to doing so. Humans can leverage their memory to gain insights when solving novel problem. After working on the Computer Science and Artificial Intelligence Laboratory (CSAIL) Genesis story-understanding system, I hope to learn about how I can give it this ability as well.

Back to Scholars