Daniel Arista and Catherine Sibert, RPI Graduate Students

 

Daniel Arista and Catherine Sibert, RPI Graduate Students

Sage 4101

February 18, 2015 12:00 PM - 1:30 PM

Reasonableness in Agency: 

For an AI agent to be trusted to act autonomously and exercise discretion, it is going to have demonstrate reasonable interpretations of the communications it receives. What does it mean to interpret reasonably? Can we expect reasonableness to reduce to a computational model? To answer the first question we'll look at US Agency Law, and borrow some of it's key assumptions necessary to establish a Principal-Agent model. From this we find that the capacity for commonsense interpretations of conditional statements and a procedure for introducing new conditional knowledge is inevitable. We briefly refer to Mental Model Theory's claims about the modulating effects context has on conditional reasoning, and leverage them as a paradigm of 'reasonable interpretation.'  In response to the second question, we propose an algorithm for processing these interpretations using extensions of the Event Calculus developed in the RAIR lab. 

 

Tetris:

The game of Tetris, in addition to being one of the most played video games in the world, enjoys a unique position as an experimental tool for both psychologists and computer scientists. However, very little research has been done that links the two approaches. We use cross-entropy reinforcement learning (CERL),(Szita & Lorincz, 2006; Thiery & Scherrer, 2009), an approach from the computer science community, to explore (a) the utility of high-level strategies (goals or objective functions) for maximizing performance under human-like constraints and (b) a variety of features and feature-weights (methods) for optimizing a low-level, one-zoid optimization strategy. The resulting models were then used to predict the placement decisions of 67 human subjects, and were able to correctly choose the human's move between 43% of the time, for novice players, and 65% of the time, for expert players. Moving forward, we hope to use these models as tutors in an effort to improve the performance of novice or intermediate Tetris players. 

 

HD VIDEO LINK

Add to calendar
Share|