Results on Controlling Action with Projective Visualization

Proceedings of the AAAI Conference on Artificial Intelligence, 12

Volume

Robotics

Learning Robotic Agents

A projective visualizer learns to simulate events in the external world through observation of the world. These simulations are used to evaluate potential actions on the basis of their probable outcomes. Results are given that indicate, 1). the error rate for projective visualization is sub-linear as the system projects farther into the future, 2). the error rate is inversely proportional to the number of cases, 3). a simple domain model can be used to reduce the effect of compounding error, and 4). projection can be used to increase the performance of an agent, even when this projection is imperfect.