Zoom Link: https://cmu.zoom.us/j/91465601940

Title: Solving Puzzles Like A Human With Two Stage Random Search

Abstract: Humans are remarkably good at solving novel physical puzzles and tasks, with only a basic understanding of abstract concepts like kinematics, gravity, mass, friction, and inertia. We wanted to replicate how a human would explore the search space of a problem. To do so, we use a 2 step algorithm that first randomly explores the entire search space with some simple heuristics, and then uses local optimization to refine the solution. This replicates how we often solve difficult problems as humans, where we first explore possible solutions using our prior general knowledge, and then refine the promising solutions. We tes this method on the PHYRE (PHYsical REasoning) benchmark, a set of simple mechanics puzzles in a 2D environment, showing that we outperform the simple baseline agents, and have comparable performance to the more complicated agents.

Committee:

Chris Atkeson (advisor)

Aikaterini Fragkiadaki

Leonid Keselman