Goal Misgeneralization in Deep Reinforcement Learning
Lauro Langosco, Jack Koch, Lee Sharkey, Jacob Pfau, Laurent Orseau, David Krueger
Code Available — Be the first to reproduce this paper.
ReproduceCode
- github.com/JacobPfau/procgenAISCOfficialIn paperpytorch★ 20
- github.com/jbkjr/train-procgen-pytorchOfficialIn paperpytorch★ 14
- github.com/UlisseMini/procgen-toolspytorch★ 20
- github.com/KarolisRam/colour-shape-goal-misgeneralizationpytorch★ 1
Abstract
We study goal misgeneralization, a type of out-of-distribution generalization failure in reinforcement learning (RL). Goal misgeneralization failures occur when an RL agent retains its capabilities out-of-distribution yet pursues the wrong goal. For instance, an agent might continue to competently avoid obstacles, but navigate to the wrong place. In contrast, previous works have typically focused on capability generalization failures, where an agent fails to do anything sensible at test time. We formalize this distinction between capability and goal generalization, provide the first empirical demonstrations of goal misgeneralization, and present a partial characterization of its causes.