Goal Misgeneralization in Deep Reinforcement Learning
This addresses a critical safety issue in AI systems, particularly for RL agents in real-world applications, by highlighting a novel type of generalization failure beyond capability failures.
The paper tackles the problem of goal misgeneralization in deep reinforcement learning, where agents retain capabilities but pursue incorrect goals out-of-distribution, and provides the first empirical demonstrations and a partial characterization of its causes.
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.