Deep RL algorithms learn distinct representational invariances

Researchers have analyzed deep reinforcement learning representations using MDP reduction theory, finding that different algorithms learn distinct types of invariances. Specifically, DQN learns representations invariant to MDP homomorphism symmetries, while PPO learns representations invariant to action symmetries, even when performance is similar. These representational differences have implications for transfer learning and may also be observed in large language models in a prompt-dependent manner. AI

IMPACT Different RL algorithms learn distinct representational invariances, impacting transfer learning and potentially LLM behavior.

RANK_REASON The cluster contains an academic paper detailing novel research findings in deep reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• LLMs

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →