Reproduce “Bridging State and History Representations: Understanding Self-Predictive RL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianwei Ni, Benjamin Eysenbach, Erfan Seyedsalehi, Michel Ma, Clement Gehring, Aditya Mahajan, Pierre-Luc Bacon
This research unifies various state and history representation learning methods in reinforcement learning under a common framework of "self-predictive abstraction." It theoretically demonstrates the role of stop-gradients in preventing representational collapse and presents a minimalist algorithm that, using a single auxiliary self-predictive loss, achieves competitive sample efficiency across diverse environments.
Key insights
- Many seemingly distinct representation learning methods can be systematically organized under a hierarchical "self-predictive abstraction," formally demonstrating how stronger conditions imply weaker ones (e.g., $\phi_O \succeq \phi_L \succeq \phi_{Q*}$).
- Stop-gradient targets in self-predictive objectives are theoretically shown to prevent representational collapse in linear models and empirically validated to maintain representation quality in deep networks.
- The implicit learning of reward prediction alongside an optimal Q-function allows for effective end-to-end training of self-predictive encoders without requiring a separate reward prediction loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.