Reproduce “What deep reinforcement learning tells us about human motor learning and vice-versa” 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

Michele Garibbo, Casimir Ludwig, Nathan Lepora, Laurence Aitchison

MB-DPG combines a learned action-to-outcome model with an error signal evaluated at the observed, not predicted, endpoint. In simulations, it reproduces human-like worsening after mirror reversal when slower model learning, update saturation, and nonvisual feedback are added. Under tested model biases, it roughly maintains performance while model-based reinforcement learning is impaired; findings are limited to these simulations and conditions.

Key insights

  • Standard deep RL algorithms struggle to adapt to sensory perturbations in the same way humans do, failing to show initial performance degradation in mirror-reversal tasks.
  • Human error-based learning likely benefits from processing observed rather than predicted outcomes to compute directed error gradients for adaptation.
  • Integrating principles from human motor learning, specifically relying on observed outcomes for error calculation, can lead to more robust and biologically plausible RL agents.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.