The authors developed hyperbolic contrastive reinforcement learning, a drop-in change to goal-conditioned CRL.
Hyperbolic geometry beats extreme network depth in goal-conditioned maze navigation
A 64-layer hyperbolic agent surpassed 1024-layer Euclidean models on difficult humanoid mazes while training faster with fewer parameters.
Research Lab
Timo Klein · Thomas Lang · Andrii Shkabrii · Lukas Miklautz · Yllka Velaj · Sebastian Tschiatschek
University of Vienna · UniVie Doctoral School Computer Science · Max Planck Institute of Biochemistry · ds:UniVie
Research Digest··2 min read
Klein et al.
Why this paper
From Max Planck Institute of Biochemistry and 3 others
In one line
Using hyperbolic geodesic distances in contrastive RL lets depth-64 agents outperform 1024-layer Euclidean networks on hard maze navigation, training over 12x faster with 13x fewer parameters.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§