Computer Science

Simplifying Latent Dynamics with Softly State-Invariant World Models

T. Saanum, P. Dayan, et al.

To solve control problems an agent must predict how its actions change the world. The Parsimonious Latent Space Model (PLSM) regularizes latent dynamics by minimizing mutual information between latent states and action-induced changes, producing softly state-invariant world models that improve prediction, generalization, and downstream performance. This research was conducted by Tankred Saanum, Peter Dayan, and Eric Schulz.