dynamics

Definition

The rule governing how a system's state evolves over time. In reinforcement learning it is the transition from current state and action to the next state; learning an approximation of it and planning through that approximation is what makes a method model-based. The word is also used for training dynamics — how loss, gradients, and internal representations move over the course of a run — so check whether the changing thing is the world or the model.

Explainers using this term

• Paper Walkthrough: EnvACE — Agents That Rehearse the World Instead of Calling ItEnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning