Menu

Active Inference — The Learn Arc, Part 40: Session §8.1 — Generalized coordinates
📰

Active Inference — The Learn Arc, Part 40: Session §8.1 — Generalized coordinates

DEV Community: elixir·ORCHESTRATE·7 months ago
#VNDKiZEB
#dev#code#strong#generalized#derivatives
Reading 0:00
15s threshold

Series: The Learn Arc — 50 posts through the Active Inference workbench.

Previous: Part 39 — Session §7.5: Worked example

Hero line. Generalized coordinates replace the single state s_t with a tower: s , s' , s'' , s''' , … The belief is over the state and all of its derivatives at an instant . That is what makes continuous-time inference tractable.

Why a tower

Discrete-time active inference gets a free lunch: every t gives you a fresh observation, and Eq 4.13 is a single softmax. In continuous time there is no clean tick. You need a representation that captures both where the state is and where it is heading at the same moment.

Generalized coordinates solve this. Instead of s_t , the agent carries (s, s', s'', s''', …) — position, velocity, acceleration, jerk, and so on — as its instantaneous estimate. The gaussian prior binds them together with a characteristic smoothness.

Five beats

Generalized state = position + derivatives. The belief is a multivariate gaussian over (s, s', s'', …) .…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.