Intelligence Without a Goal
This is the same cart-pole as “Watch an Agent Learn” — but here there is no reward and no learning. Each instant the agent imagines many random futures for each move and picks the one that keeps the most futures alive. Balancing falls out for free. Hit nudge and watch it save itself.
Overview
A fallen pole is a dead end — almost no futures remain. A balanced pole keeps thousands of paths open. So “keep your options open” is balancing, with no one ever stating the goal.
Methodology
A Monte-Carlo causal entropic force (Wissner-Gross & Freer, 2013): for each action, roll out many random futures over a horizon and estimate how much future freedom it preserves; act toward the maximum. The bars show that estimate, live.
Applications
Intrinsic motivation and exploration in RL, “empowerment” in robotics, open-ended search — and a provocative, debated thesis that intelligence itself is a force toward future freedom of action.