Chapter 03 of 08 · teaching experiments

The teacher, and where labels come from

A policy is any rule from observations to actions. Before learning one, it pays to measure how far non-learned rules get, and to meet the teacher whose answers this study spends as currency.

The policy spectrum, measured

Three policies, identical scenarios. A uniformly random policy succeeds in of episodes, more than expected: the maze is small and steps is a long time, so blind wandering does stumble onto the target, hopelessly inefficiently. A hand-written wall follower reaches . The scripted oracle solves in on average.

What the oracle is

BabyAIBot, shipped with MiniGrid, is a scripted planner with privileged access: it reads the full grid, the true state that chapter 02 proved the agent never observes, and replans from a subgoal stack at every step. It is not learned, and it never acts at evaluation time. It exists to answer one question at any visited state: what should be done here? That answer is an expert label. A demonstration is nothing more than a sequence of such labels along the expert's own path, exactly what the video and scrubber in chapter 01 showed.

Trying to break the teacher

The bot tracks the world by being told which action was executed. Two questions, measured on identical scenarios with one forced off-plan action each. First, the fact the whole study rests on: an honestly informed oracle recovers, success after the with the deviation, which is what lets it label states a learner reaches instead of the expert. Second, a falsification attempt: lying to it, never informing it, double-calling its planner. Nothing broke.

One demonstration, frame by frame