Subjects › AGI Agents (6-week course)
W1 · World models & planning
A model of how the world responds to actions, and searching it (BFS, A*) before acting.
The idea
A world model is a function from (state, action) to predicted outcomes — here a learned lookup table with counts, which is exactly why it cannot generalise to a new layout. Planning is search over that model: BFS for fewest steps, A* when an admissible heuristic (Manhattan distance on the grid) lets you skip hopeless branches. In your repo: Week01_Foundations_Cognition/world_model.py and planner.py; Exercises 1.1 (uncertainty as entropy) and 1.2 (safe_astar_plan).
Lesson
No lesson built yet. The worked solutions below are the material it would be assembled from.
Questions (5)
| Question | Type | Difficulty | Drift | Accuracy |
|---|---|---|---|---|
| Your agent's LearnedWorldModel has never tried moving left from cell (2,3). During planning, A* calls predict(… | mcq | Easy -1.30 | — | — (0) |
| A* is planning towards the goal at (5,7) in the course GridWorld (unit step costs, Manhattan-distance heuristi… | numeric | Easy -0.60 | — | — (0) |
| Exercise 1.1 defines uncertainty(s, a) as the Shannon entropy, in bits, of the model's outcome distribution. T… | numeric | Moderate 0.20 | — | — (0) |
| Exercise 1.2's safe_astar_plan refuses any (s, a) with confidence(s, a) < 0.6. confidence() is the probability… | mcq | Hard 0.90 | — | — (0) |
| The agent's world model is a table of transition counts. Explain what exactly stops it from generalising to a … | short | Very hard 1.60 | — | — (0) |