---

name: grid-world-dp-and-control description: "Use and adapt GridWorld dynamic programming, tabular control, Deep SARSA, and REINFORCE workflows distilled from the reinforcement-learning examples." metadata: disco-role: operating disable-model-invocation: true license: MIT

GridWorld DP and Control

Use this sub-skill when the task is about the repository's 5x5 GridWorld workflows: policy iteration, value iteration, tabular SARSA, tabular Q-learning, Deep SARSA, or REINFORCE. It is self-contained; do not depend on the original checkout at runtime.

Route here for

• Explaining, adapting, or debugging GridWorld dynamic programming (1-policy_iteration.py, 2-value_iteration.py) as provenance workflow labels.
• Explaining, adapting, or debugging tabular control (3-sarsa.py, 4-q_learning.py) as provenance workflow labels.
• Making tiny CPU checks for Deep SARSA and REINFORCE neural update logic (5-deep_sarsa.py, 6-reinforce.py) as provenance workflow labels.
• Diagnosing row/column versus column/row coordinate mistakes, action-order mismatches, and render/headless behavior.

Route away from

• Gymnasium CartPole DQN/A2C/PPO: use the CartPole sub-skill.
• Atari Breakout/Pong DQN/PPO, preprocessing, ROMs, or W&B: use the standard Atari sub-skill.
• Montezuma/Pitfall/PrivateEye hard-exploration, RND, Go-Explore, or robustification: use the hard Atari sub-skill.

Start here

1. Read references/algorithm-guide.md for the MDP layout, coordinate conventions, formulas, API-shaped examples, and safe adaptation patterns.
2. Read references/troubleshooting.md when a grid policy points in the wrong direction, rendering blocks on a server, tabular values do not update, or Torch neural checks fail.
3. Run the bundled smoke script when you need a self-contained, non-rendering API check:

python scripts/grid_world_smoke.py --help
python scripts/grid_world_smoke.py
python scripts/grid_world_smoke.py --section dp
python scripts/grid_world_smoke.py --section neural --strict-torch

The smoke script intentionally reimplements tiny fixtures and update steps. It does not import source repository files, open Pygame, train to convergence, or require a display.

Critical facts to preserve

• DP workflows use state as [row, col]; drawing helpers usually need (col, row), so swap before rendering arrows or markers.
• Tabular Env-style workflows use state as [col, row] and actions 0=up, 1=down, 2=left, 3=right.
• Dynamic DynamicEnv-style workflows also use [col, row], but actions are 0=up, 1=down, 2=right, 3=left; the distilled source agents expose a fifth network output/action that behaves as a no-op in the dynamic environment.
• GUI examples are interactive and render-oriented. For validation or CI, use headless logic equivalent to render_mode=None rather than opening Pygame loops.