ARC-AGI Challenge 3 – StochasticGoose

StochasticGoose is an action-learning agent for the ARC-AGI-3 Agent Preview Competition. It uses a simple reinforcement learning approach to predict which actions will cause frame changes, enabling more efficient exploration than random selection.

Authors

• Lead Developer: Dries Smit
• Adviser/Reviewer: Jack Cole

Overview

The action learning agent uses a CNN-based model to predict which actions (ACTION1-ACTION6) will result in new frame states. This enables more precise exploration by biasing action selection toward actions predicted to cause changes.

Key Features:

• CNN with shared backbone for action and coordinate prediction
• Binary classification: predicts if actions will change the current frame
• Hierarchical sampling: first select action type, then coordinate if needed. The coordinate sampling is done purely through convolution to retain the 2D grid bias.
• Efficient experience buffer that stores all experiences with hash-based deduplication for maximum sample efficiency given the ~200k sample constraint
• Dynamic model reset when reaching new levels

Setup Instructions

Prerequisites

• Python 3.10+
• CUDA-capable GPU (recommended)
• uv package manager

Step 1: Clone Repository

git clone --recurse-submodules [email protected]:DriesSmit/ARC3-solution.git
cd ARC3-solution

Step 2: Create Environment File

Copy the example environment file and set your API key (get your API key from https://three.arcprize.org/user):

cd ARC-AGI-3-Agents
cp .env-example .env
# Then edit .env file and replace the empty ARC_API_KEY= with your actual API key
cd ..

Step 3: Install Dependencies

make install

Step 4: Configure Submodule

Add the following code to ARC-AGI-3-Agents/agents/__init__.py (under the imports and before load_dotenv()):

import sys
import os
sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
from custom_agent import *

Also, add the following field to the FrameData class in ARC-AGI-3-Agents/agents/structs.py (after the full_reset field):

available_actions: list[GameAction] = Field(default_factory=list)

Step 5: Run the Action Agent

make action

Architecture

ActionModel (CNN)

• Input: 16-channel one-hot encoded frames (64x64)
• Backbone: 4-layer CNN (32→64→128→256 channels)
• Action Head: Predicts ACTION1-ACTION5 probabilities
• Coordinate Head: Predicts 64x64 click position probabilities for ACTION6 with 2D inductive bias using convolutional layers instead of flattened representations

Training

• Supervised Learning: (state, action) → frame_changed labels
• Experience Buffer: 200K unique state-action pairs with hash-based deduplication
• Dynamic Reset: Clears buffer and resets model when reaching new levels
• Loss: Binary cross-entropy with light entropy regularization

Exploration Strategy

• Stochastic Sampling: Uses sigmoid probabilities for action selection
• Hierarchical Selection: First sample action type, then coordinates if ACTION6
• Change Prediction: Biases exploration toward actions predicted to cause changes

Monitoring

The agent generates comprehensive logs and TensorBoard metrics:

# View training metrics
make tensorboard
# Open http://localhost:6006 in browser

Files Structure

ARC3/
├── ARC-AGI-3-Agents/      # Competition framework (submodule)
├── custom_agents/
│   ├── __init__.py        # Agent registration
│   ├── action.py          # Main action learning agent
│   └── view_utils.py      # Visualization utilities
├── custom_agents.py       # Agent imports
├── Makefile               # Build commands
├── README.md              # This file
├── requirements.txt       # Python dependencies
└── utils.py               # Shared utilities

Additional Usage Examples

# Standard competition run
make action

# Run with specific game ID
uv run ARC-AGI-3-Agents/main.py --agent=action --game=vc33

# View logs and metrics
make tensorboard

# Clean generated files
make clean