Views
No views yet
| Module | Role | Capability |
|---|---|---|
| Perception Encoder | CNN over one-hot 64×64×16 frames → compact latent | — |
| World/Transition Model | Self-supervised P(frame changes | state, action) + forward model |
| State Memory Graph | Hash-deduplicated directed graph of observed states | Exploration |
| Intrinsic Reward | Extrinsic Δscore + graph novelty + prediction-error curiosity | Goal-setting |
| Planner | Short-horizon tree search using world model as simulator | Planning |
| Action Head | Hierarchical: action-type softmax + conv coordinate head for ACTION6 | — |
1# Install
2uv pip install -e ".[dev]"
3
4# Run against a public game
5uv run main.py --agent=wayfinder --game=ls20
6
7# Run tests
8uv run pytest
9
10# Offline evaluation
11uv run python eval/run_local_eval.py --agent=wayfinder --games=ls20,ls21,ls22agents/wayfinder/ # Core agent modules (perception, world_model, memory_graph, ...)
training/ # Replay buffer, training loops, configs
eval/ # Offline evaluation harness, metrics
notebooks/ # Kaggle submission notebook
tests/ # Unit + integration testsuv pip install -e ".[dev]"uv run python eval/run_local_eval.py --agent=wayfindereval/results/ with per-game/level breakdowns.