RS10Env

English | 简体中文

Gymnasium-compatible RS10 board game environment and heuristic strategies (PyTorch).

PyPI: rs10env · GitHub: xx025/rs10env

Install

By default the project uses CPU-only PyTorch (smaller install; no GPU/CUDA needed if you are not training):

# With uv (recommended; installs torch from PyTorch CPU index)
uv add rs10env

# Or from source: clone repo, then sync in project root
git clone https://github.com/xx025/rs10env.git
cd rs10env
uv sync

For GPU/CUDA, install the matching torch first, then rs10env, e.g.
uv pip install torch --index-url https://download.pytorch.org/whl/cu124, then uv sync.

Environment

RS10Env is a turn-based env: each step selects a rectangle whose cells sum to a target (default 10) and clears it. Board: H×W (default 16×10), cells 0–9. Full spec (observation/action space, mask, reward): API reference.

API usage

See docs/API.md for environment API and code examples (single board, multi-game comparison, low-level env + strategy).

Streamlit app

Two modes: single board (multiple strategies) and multi-game comparison (with progress). Best strategy is highlighted.

uv add rs10env[app]
rs10env-app
# or from repo root
uv run streamlit run app.py

Strategies

• random — uniform over valid actions
• greedy — clear as many cells as possible
• center_bias — prefer rectangles closer to board center
• large_rect / small_rect — prefer larger / smaller area
• center_small_rect — center + small area
• epsilon_greedy — ε-greedy
• max_future_moves — choose action that maximizes valid moves on the next state

Benchmark (strategy comparison)

Batch runs on a fixed set of boards (16×10, target_sum=10). Below: 100k games per strategy. “Best count” / “Share” = times that strategy cleared the most cells in a game.

Strategy	Tests	Avg steps	Avg removed	Time/game (s)	Best count	Share (%)
random	100,000	40.72	96.82	0.32	1,109	1.11
greedy	100,000	36.38	92.11	0.31	244	0.24
center_bias	100,000	43.15	101.15	0.31	5,185	5.19
large_rect	100,000	36.33	91.57	0.31	212	0.21
small_rect	100,000	45.77	103.44	0.39	7,316	7.32
center_small_rect	100,000	46.31	104.65	0.52	10,643	10.64
max_future_moves	100,000	50.07	113.54	8.12	80,793	80.79

max_future_moves clears the most on average and wins most often, at higher per-game cost; center_small_rect and small_rect offer a good trade-off.

Cells removed by strategy (boxplot):

Best-strategy share (who clears the most in each game):

Cumulative average cells removed over games:

Dependencies

• Python >= 3.10
• PyTorch >= 2.0
• Gymnasium >= 1.0
• NumPy >= 1.24

Related projects

This repo focuses on simulation environment and strategies (Gymnasium + heuristics). Other open-source projects that implement automation or assist tools for similar number-sum-elimination mechanics (various platforms):

Project	Platform	Description
nusery (longlifedahan)	Web	HTML/JS frontend, playable demo
Opening_Nursery_For_Mac	macOS	Python, pyautogui + OpenCV + Tesseract
nursery-bot	Windows	Python bot, Tesseract OCR
KaiJuTuoErSuo	Android	Java + ADB, OpenCV, OCR, DFS for elimination path
tuoersuo	—	Automation script (Gitee)