APV Solver Challenge — Call for Engine Contributions

[andrewrgarcia]

APV Solver Challenge — Engine Development Call

This repository includes a baseline self-play agent for APV (Axis · Pivot · Veil),
a deterministic, perfect-information hex strategy game.

This challenge invites contributors to develop stronger APV engines —
comparable in spirit to chess engines like Stockfish or Go engines like AlphaZero —
and benchmark them against the baseline greedy agent included in this Simulator.

The goal is straightforward:

Build an APV solver that consistently beats the Simulator’s baseline agent.

---

Requirements

Your solver must:

• Comply with all APV rules:

• Hexagonal radius-4 board (61 cells)
• Piece types: Pivot, Axis, Veil, Veil90
• No captures; victory only by immobilizing the opposing Pivot
• One move per turn; perfect information

• Provide a move-selecting function compatible with:

choose_best_move(for_player, pieces, grid, ...)

• Operate deterministically (no stochastic policy hidden inside search)
• Never modify game state outside legal movement rules

---

Evaluation Protocol

Benchmark using:

python simulator/run_sim.py

Default experimental setup:

• 200 games per batch
• Randomized starting player
• 50% mirrored boards to cancel spatial bias
• Turn cap: 500 (draw condition)
• Pivot containment as win condition

Success criterion:

Demonstrate a statistically meaningful performance advantage vs. the baseline.

You may increase number of games for higher statistical confidence.

---

Submissions

When ready:

1. Publish your solver in your own public repository

2. Reply to this Issue with:

   • Repository link
   • Win-rate results vs. baseline agent
   • Short description of your method (search depth, pruning, learning model, etc.)

Optional improvements:

• Elo-style strength estimates
• Multi-seed tournament logs
• Visualizations or evaluation diagnostics

---

Recognition & Deployment

Approved engines will be:

✔ Listed under Recognized Engines in the Simulator README
✔ Included in future engine-vs-engine tournaments
✔ Considered for deployment on the official APV playable site

Top-performing ML-based engines may additionally be:

✔ Exported to ONNX
✔ Served as a selectable AI opponent in the live web client

(With credit and reference to the author.)

---

Why Participate?

This challenge supports research in:

• Engine design for new perfect-information games
• Geometric containment as a strategic model
• Practical comparisons of heuristic vs. learning-based systems
• Compact competitive agents that scale to strong play

APV offers minimalism with depth — no combat, only positional dominance.

---

Resources

Simulator + evaluation API:
https://github.com/andrewrgarcia/apv/blob/main/simulator

Official APV rules + design context:
https://github.com/andrewrgarcia/apv/tree/main