| # 🛸 Orbit Wars - Kaggle Competition Agent (v2 Adaptive) |
|
|
| **Competition:** [Orbit Wars](https://www.kaggle.com/competitions/orbit-wars) ($50,000 prize pool) |
| **Deadline:** June 23, 2026 |
| **Current ELO:** ~1100 and climbing |
|
|
| ## Overview |
|
|
| This is a highly competitive **adaptive** agent for the Orbit Wars Kaggle competition — a real-time strategy game where 2 or 4 AI agents compete to conquer planets orbiting a central sun in continuous 2D space. |
|
|
| **v2 adds real-time in-match opponent profiling and adaptive parameter tuning — the agent learns opponent playstyle during each game and adjusts its strategy accordingly.** |
|
|
| ## Architecture |
|
|
| ### Core: Composite Rule-Based Engine |
| Combined from the **5 top public agents** on the leaderboard: |
|
|
| | Source | LB Rating | Key Feature | |
| |--------|-----------|-------------| |
| | tamrazov-starwars (base) | LB 1224 | Gang-up attacks, weakest enemy targeting, elimination missions | |
| | ykhnkf | LB #1 | Hostile reinforcement prediction | |
| | pascal v14 | High-rated | 4-source coordinated swarm attacks | |
| | pilkwang | LB ~1000 | Structured decision architecture | |
| | yuriygreben | Architect | Physics-aware multi-phase strategy | |
|
|
| ### v2: In-Match Opponent Profiling & Adaptation |
|
|
| New adaptive layer that monitors opponent behavior in real-time: |
|
|
| **What it tracks (EMA-smoothed):** |
| - **Aggression** — fleet launch frequency (how often they attack) |
| - **Expansion rate** — planet capture speed |
| - **Relative strength** — ship/planet differential |
|
|
| **How it adapts:** |
|
|
| | Opponent Style | Agent Response | |
| |---------------|---------------| |
| | Very aggressive (aggression > 0.6) | ↑ defense ratios, ↑ reinforcement priority, ↓ attack aggression | |
| | Passive/turtle (aggression < 0.3) | ↑ attack multipliers, ↑ elimination bonus, ↑ expansion pressure | |
| | We're ahead | Play safe, consolidate, higher attack cost weighting | |
| | We're behind | Take risks, ↑ snipe values, ↑ finishing bonuses, lower defense | |
| | Enemy expanding fast | Contest neutrals more aggressively, ↓ target margins | |
| | Late game (step > 350) | Maximum elimination drive, ↑ finishing multipliers | |
|
|
| **20 parameters dynamically tuned** during each match based on game state. |
|
|
| ### Key Technical Features |
|
|
| 1. **Hostile Reinforcement Prediction** — estimates enemy counterattack potential before committing fleets |
| 2. **4-Source Coordinated Swarms** — synchronizes 4 fleets to overwhelm defended targets |
| 3. **Multi-Phase Target Selection** — opening expansion → mid-game optimization → late-game elimination |
| 4. **Sun-Aware Fleet Routing** — avoids solar destruction with safe detour angles |
| 5. **Crash Exploit Detection** — captures planets weakened by enemy fleet collisions |
| 6. **Doomed Planet Evacuation** — retreats from unsaveable positions to useful targets |
|
|
| ## Performance |
|
|
| | Opponent | Win Rate | Notes | |
| |----------|----------|-------| |
| | Random | **100%** (3/3) | Eliminated by step ~94-150 | |
| | Nearest-Sniper | **100%** (4/4) | Eliminated by step ~88-152 | |
| | 3× Random (4P) | **100%** | All eliminated by step ~123 | |
|
|
| ## Usage |
|
|
| ### Direct Kaggle Submission |
| ```bash |
| wget https://huggingface.co/Builder-Neekhil/orbit-wars-agent/resolve/main/submission.py |
| kaggle competitions submit orbit-wars -f submission.py -m "v2 adaptive agent" |
| ``` |
|
|
| ### Local Testing |
| ```python |
| from kaggle_environments import make |
| exec(open('submission.py').read(), globals()) |
| |
| env = make("orbit_wars", configuration={"seed": 42}, debug=False) |
| env.run([agent, "random"]) |
| final = env.steps[-1] |
| print(f"P0: {final[0].reward}, P1: {final[1].reward}") |
| ``` |
|
|
| ## Self-Play PPO Training (Optional) |
|
|
| The repo includes a PPO self-play training pipeline for further improvement: |
|
|
| ```bash |
| pip install torch numpy pyyaml kaggle-environments huggingface_hub |
| |
| # Train (requires GPU for reasonable speed, ~10h on T4) |
| TOTAL_UPDATES=500 EPISODES_PER_UPDATE=4 python train_efficient.py |
| ``` |
|
|
| **Training approach** (based on the Artificial Generals Intelligence paper, arXiv:2507.06825): |
| - **Phase 1 (0-20%)**: Train vs random opponents (fast, learn basics) |
| - **Phase 2 (20-50%)**: Train vs baseline agent (harder, learn tactics) |
| - **Phase 3 (50-100%)**: Self-play with opponent pool (N=3, argmax opponents) |
| - **Reward**: Potential-based shaping (planets + ships + production differential) |
| - **Architecture**: 128-d MLP controller that outputs 20 parameter adjustments |
|
|
| ## Files |
| - `submission.py` — Complete adaptive agent (single-file, ready for Kaggle) |
| - `train_efficient.py` — PPO self-play training script |
| - `generate_submission.py` — Packages trained controller into submission file |
|
|
| ## License |
| MIT |
|
|
