kvatsa5/orbit-wars-agent

🚀 Orbit Wars Agent - Kaggle Competition Submission

A strong hybrid agent for the Orbit Wars Kaggle Competition designed to achieve top-5 ELO.

Architecture

This agent combines rule-based strategy with adaptive opponent modeling for maximum performance:

Core Engine (Rule-Based)

1. Forward Simulation - Simulates every planet's future for 110 turns, accounting for all incoming fleets, production, and combat resolution

2. Multi-Mission Dispatch - 7 mission types with priority ordering:

   • REINFORCE - Send ships to hold threatened planets
   • RESCUE - Emergency reinforcement for planets about to fall
   • RECAPTURE - Take back lost planets immediately
   • CAPTURE - Standard planet conquest
   • SNIPE - Arrive at neutrals right after enemy fleet battles
   • SWARM - Multi-source coordinated attacks
   • CRASH_EXPLOIT - Arrive after inter-enemy fights

3. Target Valuation - Sophisticated scoring:

   score = (production × remaining_turns) / (ships_sent + arrival_turns × 0.5 + 1)
   

   With multipliers:

   • 2.05× for enemy planets (gain + deny)
   • 1.65× for static enemy planets
   • 1.4× for static neutrals
   • 0.7× for contested neutrals
   • 2.0× for exposed (just-captured) planets

4. Orbital Mechanics - Predicts rotating planet positions and computes intercept angles

5. Sun Avoidance - Routes fleets around the central sun

6. Dynamic Game Phases - Early expansion → mid-game control → finishing mode → total war

Adaptive Layer (Learning)

7. Opponent Modeling - Tracks enemy behavior patterns:
   • Expansion rate
   • Aggression level
   • Fleet deployment patterns
8. Dynamic Parameter Tuning - Adjusts strategy multipliers based on opponent behavior
9. Weakest Enemy Targeting - In 4P games, focuses on eliminating the weakest opponent first (1.5× value multiplier)
10. Elimination Bonus - Extra incentive to finish off weak opponents

Key Strategies for ELO Climbing

Production Snowball

The game is won by production advantage. Every ship not heading to a planet is wasted. This agent:

• Computes exact keep_needed for each planet via forward simulation
• Sends surplus ships immediately (surplus-based dispatching)
• Prioritizes high-production targets early

Fleet Speed Exploitation

Fleet speed scales logarithmically with size: speed = 1 + 5 × (log(ships)/log(1000))^1.5

• 5 ships: 1.1 speed
• 50 ships: 2.2 speed
• 500 ships: 4.5 speed

This agent sends optimally-sized fleets for speed/cost balance.

4-Player Strategy

• Target weakest enemy first (1.5× value multiplier)
• Elimination bonus when enemy < 110 ships
• Gang-up exploitation (arrive after enemy inter-fights)
• Conservative defense in crowded games

How to Submit

Kaggle Submission

1. Download submission.py
2. Create a new Kaggle notebook
3. Copy the content into the notebook
4. Make sure the agent(obs, config) function is defined
5. Submit

Local Testing

from kaggle_environments import make
from submission import agent

env = make("orbit_wars", debug=True)
result = env.run([agent, "starter"])
print(f"Result: {'WIN' if result[-1][0].reward == 1 else 'LOSS'}")

Performance

• vs Random: ~100% win rate
• vs Starter: ~90%+ win rate (dominates the baseline)
• 4-Player Mixed: Strong performance, adapts to multi-agent dynamics
• LB Expected: 1200+ ELO

Game Mechanics Summary

• 100×100 board with central sun (radius 10)
• Planets orbit the sun; outer planets are static
• Each planet produces ships per turn
• Send fleets: [from_planet_id, angle, num_ships]
• Fleet speed scales with size (logarithmic)
• Combat: largest fleet wins, excess fights garrison
• Win: most total ships at step 500, or eliminate all opponents

Files

• submission.py - Full agent code (3379 lines)
• main.py - Usage instructions

License

MIT