Bot Tournaments
Welcome to the Reinforce Tactics Bot Tournament page! This page tracks official tournament results between different bot types, showcasing the performance of rule-based bots, LLM-powered bots, and trained RL models.
What are Tournaments?
Tournaments in Reinforce Tactics pit different bot types against each other in a round-robin format. Each matchup consists of multiple games where bots alternate playing as Player 1 and Player 2 to account for first-move advantage.
Tournament Format
- Round-Robin: Every bot plays against every other bot
- Fair Play: Equal games with each bot as Player 1 and Player 2
- Map: Typically played on the 6x6 beginner map for quick matches
- Results: Win/Loss/Draw records with win rate statistics
Bot Types
SimpleBot
A rule-based AI that follows a simple strategy:
- Purchases the most expensive units first (one per turn)
- Prioritizes capturing buildings and towers
- Attacks nearby enemy units
- Always included in tournaments
MediumBot
An improved rule-based AI with more sophisticated strategies:
- Maximizes unit purchases: Buys multiple units per turn instead of just one
- Coordinated attacks: Plans focus-fire to kill enemies in one turn
- Smart prioritization: Captures structures closest to its own HQ first
- Interrupts captures: Attacks enemies that are capturing structures
- Efficient combat: Evaluates damage-per-cost to make better attack decisions
- Included in tournaments by default alongside SimpleBot
AdvancedBot
A highly sophisticated AI extending MediumBot with enhanced strategic capabilities:
- Inherits from MediumBot: Builds upon MediumBot's proven coordinated attack and capture strategies
- Enhanced unit composition: Optimized purchasing (35% Warriors, 30% Archers, 25% Mages, 10% Clerics)
- Mountain positioning: Positions units on mountains before attacking for bonus damage
- Smart ranged combat: Prioritizes ranged attacks with Archers and Mages to minimize counter-attack damage
- Special ability usage: Effectively uses Mage Paralyze on high-value targets and Cleric Heal on damaged frontline units
- Map awareness: Analyzes HQ locations and defensive mountain positions on first turn
- Aggressive tactics: More aggressive attack prioritization and pursuit of enemies
- Included in tournaments by default alongside SimpleBot and MediumBot
LLM Bots
AI-powered by Large Language Models:
- OpenAIBot: Powered by GPT models
- ClaudeBot: Powered by Anthropic's Claude
- GeminiBot: Powered by Google's Gemini
- Uses natural language reasoning to make strategic decisions
Model Bots
Trained using Reinforcement Learning:
- Uses Stable-Baselines3 (PPO, A2C, or DQN)
- Trained through self-play and opponent challenges
- Learns optimal strategies through experience
Running Your Own Tournament
Want to run a tournament? It's easy! See the Tournament System guide for detailed instructions.
Quick start:
python3 scripts/tournament.py
Official Tournament Results
Below are the results from official tournaments run on the Reinforce Tactics platform. Results are organized by game version.
Version 0.1.0
Tournament #1 - December 19, 2025
Configuration:
- Games per matchup: 2 per side (16 total games per bot pair across all maps)
- Map pool mode: All maps played
- Starting ELO: 1500
Maps Used:
beginner.csv
6×6 • 2 Players
funnel_point.csv
13×11 • 2 Players
center_mountains.csv
18×18 • 2 Players
corner_points.csv
12×10 • 2 Players
Map Legend:
- 🟩 Green - Plains (walkable terrain)
- 🟦 Blue - Ocean/Water (impassable)
- ⬜ Light Gray - Mountains (impassable, provides attack bonus)
- 🌲 Dark Green - Forest (provides defense bonus)
- 🟫 Brown - Roads (faster movement)
- 🔴 Red - Player 1 structures (HQ, Buildings)
- 🔵 Blue Tint - Player 2 structures (HQ, Buildings)
- ⬜ White/Gray - Neutral Towers/Buildings
Participants:
- SimpleBot (rule-based)
- MediumBot (rule-based)
- AdvancedBot (rule-based)
- Claude Haiku 4.5 (claude-haiku-4-5-20251001)
Final Standings
| Rank | Bot Name | Wins | Losses | Draws | Total Games | Win Rate | ELO | ELO Change |
|---|---|---|---|---|---|---|---|---|
| 🥇 1 | AdvancedBot | 36 | 2 | 10 | 48 | 75.0% | 1693 | +193 |
| 🥈 2 | MediumBot | 19 | 10 | 19 | 48 | 39.6% | 1575 | +75 |
| 🥉 3 | SimpleBot | 1 | 20 | 27 | 48 | 2.1% | 1405 | -95 |
| 4 | Claude Haiku 4.5 | 2 | 26 | 20 | 48 | 4.2% | 1327 | -173 |
Head-to-Head Results
| Matchup | Wins | Losses | Draws | Notes |
|---|---|---|---|---|
| SimpleBot vs MediumBot | 0 | 6 | 10 | MediumBot dominates |
| SimpleBot vs AdvancedBot | 0 | 12 | 4 | Complete sweep by AdvancedBot |
| SimpleBot vs Claude Haiku 4.5 | 1 | 2 | 13 | Mostly draws, slight LLM edge |
| MediumBot vs AdvancedBot | 2 | 10 | 4 | AdvancedBot clearly superior |
| MediumBot vs Claude Haiku 4.5 | 11 | 0 | 5 | MediumBot perfect record |
| AdvancedBot vs Claude Haiku 4.5 | 14 | 0 | 2 | Complete domination |
Per-Map Performance
beginner.csv (6×6)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| AdvancedBot | 12 | 0 | 0 |
| MediumBot | 4 | 4 | 4 |
| Claude Haiku 4.5 | 2 | 7 | 3 |
| SimpleBot | 1 | 8 | 3 |
funnel_point.csv (13×11)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| MediumBot | 8 | 0 | 4 |
| AdvancedBot | 6 | 2 | 4 |
| SimpleBot | 0 | 4 | 8 |
| Claude Haiku 4.5 | 0 | 8 | 4 |
center_mountains.csv (18×18)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| AdvancedBot | 6 | 0 | 6 |
| MediumBot | 1 | 2 | 9 |
| SimpleBot | 0 | 2 | 10 |
| Claude Haiku 4.5 | 0 | 3 | 9 |
corner_points.csv (12×10)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| AdvancedBot | 12 | 0 | 0 |
| MediumBot | 6 | 4 | 2 |
| SimpleBot | 0 | 6 | 6 |
| Claude Haiku 4.5 | 0 | 8 | 4 |
Analysis
Key Findings:
-
AdvancedBot dominance: The AdvancedBot performed exceptionally well, achieving a 75% win rate and gaining +193 ELO. Its sophisticated strategies (mountain positioning, ranged combat, special abilities) proved highly effective.
-
MediumBot solid performance: With a 39.6% win rate and +75 ELO gain, MediumBot showed the value of coordinated attacks and maximized unit production over SimpleBot's basic approach.
-
LLM bot struggles: Claude Haiku placed last with only a 4.2% win rate (-173 ELO). This suggests that current LLM reasoning may not be optimized for the tactical decision-making required in this game format.
-
High draw rates: Many matches ended in draws, particularly on center_mountains.csv (defensive terrain) and between lower-performing bots. This indicates the game's strategic depth and defensive viability.
-
Map influence: AdvancedBot dominated on beginner.csv and corner_points.csv (100% win rate), while MediumBot actually outperformed on funnel_point.csv.
Notable Observations:
- AdvancedBot never lost on 3 of 4 maps
- The hierarchy is clear: AdvancedBot > MediumBot > SimpleBot/Claude Haiku
- Center mountains map produces the most draws (defensive terrain advantage)
Version 0.1.1
Tournament #1 - December 29, 2025
Configuration:
- Games per matchup: 8 games per bot pair (1 per side per map)
- Map pool mode: All maps played
- Starting ELO: 1500
- Bots: 8 participants (3 rule-based, 5 LLM-powered)
Maps Used:
beginner.csv
6×6 • 2 Players
funnel_point.csv
13×11 • 2 Players
center_mountains.csv
18×18 • 2 Players
corner_points.csv
12×10 • 2 Players
Participants:
- SimpleBot (rule-based)
- MediumBot (rule-based)
- AdvancedBot (rule-based)
- Claude Haiku 4.5 (claude-haiku-4-5-20251001)
- Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
- ChatGPT 5 Mini (gpt-5-mini-2025-08-07)
- Gemini 2.0 Flash (gemini-2.0-flash)
- Gemini 3.0 Flash (gemini-3-flash-preview)
Final Standings
| Rank | Bot Name | Wins | Losses | Draws | Total Games | Win Rate | ELO | ELO Change |
|---|---|---|---|---|---|---|---|---|
| 🥇 1 | AdvancedBot | 34 | 1 | 21 | 56 | 60.7% | 1702 | +202 |
| 🥈 2 | Gemini 3.0 Flash | 33 | 4 | 19 | 56 | 58.9% | 1646 | +146 |
| 🥉 3 | MediumBot | 26 | 8 | 22 | 56 | 46.4% | 1618 | +118 |
| 4 | Claude Sonnet 4.5 | 19 | 13 | 24 | 56 | 33.9% | 1508 | +8 |
| 5 | SimpleBot | 1 | 20 | 35 | 56 | 1.8% | 1410 | -90 |
| 6 | Claude Haiku 4.5 | 3 | 22 | 31 | 56 | 5.4% | 1389 | -111 |
| 7 | ChatGPT 5 Mini | 2 | 22 | 32 | 56 | 3.6% | 1386 | -114 |
| 8 | Gemini 2.0 Flash | 0 | 28 | 28 | 56 | 0.0% | 1341 | -159 |
Head-to-Head Results
| Matchup | Wins | Losses | Draws | Notes |
|---|---|---|---|---|
| AdvancedBot vs Gemini 3.0 Flash | 4 | 0 | 4 | AdvancedBot undefeated |
| AdvancedBot vs MediumBot | 5 | 1 | 2 | AdvancedBot dominant |
| AdvancedBot vs Claude Sonnet 4.5 | 5 | 0 | 3 | Complete domination |
| Gemini 3.0 Flash vs MediumBot | 3 | 0 | 5 | Gemini edges out |
| Gemini 3.0 Flash vs Claude Sonnet 4.5 | 4 | 0 | 4 | Gemini superior |
| MediumBot vs Claude Sonnet 4.5 | 4 | 0 | 4 | MediumBot perfect |
| Claude Sonnet 4.5 vs Claude Haiku 4.5 | 5 | 0 | 3 | Sonnet dominates Haiku |
| Claude Sonnet 4.5 vs ChatGPT 5 Mini | 4 | 0 | 4 | Sonnet edges GPT |
| Claude Sonnet 4.5 vs Gemini 2.0 Flash | 6 | 0 | 2 | Sonnet sweeps |
| Claude Haiku 4.5 vs ChatGPT 5 Mini | 0 | 0 | 8 | Complete stalemate |
| Gemini 2.0 Flash vs SimpleBot | 0 | 0 | 8 | All draws |
Per-Map Performance
beginner.csv (6×6)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| AdvancedBot | 14 | 0 | 0 |
| Gemini 3.0 Flash | 12 | 2 | 0 |
| Claude Sonnet 4.5 | 7 | 5 | 2 |
| MediumBot | 6 | 4 | 4 |
| ChatGPT 5 Mini | 2 | 8 | 4 |
| Claude Haiku 4.5 | 2 | 8 | 4 |
| SimpleBot | 1 | 8 | 5 |
| Gemini 2.0 Flash | 0 | 9 | 5 |
funnel_point.csv (13×11)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| Gemini 3.0 Flash | 11 | 1 | 2 |
| MediumBot | 9 | 1 | 4 |
| Claude Sonnet 4.5 | 8 | 5 | 1 |
| AdvancedBot | 4 | 1 | 9 |
| Claude Haiku 4.5 | 1 | 6 | 7 |
| SimpleBot | 0 | 6 | 8 |
| ChatGPT 5 Mini | 0 | 6 | 8 |
| Gemini 2.0 Flash | 0 | 7 | 7 |
center_mountains.csv (18×18)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| AdvancedBot | 4 | 0 | 10 |
| Claude Sonnet 4.5 | 3 | 0 | 11 |
| Gemini 3.0 Flash | 2 | 0 | 12 |
| MediumBot | 2 | 1 | 11 |
| SimpleBot | 0 | 1 | 13 |
| ChatGPT 5 Mini | 0 | 2 | 12 |
| Claude Haiku 4.5 | 0 | 2 | 12 |
| Gemini 2.0 Flash | 0 | 5 | 9 |
corner_points.csv (12×10)
| Bot | Wins | Losses | Draws |
|---|---|---|---|
| AdvancedBot | 12 | 0 | 2 |
| MediumBot | 9 | 2 | 3 |
| Gemini 3.0 Flash | 8 | 1 | 5 |
| Claude Sonnet 4.5 | 1 | 3 | 10 |
| SimpleBot | 0 | 5 | 9 |
| ChatGPT 5 Mini | 0 | 6 | 8 |
| Claude Haiku 4.5 | 0 | 6 | 8 |
| Gemini 2.0 Flash | 0 | 7 | 7 |
LLM Token Usage
| Model | Input Tokens | Output Tokens | Total Tokens | Games |
|---|---|---|---|---|
| Gemini 3.0 Flash | 24,174,240 | 1,295,618 | 25,469,858 | 58 |
| ChatGPT 5 Mini | 13,899,718 | 4,508,521 | 18,408,239 | 57 |
| Claude Sonnet 4.5 | 16,585,568 | 695,199 | 17,280,767 | 56 |
| Claude Haiku 4.5 | 16,116,414 | 552,455 | 16,668,869 | 56 |
| Gemini 2.0 Flash | 12,057,436 | 402,555 | 12,459,991 | 56 |
Analysis
Key Findings:
-
AdvancedBot maintains dominance: With a 60.7% win rate and +202 ELO, AdvancedBot proves its sophisticated strategies remain effective even against a larger, more diverse field of competitors.
-
Gemini 3.0 Flash emerges as top LLM: Gemini 3.0 Flash achieved an impressive 58.9% win rate (+146 ELO), becoming the first LLM bot to finish in the top 3. It went undefeated against all other LLM bots.
-
Claude Sonnet 4.5 shows promise: With a 33.9% win rate, Claude Sonnet significantly outperformed other Claude and OpenAI models, demonstrating improved tactical reasoning over Haiku.
-
LLM tier separation: Clear performance tiers emerged among LLMs: Gemini 3.0 Flash (top tier), Claude Sonnet 4.5 (mid tier), and Claude Haiku/ChatGPT 5 Mini/Gemini 2.0 Flash (bottom tier).
-
Map-specific patterns:
- beginner.csv: Decisive outcomes with few draws - fast-paced combat favors aggressive bots
- center_mountains.csv: Extremely high draw rate (75%+ for most bots) - defensive terrain creates stalemates
- funnel_point.csv: Gemini 3.0 Flash and MediumBot excelled with 11 and 9 wins respectively
Notable Observations:
- AdvancedBot's only loss came from MediumBot (1 game out of 56)
- Gemini 2.0 Flash won 0 games across all matchups - the only bot with no wins
- Claude Haiku 4.5 vs ChatGPT 5 Mini resulted in 8 draws - complete stalemate
- Token efficiency: Claude models used ~300K tokens per game vs ChatGPT's ~323K per game
How to Contribute Results
Have you run a tournament? Share your results with the community!
- Run a tournament using
scripts/tournament.py - Save the results (CSV/JSON files)
- Submit a pull request with your results
- Include replay files for verification
Insights and Analysis
Strategy Patterns
As tournaments are completed, we'll analyze:
- Opening strategies and unit compositions
- Economic vs. military balance
- Successful tactical patterns
- Common mistakes and pitfalls
Bot Strengths and Weaknesses
Each bot type has different characteristics:
- SimpleBot: Predictable but consistent; makes basic strategic errors
- MediumBot: More strategic than SimpleBot; uses coordinated tactics and better unit purchasing
- AdvancedBot: Extends MediumBot with enhanced unit composition, mountain positioning, ranged combat prioritization, and special ability usage
- LLM Bots: Creative but sometimes unpredictable
- Model Bots: Optimized but may overfit to training conditions
Map-Specific Performance
Different maps favor different strategies:
- Small maps (6x6): Fast, aggressive play
- Medium maps (10x10): Balanced economic/military
- Large maps (24x24+): Long-term strategy and positioning
Tournament Archive
All tournament replays are saved and can be watched using the game's replay system:
# Load a replay file
python3 main.py --mode play --replay path/to/replay.json
Replay files include:
- Complete action history
- Game metadata (bots, map, duration)
- Can be analyzed programmatically
Future Tournaments
Upcoming tournament ideas:
- Map Variety: Tournaments on different map sizes and layouts
- Specialized Competitions: Economy-focused, combat-focused, speed-run
- Evolution: Tournament of evolved/improved model versions
- Human vs. Bot: Special exhibitions with human players
Resources
- Tournament System Guide - Technical documentation
- Implementation Status - Current features
- Getting Started - Learn how to play
Last updated: December 2025
Tournament results will be updated as official competitions are completed.