LLM Evaluation
Benchmark GPT, Claude, Gemini, and custom models on strategic reasoning, spatial awareness, and multi-step planning.
RL Training
Full Gymnasium environment with multi-discrete action space, configurable reward shaping, and headless mode for fast training.
Tournaments
Automated round-robin tournaments with ELO ratings, replay recording, and detailed performance analytics.
Diverse Tactical Maps
14 hand-crafted maps across 1v1, 1v1v1, and 2v2 formats with varied terrain, chokepoints, and strategic objectives.
Crossroads
Island Fortress
Tower Rush
Center Mountains8 Unique Unit Types
Each unit has distinct stats, abilities, and roles — creating a rich decision space for AI agents to master.
How It Works
Install
Install via pip with optional GPU, GUI, and LLM extras. Works on Python 3.10+.
pip install reinforcetactics[llm]Configure
Pick your agents — LLM bots, RL models, rule-based bots, or your own custom agent.
--agents gpt-4o claude-sonnetCompete
Run tournaments, compare ELO ratings, analyze replays, and iterate on your models.
python -m reinforcetactics tournamentBuilt for AI Research
Gymnasium Compatible
Standard RL interface with observation and action spaces, reward shaping, and episode management.
Multi-Agent Support
PettingZoo integration for multi-agent RL. Train cooperative and competitive policies.
Replay & Analysis
Record battles, export to video, and analyze decision patterns for model interpretability.
Extensible Architecture
Add custom units, maps, reward functions, and AI agents with a clean Python API.
Multiple AI Backends
OpenAI, Anthropic, and Google Gemini SDKs built-in. Plug in any LLM via API.
Docker Tournaments
Containerized tournament runner for reproducible benchmarks at scale.
Ready to benchmark your AI?
Open source and ready for research. Clone the repo and run your first tournament in minutes.