Training neural network agents to play Slay the Spire using supervised learning.

This project uses sts-lightspeed, a high-performance RNG-accurate C++ simulator, to generate training data and evaluate agents. The ultimate goal is to build an agent that can eventually play Ascension 20 Ironclad.

This currently is a supervised learning pipeline that trains neural network agents to play Slay the Spire:

1. Generate diverse battle scenarios from real game data
2. Run a baseline agent to collect demonstrations
3. Extract state-action pairs from battle snapshots
4. Train a neural network to imitate the baseline agent's behavior
5. Export the model for C++ inference and evaluation

Currently focused on benchmarking different approaches.

• C++17 compiler and CMake 3.10+ (for the sts-lightspeed simulator)
• just (recommended for running commands)
• Python 3.8+ with uv

just build

Create randomized battle scenarios from real game data:

uv run randomize_scenarios.py --count 20

This creates variations of base scenarios by randomizing HP, deck composition, and RNG seeds.

Run SimpleAgent on scenarios and capture battle snapshots:

just run-agent simple --snapshot --scenario=jaw_worm

Battle progression data is saved to data/agent_battles/simpleagent/.

Extract state-action pairs from battle snapshots:

cd AutoClad
uv run data_parser.py

Creates jaw_worm_data.npz with feature vectors (game state) and action labels (which card was played).

cd AutoClad
# Train with plotting and early stopping
uv run main.py --plot --early-stopping
# Or with default settings
uv run main.py

The trained model is exported as jaw_worm_model_traced.pt (TorchScript format) for C++ inference.

# Requires LibTorch
LIBTORCH_PATH=~/Downloads/libtorch just run-agent neural --scenario=jaw_worm

Working:

• Data generation from baseline agent demonstrations
• Neural network training on Jaw Worm encounters
• C++ inference with trained models

In Progress:

• Expanding to more encounter types
• Improving agent performance metrics
• Recording intermediate combat states for richer training data

The end goal is to train an agent that can play the full game at a high level. The approach:

1. Start with battles: If you can predict whether you'll win specific fights, you can make better decisions about everything else (pathing, card choices, shops, etc.)

2. Generate quality data: Available human gameplay data is incomplete and outdated. Instead, use tree search and simpler agents to generate training data.

3. Scale up with RL: Once we have a decent baseline from supervised learning, use reinforcement learning to reach superhuman play.

├── AutoClad/                    # Neural network training code (Python)
│   ├── main.py                  # Training script
│   ├── data_parser.py           # Parse battle snapshots to training data
│   └── *.pth, *.pt              # Trained models
├── randomize_scenarios.py       # Generate randomized battle scenarios
├── battle/
│   ├── generated_scenarios/     # Base scenarios from real games
│   └── randomized_scenarios/    # Training scenario variations
├── data/agent_battles/          # Battle snapshots from agent runs
└── [sts-lightspeed files]       # C++ simulator (apps/, src/, include/, etc.)