Detail	Value
Type	`Pipeline(StandardScaler + RandomForestRegressor)`
Task	Tabular regression - predict average win rate
Framework	scikit-learn 1.8.0
Training data	Simulated gameplay (custom engine)

Features (inputs)

#	Feature	Description
0	AI	AI-themed card count
1	Nano	Nano-themed card count
2	Quantum	Quantum-themed card count
3	Base	Base card count
4-8	0_cost to 4_plus_cost	Cost curve distribution
9-11	equipment, item, action	Card type counts
12	avg_turns	Average game length
13	avg_evos	Average Evo equipment used
14	avg_teklo_energy	Avg Teklo Energy generated
15	avg_nanite_counters	Avg Nanite counters
16	avg_quantum_charges	Avg Quantum charges
17	consistency_score	Deck consistency (0-1)
18	avg_expert_balance_score	Expert balance score (1-10)

Output

avg_win_rate - float in [0, 1] representing predicted deck win rate.

Metrics

Metric	Value
Train R2	0.917
CV R2 (mean +/- std)	-1.3332 +/- 2.1072
Train MSE	0.0049
Train MAE	0.042
Samples	9
Features	17

Quick Start

python

from huggingface_hub import hf_hub_download
import joblib, numpy as np

path = hf_hub_download(repo_id="4math/FAB_Prediction_Model", filename="model.joblib")
model = joblib.load(path)

# Example: a Nano-focused deck
features = np.array([[9,0,0,2, 5,2,1,0, 9,9,12, 0.6, 9.4,13.6,7.3, 0.74, 6.7, 0,0]])
print(model.predict(features))  # predicted win rate

About This Project

This is a fan-made educational project exploring ML for TCG game balance. The model is trained on simulated gameplay data from a custom Flesh and Blood game engine featuring Teklovossen, a Mechanologist hero with custom Evo cards across four tech themes: AI, Nano, Quantum, and Biomancy.

Limitations

Trained on simulated data, not real tournament results
Small dataset (9 samples) - treat predictions as rough estimates
Simplified card mechanics in the simulation engine
Fan-made cards only - does not cover the official card pool

License

MIT