KQuity

KQuity powers live win-probability predictions in Hivemind and provides game analysis for Killer Queen — a 10-player arcade strategy game where teams of five race to win by military dominance, economic victory (berries), or snail ride.

Models

Win-probability model

A LightGBM classifier that predicts P(gold wins) from 52 in-game state features (berry counts, snail position, kills, warrior upgrades, etc.) extracted at each game event. A typical game produces 100–300 events, giving a real-time probability curve from start to finish. Trained on quality-filtered game data with symmetry augmentation.

Game quality classifier

Not all recorded games are competitive — many are casual warm-ups, kids mashing buttons, or half-empty cabinets. The quality classifier separates real games from junk using 69 hand-crafted features computed over the full event stream. It achieves an AUC of ~0.908 using logged-in games as positive examples and the unfiltered population as negatives, with tournament games anchoring the decision threshold. Its primary role is curating clean training data for the win-probability model.

Analysis

• Worker State Values — How much does worker composition matter? Bradley-Terry linearization to isolate the effect of upgrades (warrior, speed drone, speed warrior) on win probability.
• Lockout Analysis — When do teams lockout, how long does it last, and does it matter?

Documentation

• Training Guide — How the win-probability model is trained end-to-end
• Quality Classifier Report — Design and evaluation of the game quality classifier
• Experiment Log — Chronological log of modeling experiments
• Data Quality Report — Analysis of data filtering strategies
• Combined Scaling Experiment — Scaling behavior across data sources
• Symmetry Augmentation — Gold/blue swap augmentation results
• Data: Logged-In Games | Data: Quality Filtered — Dataset documentation