Predicting Game Difficulty and Churn Without Players
summary
Combines Deep Reinforcement Learning gameplay with a simple population simulator (skill, persistence, boredom) to predict per-level pass and churn for Angry Birds Dream Blast using a dataset of over 90,000 players and 100+ levels. DRL agents are trained once per level to produce difficulty estimates, a population model simulates churn (players leave if attempts exceed persistence or from boredom), and 5-fold cross validation shows churn prediction improves over a baseline by about 2 standard deviations.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:37The Problem
- 1:03The Data
- 1:35AI Agents
- 2:48Neural Network
- 3:36Input Features
- 5:17Summary
related talks
23mCustom-made games with machine learning and big data
01h 47mGamified Reinforcement Learning: Mastering AI through Video Games
27mUsing Reinforcement Learning to Optimize IAP Offer Recommendations in Mobile Games
39mMachine Learning and Online Experiments to Prevent Churn on GloboPlay
31mDesigning game AI with reinforcement learning
49mData Science in Gaming
47mMachine Learning for Gamers - Dungeon Forecasts & Dragon Regressions
30mDetecting Hidden Bias in Datasets Before Models Fail