Raheel Yawar, Game Programmer | Flying Sheep Studios, Germany
In-game Content Generation using Machine Learning
Adapting gameplay difficulty and content based on player profiles is vital for better retention. The Free-to-play Dreamwork’s TrollhuntersRPG is taken as an example for an in-depth look at how Machine Learning algorithms can be employed to discover player profiles and present profile relevant content. The discussion includes implementation details on how these methods can be crafted and adapted for other game genres or use cases and how they can positively impact player retention.
Primarily the talk deals with a special breed of Machine Learning methods called Recommender Systems that can be used to take the load off the designers and learn from players about their skill level and preferences and use that information to tailor in-game content.
Presented by the
Serious Play Conference
seriousplayconf.com
at
Orlando,
University of Central Florida,
UCF,
July 24-26, 2019
2. Who am I?
● Raheel Yawar
○ Game Developer @ Flying Sheep Studios
○ Mobile and HTML5 game development
○ MSc Media Informatics
○ ML/ AI Enthusiast
● @raheelyawar
3. Flying Sheep Studios
● Founded in 2014 in Cologne Germany
● Over 150 cross-platform HTML5 games
● Worked with over 50 brands
● Team of 17 people
○ 47% women
○ 35% internationals
6. What is this about?
● Motivation
● Action RPG Introduction
● Machine Learning Methods
● Pros and Cons
● Results and Findings
Note: This talk is beginner friendly
10. The Game
● Single Player Hack n’ Slash Dungeon
Crawler Role Playing Game
● Free-2-Play
● Cross-platform HTML5
● Created using Three.js and Phaser
● Gameplay spans over
○ 30 levels
○ 9 story quests
○ Virtually infinite side-quests
http://www.toggo.de/spiele/trolljaeger/abenteuer-in-den-trollhoehlen-4774.htm
29. Collaborative Filtering
Users/ Movie Lego Movie Expendables Notebook
User 1 5 5 1
User 2 3 1 5
User 3 5 - 1
User 4 3 1 -
User/ Movie Ratings - Utility Matrix
30. Collaborative Filtering
Users/ Movie Lego Movie Expendables Notebook
User 1 5 5 1
User 2 3 1 5
User 3 5 - 1
User 4 3 1 -
Utility Matrix
Cosine Similarity of User 1 and User 3
41. Results – Retention Rate
Baseline
(PCG)
Content Based
(CB)
Neighbourhoo
d Oriented
(NO)
Matrix
Factorization
(MF)
Tensor
Factorization
(RTDD)
Day 1 8.21 8.64 7.79 8.92 8.93
Day 2 5.62 6.62 5.90 6.49 6.78
Day 3 3.96 4.64 4.40 4.55 4.74
Day 4 2.88 3.49 2.89 3.30 3.48
Day 5 2.02 2.44 1.82 1.94 2.30
Day 6 1.19 1.55 1.11 1.25 1.51
Day 7 0.25 0.54 0.61 0.47 0.72
Average 3.45 3.99 3.50 3.85 4.07
42. Summary
● Need for better tailored content
● ML/ AI algorithms can be used to:
○ Better identify player profiles
○ Tailor procedurally generated content
○ Matrix/ Tensor based approaches outperform rule-based approaches
43. References
● Matrix and Tensor Factorization Based
Game Content Recommender Systems:
A Bottom-Up Architecture and a
Comparative Online Evaluation, AIIDE
2018
● Clip Art: https://flaticon.com
44. Thank You
● Raheel Yawar
○ Game Developer
○ Flying Sheep Studios
● Website: http://raheelyawar.com/
● Twitter: @raheelyawar