1. Adaptive Games
Content Generation
“Mario”
Mohammad Shaker
Department of Artificial Intelligence
IT University of Damascus
Seminar of Artificial Neural Networks
ZGTR
2. Outline
• Readings
• Motivation
• The proposed approach
• Experiments
• ANN Implementation
• Results
• Conclusion
3. Readings
• Towards Automatic Personalized Content
Generation for Platform Games
Noor Shaker, Georgios N. Yannakakis, Member, IEEE, and Julian Togelius,
Member, IEEE
• Feature Analysis for Modeling Game Content
Quality
Noor Shaker, Georgios N. Yannakakis, Member, IEEE, and Julian Togelius,
Member, IEEE
21. Open Questions!
Session period? (frequency of adaptation)
The most useful information about game content?
Game aspects with major affect on player
experience?
22. Open Questions!
Session period? (frequency of adaptation)
The most useful information about game content?
Game aspects with major affect on player
experience?
25. Approach
Model
Design Collect Player’s
Data Emotion
26. Data Collection
40 small levels
(one-third of usual size)
600 game pairs
Features
Six controllable features
Players preferences of engagement
27. Data Collection
40 small levels
(one-third of usual size)
600 game pairs
Features
Six controllable features
Players preferences of engagement
28. Data Collection - Controllable Features
number of gaps
average width of gaps
number of enemies
number of powerups
number of boxes
Enemies placement
Around horizontal boxes
Around gaps
Random placement
35. Content-Driven Preference Learning
• It’s the use of genetic algorithms to evolve the
weight of neural networks to learn preference data.
Levels
36. Content-Driven Preference Learning
• It’s the use of genetic algorithms to evolve the
weight of neural networks to learn preference data.
Levels Segmentation
37. Content-Driven Preference Learning
• It’s the use of genetic algorithms to evolve the
weight of neural networks to learn preference data.
Feature
Levels Segmentation
extraction
38. Content-Driven Preference Learning
• It’s the use of genetic algorithms to evolve the
weight of neural networks to learn preference data.
NeuroEvolutionary
Feature preference
Levels Segmentation
extraction learning
39. Content-Driven Preference Learning
• It’s the use of genetic algorithms to evolve the
weight of neural networks to learn preference data.
NeuroEvolutionary
Feature Player’s
Levels Segmentation preference
extraction Engagement
learning
45. Game Content Representation
Statistical features
Six controllable features
Used for level generation
Sequences
46. Game Content Representation
Statistical features
Six controllable features
Used for level generation
Sequences
Numbers representing different types of game content
o Platform structure, S
o Enemies placement, Ep
o Enemies and items placement, D
55. ANN Implementation
• Multilayer perceptrons (MLPs)
o ANN inputs
• Controllable features
• Sequences as features
o ANN output
• Value of the engagement preference
58. ANN Training
• Genetic algorithms (GAs)
o No prescribed target outputs
• How it works?
players’ magnitude of
reported corresponding
emotional model (ANN)
preferences output
59. ANN Training
• Genetic algorithms (GAs)
o No prescribed target outputs
• How it works?
players’
reported
emotional
preferences - magnitude of
corresponding
model (ANN)
output
77. Game Content Representation
The best-performing MLP models evaluated on occurrences
of frequent subsequences of length three extracted from the 40 levels
78. MLPs Performance on Full Information
about Game Content
The topology and performance of the best MLP models evaluated on full and
partial information about game content. the MLP performance presented is the
average performance over 20 runs.
79. Results
The performance and topologies of MLP models evaluated on full and partial
information of game content using statistics from the game window and from two
and three segments to which the window has been divided. The performance
presented is the average over five runs.
81. Conclusion
Combining both sequential and statistical features
gives better results in predicting players' reported
emotional state.
Partitioning the level causes a significant decrease
(p < 0.05) in the accuracy of predicting player’s
reported engagement. This suggests that there
might be information loss because of decomposing
the data and that this loss causes a performance
decrease.
Multiple perspectives can be done in reference to
this study which is already going on!