This is a presentation me and a friend created in the context of Multiagent Systems. We were asked to present the paper "A Sequential Recommendation Approach for Interactive
Personalized Story Generation" by Hong Yu and Mark O.Riedl
Climate Change Impacts on Terrestrial and Aquatic Ecosystems.pptx
A Sequential Recommendation Approach for Interactive Personalized Story Generation paper presentation
1. Παντελιδάκης Μηνάς
Θεοδωρής Χαράλαμπος
Παρουσίαση θεωρητικής εργασίας
A Sequential Recommendation Approach for Interactive
Personalized Story Generation
Hong Yu and Mark O. Riedl
College of Computing
Georgia Institute of Technology
Atlanta, GA 30332
{hong.yu, riedl}@cc.gatech.edu
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
2. There is a whole genre of games known as interactive
story games. To modernize and make them more
appealing, developers have been using agents (better
known as Drama Managers) that build a player profile
deriving from the player’s in game actions.
Hong Yu and Mark O. Riedl approach this issue by
constructing the player model based on the user’s
preferences (ratings) instead of their actions.
Introduction
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
3. In this paper…
● Drama Manager as surrogate for both the human designer and
the user
● Prefix-based collaborative filtering for modeling the user’s
preference profile
● Two learning algorithms, probabilistic Principal Component
Analysis (pPCA) and Non-negative Matrix Factorization (NMF)
● Validation testing on both human and simulated users
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
4. ●
Drama Manager
● Story Library
● Player Model
● User
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Setup
5. But what is a Drama Manager?
An omniscient agent that acts to
bring about a particular sequence
of plot points for the user to
experience
● Monitors the system state
● Selects the optimal next plot point from the story library
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
6. Story Library
● Plot points form story graphs after termporal and semantic
constraints are applied
●
Story graphs are later transformed into prefix graphs and stored
in the story library
● Notice that the prefix graph will be a tree or forest since any
prefix cannot have more than one parent node
Story graph prefix graph → tree
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
7. Player Model
●
Extracts the dimensions of the users’ preferences from the users’
ratings
●
It is not constrained by pre-defined dimensions
●
Assumes that people who share similar preference in the past tend to
share it again in the future, hence it utilizes a form of collaborative
filtering
●
Stories are presented plot point by plot point and a preference rating
for the story-so-far is collected after every plot point
●
Ratings are stored in the story library along with the story prefix
●
The prefix-rating matrix is obtained
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
8. Player Model+
The prefix-rating matrix
● Stars (*) represent missing ratings.
● Dimensions nxm (n prefixes, m users)
●
Input for both pPCA and NMF
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
9. Principal Component Analysis
● A technique of dimensionality reduction
● Aims at finding some correlation among data of a large
data set
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
10. Probabilistic
Principal Component Analysis
● Assumes the prefix-rating vector obeys a multi-
dimensional Gaussian distribution
● If the user-prefix matrix contains missing values, the EM
algorithm can be used to compute W and σ.
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
11. Expectation Maximization
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
●
Iterative algorithm with two phases
●
Starts with random Gaussian distributions
●
Phase 1: correlate data to given Gaussian distributions
●
Uses soft clustering to assign data to Gaussian distributions
●
Phase 2: estimate new Gaussian distributions given the data
12. Non-negative Matrix Factrorization
● Constrains data to non-negative values
● Constrains only to additive combinations
● Given t types, creates W (weights-by-types)
● and H (types-by-users) vectors.
● Uses EM to determine best W and H values, using existing data.
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
13. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Player Modeling Steps
● Build the story library
● Collect the ratings and populate the matrix R
● Compute W, σ and μ for the pPCA, or W for NMF
For A new user:
● Model his preference using r, computing x for pPCA or h for NMF
● Calculate the full rating vector r’ from x
● Select the best full-length story, descendant of the current prefix
● Collect the user’s rating on the story-so-far
● Include the new rating into r and goto step 1
Training the model:
14. ● A story library based on Choose-Your-Own-
Adventure books was transcribed
● Both human and simulated testers
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Testing
15. ● Four choose-your-own-adventure books for better story quality
● The Abominable Snowman, Journey Under the Sea, Space and Beyond,
●
The Lost Jewels of Nabooti
● Transformed into prefix graphs, containing 154 possible stories and 326 prefixes
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Story Library Components
17. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Model Training Phase
●
31 users, only 5 exposed to similar games
●
Each user reads and rates 10 random stories for initialization
●
Obtained a 326-by-31 prefix matrix R with ~86% ratings missing
●
Split R to 90% training set / 10% validation set randomly 10 times
●
Each time the RMSE is calculated after training
● The average of all the training data for that session suggested there
were 6 types of users
18. ● Represents the sample standard deviation of
● the differences between predicted values and observed values
● Intuitively, finds out how good the expectation is by calculating the
distance of real datapoints from the expected function
● The smaller the RMSE, the better our expectation
ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Root Mean Square Error
19. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Procedure & results
●
22 new users alongside 11 training ones
● 5 random stories read and rated
● DM chooses personalized stories based on user preference
● Final NMF and pPCA models are trained
● The stories are presented to the users and rated
●
DM serves another 5 personalized stories to the users
●
followed by 5 random stories to eliminate prejudice
20. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Experiments with simulated users
Same steps as with humans
Many more (consistent) users
21. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Model Training phase
● Built based upon Robin’s Law (5 gamer types)
● 5-dimensional characteristic vector
● Prefixes on library were labeled according to those types
● 120 randomized one-type-only users
● 10 stories read and rated for initialization
● A 326-by-120 prefix-rating matrix was populated
22. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Story Generation phase
● 1000 fully random generated users tested the algorithm exactly
as the human ones.
● Baseline P and N: PPCA and NMF without prefixes, rated only at
the end.
● Vector: A vector that simulates the user, all stats are initialized at
0, get added values as the game progresses
● PPCA: The proposed PPCA algorithm
● NFMwP, NFMwoP : The proposed NFM algorithm with and without
prior knowledge of player types (Robin’s Law)
23. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Results
● The prefix algorithms are a lot more accurate than their non-
prefix counterparts
● NMF algorithms outclass every other algorithm even without
prior knowledge
24. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Results+
● Average accuracies as the number of story changes
● Prefix-based algorithms shine from the second story forth
● In particular, NMF algorithms are very accurate from the start
25. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Results++
● Average RMSEs of the prefix-based algorithms with different
numbers of training users/data
● NMF with prior is the most accurate from start to finish
26. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Conclusion
● Future work could combine structured (user preference) and
unstructured feedback (user actions) to further improve player
modeling
● Machine learning + human intuition = Successful storytelling
27. ΠΛΗ 517 – Πολυπρακτορικά Συστήματα – Χειμώνας 2017
Congratulations!
You have endured our presentation!
We hope you found the topic as interesting as we did, and we encourage
you to engage us backstage on the topic!
Good Luck to the upcoming teams!