Xavier Anadón, Pablo Sanahuja, V. Javier Traver, A. Lopez, J. Ribelles

Previous work

Web-based cube puzzle game with set of puzzles with different visual modifications offered different challenges. But players had different skills as well.

overview

Goal

Could we characterise players from their mouse and click activities in different puzzles?

Eventually, this could lead to (dynamic) difficulty adaptation.

Overview

2-level BoW

Coding interactions

Two image representations of player behavior were explored: position-based and click-based. Both code relative time. The click-based was used in this work. 

vfly2eow 2 1 tr   vfly2eow 2 1 cl

Feature learning

 Supervised learning task: Regression of time and clicks for each interaction (puzzle & player).

Goal: learn deep features to compactly encode the interaction.

 regressingTimeClicks

The 2-level BoW

 2levelBoWMadeEasy

Results

 Player groups roughly correspond to their actual effort (time, clicks)

times clicks stvc

Tentative qualitative interpretations:

  • Group 1: took the longest, they are aware (Q3), but would play again (Q4)
    • skill and challenge are aligned; they may be offered to play the same puzzles
  • Group 3: took the least, are aware (Q1, Q3), but likely would not play again (Q4)
    • skills higher than challenge; they may be offered harder puzzles

 question 1question 2question 3question 4

Conclusions

  • Preliminary results suggest the framework is suitable to compare puzzles and players (in terms of their played puzzles)
  • The identified player profiles encode not only performance but, tentatively, also opinions and traits
  • Possible future work:
    • The trajectory-based representation might capture solving strategies
    • More data would be required from actual players or synthetic or agent-based
    • Embeddings from preferences and opinions, not (only) from behavioural data, might provide rich predictive cues
    • Use LSTM-like representation to account for the sequential nature of data, using image encoding or raw mouse data

Feedback from UMAP participants

  • How to better represent and process information from incomplete puzzles for on-line prediction?
  • How to exploit player characterisation for actual puzzle selection as a form of personalisation?
  • Do any techniques for recommender systems be adequate and easily adaptable?
  • How to plan experiments and measure the effect of the personalisation?
  • How could intelligence-like traits be estimated from players’ in-game data?
  • How could our approach be integrated with generative models?