Gameplay Analysis through State Projection Erik Andersen 1, Yun-En Liu 1, Ethan Apter 1, François Boucher-Genesse 2, Zoran Popović 1 1 Center for Game Science Department of Computer Science University of Washington 2 Department of Education Université du Québec à Montréal FDG 2010 June 21 st, 2010
We want to know how people play
?
We want to find…
Player confusion
We want to find… Player confusion Player strategies
We want to find… Player confusion Player strategies Design flaws
Patterns in data SELECT * FROM replays WHERE location=x AND time>y AND attempt>3 AND death=“grenade”…
Patterns in data SELECT * FROM replays WHERE location=x AND time>y AND attempt>3 AND death=“grenade”… Confusion? Strategies?
Traditional Playtesting
Statistical Methods Surveys In-game statistics
Statistical Methods Surveys In-game statistics
Visual Data Mining Lets people see patterns in data Bungie (Halo 3)
Visual Data Mining Lets people see patterns in data Dynamic information? Bungie (Halo 3)
Visual Data Mining Lets people see patterns in data Dynamic information? Games with no map? Bungie (Halo 3)
But what about?
“Playtraces” GoalStart
“Playtraces” GoalStart
“Playtraces” GoalStart
“Playtraces” GoalStart Confusion? Distance to goal
Refraction
Massive educational data mining
2-D projection of points in high-dimensional space Clusters game states based on some distance function Classic Multidimensional Scaling
State Distance
Action Distance d a (s 1, s 2 )
State Distance GoalStart Confusion? Distance to goal
Distance to Goal d g (s 1, s 2 ) = abs(d g (s 1, s g ) - d g (s 2, s g ))
Distance Functions Action distanceCombinedDistance to goal
Refraction Distance Function d (s 1, s 2 ) = (d a (s 1, s 2 ) + d g (s 1, s 2 )) / 2
Playtracer Framework
Easy level
Difficult level
Failure
Chance To Win
Evaluation
35 children from K12 Virtual Academies
Evaluation 35 children from K12 Virtual Academies Mostly third and fourth-graders
Evaluation 35 children from K12 Virtual Academies Mostly third and fourth-graders About 15 levels
Evaluation 35 children from K12 Virtual Academies Mostly third and fourth-graders About 15 levels The game logged all player actions
Analysis
Player confusion
Analysis Player confusion Player hypotheses
Analysis Player confusion Player hypotheses Design flaws
Analysis Player confusion Player hypotheses Design flaws
Level 2
Level 2 Solution
Level 2 Visualization
Confusion: Hitting target from wrong side
Refinement
Confusion: Using pieces incorrectly
Analysis Player confusion Player hypotheses Design flaw
Level 4
Level 4 Solution
Level 4 Visualization
Hypothesis: Satisfy bottom target
Hypothesis: Get laser near targets
Hypothesis: Overload bottom target
Analysis Player confusion Player hypotheses Design flaws
Level 4 Visualization
Design flaw: Deadly state
Refinement
Limitations Difficult to find good distance function
Limitations Difficult to find good distance function
Limitations Difficult to find good distance function
Limitations Large game spaces
Conclusions Useful for game analysis
Conclusions Useful for game analysis We are expanding and refining Playtracer
Big Open Problems How to
Big Open Problems How to – specify distances between game states
Big Open Problems How to – specify distances between game states – differentiate types of confusion
Big Open Problems How to – specify distances between game states – differentiate types of confusion – classify player strategies
Acknowledgements Marianne Lee Emma Lynch Justin Irwen Happy Dong Brian Britigan Dennis Doan François Boucher-Genesse Seth Cooper Taylor Martin John Bransford David Niemi Ellen Clark Funding: NSF Graduate Fellowship, NSF, DARPA, Adobe, Intel, Microsoft
Cycles
Acyclic Paths
Player Tracking