Data Representations
23 Jan 2012 - Dr Anton Gerdelan - apg@bth.se

• what kinds of data can we extract from games and simulations?
• choosing the right visualisation(s) for a data set
• some common mistakes to avoid
• some tools to consider

Why Visualise?

• debug complex systems during development:
  • physics middleware: suspension spring compression over race course length
  • game AI: why are characters walking into a wall at point X?
  • what else?
• explain a phenomenon from a simulated experiment:
  • CS AI: at what rate are agents learning compared to previous algorithm/environment?
  • how will an epidemic spread through a population?
• provide extra interactive information
  • medical imaging from scanner-obtained data
  • augmented reality (satellite image, HUD)
• we need to choose our visualisation pipeline carefully
• data sets to use, visualisation methods, standards, tools

Visualisation Options

• static graphs or figures
  • suitable for print media (papers, journals, theses)
  • isolate 1 type of data - easy to appreciate
  • illustrate relationships between data sets
• interactive visualisation
  • virtual camera
  • explore complex virtual world to observe phenomena from different view-points
  • interrupt to dump interesting data sets or to record stills or video
• video and screen captures
  • not valid as scientific results - noise, no error
  • replay interesting phenomena in appreciable way
  • accompany other visualisations
• sometimes it's easier to read data in a table

How to Select Visualisations

• Ask yourself:
  
  Does this visualisation communicate information:
  
  A. more clearly?
  B. less clearly?

How to Select Visualisations

• start with a table if possible
• then choose the most appropriate graphs to display data using the more clear/less clear question.
• use all scientific standards available to increase quality:
  • error measurements or measurements of uncertainty
  • scales, titles, lables, number of points
  • control plots, comparisons
• then consider interactive displays and video to support the measured results

types of data and graphs

charts

safe options line charts bar chart scatter plot keep-it-simple reduce clutter annotate split into two visualisations

trends and continuous quantitative data

• most of the data that we collect is to investigate trends over time (or the increase of some other property)
• examples from games and simulations?

source: Willys Jeep manual, ~1940

• Q. how many data sets are represented here? Q. how could this graph be better?

visualising categories - qualitative data

• sometimes we want to show results of bands or groups of data
• a pie chart is the absolute worst kind of graph
  
  "Death to pie charts" - William Eddy
  
  "...the only worse design than a pie chart is several of them" - Edward Tufte
  
• why? - note: very important point
• avoid 3D graphs
• only use bar charts for qualitative data sets

error measurements and uncertainty

scientific measurements are repeated n times we must represent uncertainty visually use standard error or standard deviation for each point or bar remember: sample standard deviation and standard error is std dev divided by square root of n .: bigger n → small standard error if an error bar is large: collect more data

standard error and standard deviation

we can plot error bars with SE or SD, but what is the difference? normally bars are ±SD or ±SE but why might we double the size? why is this a valuable visualisation tool? we represent this visually on scatter, line, and bar charts with y-axis and sometimes x-axis error bars. what text must go in the caption for the error bars to make sense to the reader?

fig: normal distribution with standard deviations fig: normal distribution with standard error

...

half way...

interpreting error bars

SE bars → 68% confidence that the true mean lies inside the error bars 2X SD bars are called the 95% confidence interval with SE bars two points are significantly different if there is at least 1/2 bar gap between the error bars. 95% confidence interval bars are longer. Less than 1/4 bar of overlap corresponds to P < 0.05 always explain the statistics used in the text or caption n if standard deviation or standard error is used discuss any unsual points or trends

any closer and not significantly different

maps and "snail trails"

• how can we demonstrate motion without a video?
• why are these kinds of visualisations only supporting results? what is the main drawback of this visualisation?

heat maps

heat maps

• illustrate frequency of something happening in geographic regions
• break environment into nodes or grid cells
• when event occurs, add value with large weight to closest node, and smaller weight to nearby nodes
• for real-time, reduce (decay) values over time
• display using colour range or 3D map peaks - what are the drawbacks of each?

Fig (left): path-finding with A* algorithm. Fig (centre): regions of undesirability around opponent soccer robot. Fig (right): 3D game activity

toolkit requirements

• collecting data points?
• working out standard error or standard deviation?
• drawing charts?
  • GNUPlot or other plotting software
  • Spreadsheets?
  • drawn by hand or annotated in publisher software?
• recording, editing, and displaying video?
  • Adobe Premier Pro (editing), regular render-to-texture?
• running interactive visualisations?
  • inside game engine...stand-alone OpenGL...2D rendering engine...Matlab, Maple, GNUPlot, TI-89...
• we will talk about doing all this with real-time data in the next lecture

evaluate these charts

• problems?

• how can we improve this?

evaluate these charts

• the previous figure as a table

evaluate these charts

• if the numbers are removed - how long does it take you to rank them all in size order?

• which is the better comparison of two data sets?

a good data representation has

• axes with consistent scales
• axis labels with units given
• clear plots with control plot or comparison
• no clutter, minimal noise
• supporting caption. explain what the reader is supposed to see. explain any interesting areas.
• error measurements where possible. explain statistics; standard error or deviation or 95% confidence, size of n
• key to explain heat map colour ranges
• key to explain what the different data sets are
• always provide and explain any formulae used - how is the distance between nodes or the colour calculated?

1. can the viewer determine the difference, in units, between data points in your visualisation?
2. have you provided enough information so that another scientist can visualise their own data set and compare results with you?

problem?

• what should be put in the caption? (hint: at least 6 things)

further reading

• any good statistics for science book
  • standard deviation, standard error
  • linear regression, least-squares
  • probability and normal distributions