Machine learning methods are powerful but are hard to understand, especially if they show complex temporal behavior as part of their training process or when being used as autonomous agents. Group member Shivam Agarwal, together with colleagues, has developed two new interactive visualization approaches that literally unroll the complex behavior on the screen and make readable what is going on across time.

The first visualization approach depicts the behavior of artificial agents playing a computer game. The specific game studied is a variant of the traditional Bomberman game, where players lay bombs on a board and try to eliminate each other. As bomb laying strategies are analyzed to get a better understanding of success factors, the approach is called “Bombalytics”. It has already been used in competitions among researchers to develop the best artificial agent.

The second visualization technique can show how an automatic method learns to get better and better in classifying image data into different categories. Changing classifications during this training process are depicted as visual streams that flow between the categories. The approach, called “Set Streams”, also supports other scenarios and applies to all sorts of overlapping categories that change across time.

Both approaches were recently presented at EuroVis 2020, and they will appear as two independent articles in the international journal Computer Graphics Forum. The conference was originally scheduled to be held in Norrköping, Sweden, but – like other conferences – had to transform into a virtual event, with interactive live video streaming sessions.