Human-Like Playtesting with Deep Learning

August 2018

DOI:10.1109/CIG.2018.8490442

Prior work is the use of monte carlo tree search agents to simulate playing through a level and using "survival analysis" to gauge difficulty of the level.

Extensive gameplay analytics are used to record players performance on levels which is then used to train agents to immitate human play.

Use of CNN (convolutional neural network) on grid representations of the gameplay levels.

Robots are trained to immitate the plays that actual humans made; done so that the robot generates player-like results that can be used to see what a human would do on a newly created level.

Creating an "interest" metric from players and training the robot to immitate the same result.