Using q-learning for sequencing level difficulties in a citizen science matching game

Abstract

Sequencing task difficulty and variety can be a powerful tool for increasing engagement in online citizen science platforms. The abundance of available participant data presents great promise for machine learning oriented approaches to making tasks more engaging for participants. We present a web game for image matching called Tile-o-Scope Grid, and explore using a Q-learning based algorithm to generate a policy for sequencing level difficulties. Recruiting players using Amazon Mechanical Turk, we gathered data to train and evaluate approaches to sequencing level difficulties in Tile-o-Scope Grid. Comparisons of our Q-learning based algorithm with uniform random and greedy baselines suggest potential for using reinforcement learning for citizen science image labeling.