Applications of the principle of least effort in data transformation

Abstract

I am interested in applying the Principle of Least Effort to data transformation in an effort to solve three important challenges in using video games as a testbed for the study of inverse reinforcement learning. These challenges are as follows: the very large state space created by high granularity of time, the very large range of feature values provided by quantitative features, and the difficulty to measure similarity of trajectories. Through exploring The Principle of Least Effort from the Social Sciences, I propose a form of data transformation that categorizes data according to the familiarity and preferences of the modeled user. Furthermore I will also show how this approach can be used to create a similarity comparison between trajectories in accordance to The Principle of Least Effort. For the collection of test data I have utilized a reinforcement learning agent to play the minigame BuildMarines for StarCraft II as provided by DeepMind.